<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by KloudExams | Google Cloud Certification Prep on Medium]]></title>
        <description><![CDATA[Stories by KloudExams | Google Cloud Certification Prep on Medium]]></description>
        <link>https://medium.com/@kloudexams?source=rss-6ef48558799e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*R_Ov0XoBOcUhNvwNdJ6K2Q.png</url>
            <title>Stories by KloudExams | Google Cloud Certification Prep on Medium</title>
            <link>https://medium.com/@kloudexams?source=rss-6ef48558799e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 22:53:17 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kloudexams/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[GCP Google Cloud Professional Cloud Developer Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-professional-cloud-developer-practice-questions-e06de2cfd18f?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e06de2cfd18f</guid>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[google-cloud]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 12:02:44 GMT</pubDate>
            <atom:updated>2026-10-07T12:02:44.921Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dEHnjcKjQOgJxg9fEKWinA.jpeg" /><figcaption>KloudExams GCP Google Cloud Professional Cloud Developer Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Professional Cloud Developer certification</strong>? Knowing what development tools and cloud services do is a starting point — understanding how to use them to build, test, and deploy secure, scalable applications is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing development, integration, and deployment approaches that fit specific application requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on application development.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>At Raventra Cloudworks, an API backend runs across multiple cloud providers. You want to collect distributed request traces and generate reports that show the latency of API requests across the backend. Which two steps should you take? (Choose two.)</strong></p><p><strong>A. Use a Zipkin-compatible collector to gather trace data.</strong></p><p><strong>B. Use the Ops Agent configured to collect only logs and infrastructure metrics.</strong></p><p><strong>C. Use Cloud Trace to generate latency reports.</strong></p><p><strong>D. Use an application debugger to generate the report.</strong></p><p><strong>E. Use Cloud Profiler to generate the report.</strong></p><blockquote><strong><em>A is correct because distributed tracing requires trace spans to be collected from the services that participate in a request, including services running outside Google Cloud. Zipkin-compatible instrumentation and collectors can gather distributed trace information, and Cloud Trace can work with trace data produced by Zipkin-compatible systems when the data is exported appropriately. For a new implementation, OpenTelemetry is also a modern standard choice, but within the listed options the Zipkin-compatible collector is the tracing-oriented collection mechanism.</em></strong></blockquote><p>B is incorrect because this option configures the Ops Agent to collect only logs and infrastructure metrics, which do not provide the trace spans and parent-child relationships needed to reconstruct distributed requests. The Ops Agent does support collecting OTLP traces from instrumented applications on supported Compute Engine instances when its OTLP receiver and trace pipeline are configured. However, that tracing configuration is excluded by this option, so it does not collect the distributed trace data required here.</p><blockquote><strong><em>C is correct because Cloud Trace is the Google Cloud distributed tracing service specifically designed to collect, store, explore, and analyze request latency data and execution paths. It can analyze trace information from applications beyond Google Cloud, making it suitable for a multi-cloud API backend when the applications are instrumented to export trace spans.</em></strong></blockquote><p>D is incorrect because an application debugger is intended to inspect program execution and state while diagnosing code behavior. It is not a distributed tracing backend and does not aggregate request spans across multiple services and cloud providers into latency reports.</p><p>E is incorrect because Cloud Profiler continuously samples CPU, memory, contention, thread, and wall-time characteristics to identify where application resources are consumed. Profiling is valuable for code-level performance optimization, but it does not reconstruct distributed request paths or produce the cross-service latency traces required by this question.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/trace/docs/trace-labels">docs.cloud.google.com/trace/docs/trace-labels</a></p><p><a href="http://docs.cloud.google.com/trace/docs/otlp">docs.cloud.google.com/trace/docs/otlp</a></p><p><a href="http://docs.cloud.google.com/profiler/docs/about-profiler">docs.cloud.google.com/profiler/docs/about-profiler</a></p><p><strong>Q2.</strong></p><p><strong>At Kelvanta Labs, you have written a Cloud Run function that accesses other Google Cloud resources. You want to secure the environment according to the principle of least privilege. What should you do?</strong></p><p><strong>A. Create a new service account that has the Editor basic role to access the resources. Give the deployer permission to obtain access tokens for the service account.</strong></p><p><strong>B. Create a new service account that has a custom IAM role containing the permissions needed to access the resources. Give the deployer permission to obtain access tokens for the service account.</strong></p><p><strong>C. Create a new service account that has the Editor basic role to access the resources. Give the deployer permission to act as the new service account.</strong></p><p><strong>D. Create a new service account that has a custom IAM role containing the permissions needed to access the resources. Give the deployer permission to act as the new service account.</strong></p><p>A is incorrect because the Editor basic role grants a very broad set of permissions and does not satisfy least privilege for a function that only needs access to specific resources or operations. In addition, permission to mint access tokens is an impersonation capability and is not the normal permission required merely to attach the runtime service account when deploying the function.</p><p>B is incorrect because restricting the runtime service account to only the permissions the function requires is appropriate, but giving the deployer access-token creation capability is not the correct minimum permission for attaching that service account to the deployed workload. The relevant deployment permission is iam.serviceAccounts.actAs, commonly provided through the Service Account User role on the specific runtime service account.</p><p>C is incorrect because allowing the deployer to act as the runtime service account is the correct type of deployment permission, but assigning the Editor basic role to the service account makes the runtime identity excessively privileged. If the function is compromised, those broad Editor permissions would increase the potential impact.</p><blockquote><strong><em>D is correct because a dedicated runtime service account with only the permissions needed by the function limits what the workload can do. If no sufficiently narrow predefined role fits the exact requirement, a custom role can contain only the required supported permissions. The deployer should have permission to act as that specific service account so it can be attached to the Cloud Run function. This combination best satisfies the least-privilege requirement among the available options.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/deploy-functions">docs.cloud.google.com/run/docs/deploy-functions</a></p><p><a href="http://docs.cloud.google.com/run/docs/configuring/services/service-identity">docs.cloud.google.com/run/docs/configuring/services/service-identity</a></p><p><a href="http://docs.cloud.google.com/iam/docs/choose-role-type">docs.cloud.google.com/iam/docs/choose-role-type</a></p><p><strong>Q3.</strong></p><p><strong>At Virelune Systems, you are developing an internal application that will allow employees to organize community events within the company. You deployed the application on a single Compute Engine instance. Your company uses Google Workspace, and you need to ensure that company employees can authenticate to the application from anywhere. What should you do?</strong></p><p><strong>A. Add an external IP address to the instance, and restrict access to the instance using VPC firewall rules. Allow the company’s proxy as the only source IP address.</strong></p><p><strong>B. Add an external Application Load Balancer in front of the instance, and set up Identity-Aware Proxy (IAP). Configure IAP access so that the company’s Google Workspace domain can access the application.</strong></p><p><strong>C. Set up a Cloud VPN tunnel between the company network and the instance’s VPC network on Google Cloud. Configure the required firewall rules and routing information for both the on-premises and Google Cloud networks.</strong></p><p><strong>D. Add an external IP address to the instance, and allow traffic from the internet. Generate a random hash, create a subdomain containing the hash that points to the instance, and distribute that DNS address to company employees.</strong></p><p>A is incorrect because restricting access to a company proxy by source IP does not provide identity-based authentication for employees who must be able to connect from anywhere. Employees would generally need to route through the proxy or another controlled corporate network path, and an IP-based restriction identifies the source network rather than the individual user. It also exposes the VM directly through an external IP instead of placing an identity-aware access layer in front of the application.</p><blockquote><strong><em>B is correct because Identity-Aware Proxy protects web applications by authenticating and authorizing users before requests reach the application. A Compute Engine backend can be placed behind an Application Load Balancer and protected with IAP, and access can be granted to principals from the organization’s Google Workspace domain. Employees can therefore authenticate with their organizational identities from different locations without requiring direct network-level access to the VM. This provides centralized identity-based access control and avoids relying on secret URLs or a corporate source IP.</em></strong></blockquote><p>C is incorrect because Cloud VPN provides encrypted network connectivity between networks, but it does not by itself provide the required per-user authentication for employees connecting from arbitrary locations. Employees outside the company network would first need access to that network or another remote-access mechanism. This introduces unnecessary network dependency when IAP can directly protect the web application with organizational identity.</p><p>D is incorrect because a random or difficult-to-guess DNS name is not an authentication or authorization control. Anyone who learns or discovers the address could attempt to access the application, and the VM remains directly exposed to the internet. Security should be based on verified identity and explicit authorization rather than obscurity.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/iap/docs/enabling-compute-howto">docs.cloud.google.com/iap/docs/enabling-compute-howto</a></p><p><a href="http://docs.cloud.google.com/iap/docs/concepts-overview">docs.cloud.google.com/iap/docs/concepts-overview</a></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/application-load-balancer">docs.cloud.google.com/load-balancing/docs/application-load-balancer</a></p><p><strong>Q4.</strong></p><p><strong>At Orbinexa Digital, you are developing a web application that will be accessible over both HTTP and HTTPS and will run on Compute Engine instances. Occasionally, you need to connect by SSH from a remote laptop to one of the instances to perform application maintenance. How should you configure the environment while following Google-recommended best practices?</strong></p><p><strong>A. Place Compute Engine web server instances that have only internal IP addresses behind an external proxy Network Load Balancer.</strong></p><p><strong>B. Configure VPC firewall rules to allow all ingress traffic to the Compute Engine web servers, and give each server a unique external IP address.</strong></p><p><strong>C. Configure Identity-Aware Proxy (IAP) TCP forwarding for SSH access. Configure the Compute Engine web servers without external IP addresses and place them behind an external Application Load Balancer for application traffic.</strong></p><p><strong>D. Configure the Compute Engine web servers without external IP addresses and place them behind an external Application Load Balancer. Set up a bastion host with an external IP address and open SSH access to the bastion host, and then connect to the web instances through the bastion host.</strong></p><p>A is incorrect because an external proxy Network Load Balancer can proxy TCP traffic, but this option does not address the administrative SSH requirement with an identity-aware access mechanism. It also uses a lower-layer load-balancing approach where an Application Load Balancer is the natural fit for HTTP and HTTPS application traffic.</p><p>B is incorrect because allowing unrestricted ingress and assigning an external IP to every server substantially increases the attack surface. Google-recommended Compute Engine access patterns avoid exposing SSH directly to the public internet when IAP TCP forwarding can provide authenticated administrative access without requiring external IP addresses on the target VMs.</p><blockquote><strong><em>C is correct because IAP TCP forwarding allows authorized administrators to establish SSH connections to Compute Engine VMs that do not have external IP addresses. Access can be controlled through IAM and appropriate firewall rules rather than opening SSH to the internet. For the web application, an external Application Load Balancer can expose HTTP and HTTPS while the backend instances retain internal-only addresses. This separates public application ingress from protected administrative access and minimizes direct VM exposure.</em></strong></blockquote><p>D is incorrect because a bastion host can provide an administrative entry point, but it introduces another internet-facing VM that must be patched, monitored, secured, and restricted. IAP TCP forwarding provides the required SSH path without maintaining a public bastion host, so option C better follows the managed, identity-aware access pattern recommended for this scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/connect/ssh-using-iap">docs.cloud.google.com/compute/docs/connect/ssh-using-iap</a></p><p><a href="http://docs.cloud.google.com/compute/docs/connect/ssh-best-practices/network-access">docs.cloud.google.com/compute/docs/connect/ssh-best-practices/network-access</a></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/application-load-balancer">docs.cloud.google.com/load-balancing/docs/application-load-balancer</a></p><p><strong>Q5.</strong></p><p><strong>At Nexquora Systems, you want to create fully baked, or golden, Compute Engine images for an application. You need to bootstrap the application so that it connects to the appropriate database for the environment in which it runs: test, staging, or production. What should you do?</strong></p><p><strong>A. Embed the appropriate database connection string in the image, and create a different image for each environment.</strong></p><p><strong>B. When creating a Compute Engine instance, add a network tag containing the name of the database to which it should connect. In the application, query the Compute Engine API for the instance’s network tags and use the tag to construct the database connection string.</strong></p><p><strong>C. When creating a Compute Engine instance, create a custom metadata item with the key DATABASE and a value containing the appropriate database connection information. In the application, read a DATABASE environment variable and use its value to connect to the database.</strong></p><p><strong>D. When creating a Compute Engine instance, create a custom metadata item with the key DATABASE and a value containing the appropriate database connection information. In the application, query the Compute Engine metadata server for the DATABASE value and use it to select the appropriate database connection configuration.</strong></p><p>A is incorrect because creating a separate machine image for every environment couples the immutable application artifact to environment-specific configuration. A golden-image approach is more maintainable when the same tested image can be promoted through test, staging, and production while deployment-time configuration selects the environment. Any sensitive database credentials should also be managed separately with an appropriate secret-management mechanism rather than baked into an image.</p><p>B is incorrect because Compute Engine network tags are primarily intended for network-related configuration such as firewall targeting and routes, not general application configuration. Querying the Compute Engine API also requires additional API access and logic when instance metadata provides a purpose-built, local mechanism for supplying VM configuration.</p><p>C is incorrect because custom Compute Engine metadata is not automatically exported to the guest operating system as an environment variable named after the metadata key. Creating a DATABASE metadata entry therefore does not make a DATABASE environment variable appear inside the application process. Additional startup logic would be required to fetch the metadata and set such a variable.</p><blockquote><strong><em>D is correct because custom instance metadata can supply environment-specific configuration at VM creation time while keeping the machine image identical across environments. Workloads on the VM can query the metadata server locally to retrieve the custom DATABASE value and select the correct database configuration. This preserves the reusable golden image and externalizes environment-specific settings. Sensitive credentials should still be stored in a dedicated secret-management service rather than treated as ordinary metadata.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/metadata/overview">docs.cloud.google.com/compute/docs/metadata/overview</a></p><p><a href="http://docs.cloud.google.com/compute/docs/metadata/setting-custom-metadata">docs.cloud.google.com/compute/docs/metadata/setting-custom-metadata</a></p><p><a href="http://docs.cloud.google.com/compute/docs/metadata/querying-metadata">docs.cloud.google.com/compute/docs/metadata/querying-metadata</a></p><p><strong>Q6.</strong></p><p><strong>At Pelvaris Tech, your team develops stateless services that run on Google Kubernetes Engine. You need to deploy a new service that will be accessed only by other services running in the GKE cluster. The service must scale quickly in response to changing load. What should you do?</strong></p><p><strong>A. Use Vertical Pod Autoscaling to scale the containers, and expose them through a ClusterIP Service.</strong></p><p><strong>B. Use Vertical Pod Autoscaling to scale the containers, and expose them through a NodePort Service.</strong></p><p><strong>C. Use Horizontal Pod Autoscaling to scale the containers, and expose them through a ClusterIP Service.</strong></p><p><strong>D. Use Horizontal Pod Autoscaling to scale the containers, and expose them through a NodePort Service.</strong></p><p>A is incorrect because ClusterIP is appropriate for cluster-internal access, but Vertical Pod Autoscaling adjusts Pod resource requests and limits rather than rapidly increasing the number of stateless replicas to handle additional concurrent load. Depending on its mode and workload, applying vertical recommendations can also require Pod recreation. Horizontal scaling is a better match for a stateless service whose request volume changes.</p><p>B is incorrect because Vertical Pod Autoscaling is not the best scaling mechanism for this stateless request-serving workload, and NodePort unnecessarily exposes the service on a port of each node. The requirement says the service is consumed only by other workloads within the cluster, so there is no need for node-level exposure.</p><blockquote><strong><em>C is correct because Horizontal Pod Autoscaling changes the number of Pod replicas based on observed metrics, making it appropriate for a stateless service that needs to respond to changing load. A ClusterIP Service exposes the service only through a cluster-internal virtual IP and DNS name, which matches the requirement that only services within the GKE cluster access it. The combination provides the appropriate scaling model and the narrowest required network exposure.</em></strong></blockquote><p>D is incorrect because Horizontal Pod Autoscaling is appropriate, but NodePort exposes the service on every node’s IP address and an allocated port, which is unnecessary when all clients are internal to the cluster. ClusterIP provides the required internal service discovery and connectivity without the extra exposure.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/horizontalpodautoscaler">docs.cloud.google.com/kubernetes-engine/docs/concepts/horizontalpodautoscaler</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/scaling-apps">docs.cloud.google.com/kubernetes-engine/docs/how-to/scaling-apps</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/service">docs.cloud.google.com/kubernetes-engine/docs/concepts/service</a></p><p><strong>Q7.</strong></p><p><strong>At Daxoria Systems, you manage an application that runs on a Compute Engine instance. You also have multiple backend services running in containers on Compute Engine instances. The backend instances are scaled by managed instance groups in multiple regions. You want the calling application to remain loosely coupled from individual backend instances. You also need to invoke different service implementations depending on the value of an HTTP header in each request. Which Google Cloud feature should you use?</strong></p><p><strong>A. Cloud Service Mesh with Google Cloud service routing APIs</strong></p><p><strong>B. Service Directory</strong></p><p><strong>C. Cloud Service Mesh with Istio APIs</strong></p><p><strong>D. A regional internal Application Load Balancer</strong></p><blockquote><strong><em>A is correct because Cloud Service Mesh’s Google Cloud service routing APIs are designed to provide service-to-service routing for workloads such as Compute Engine VM applications while keeping clients decoupled from individual backend endpoints. The routing model can use managed backend services and routes, including HTTP header match conditions, to direct requests to different backend implementations. This is well suited to Compute Engine workloads backed by managed instance groups across regions and provides centrally managed application-layer service routing rather than requiring the caller to discover and select individual instances.</em></strong></blockquote><p>B is incorrect because Service Directory provides service registration and discovery. It can help clients locate service endpoints, but it is not itself an HTTP traffic-management layer that evaluates request headers and selects one backend implementation rather than another. The caller or another proxying layer would still need to implement the routing decision.</p><p>C is incorrect because the Istio API path in Cloud Service Mesh is primarily the service-mesh configuration model for Kubernetes-based workloads. This scenario specifically describes services running on Compute Engine instances and managed instance groups, for which the Google Cloud service routing APIs are the appropriate Cloud Service Mesh control model among these choices.</p><p>D is incorrect because a regional internal Application Load Balancer is regional in scope and does not represent the intended global service-mesh routing model for backend managed instance groups distributed across multiple regions. Although Application Load Balancers can provide advanced HTTP routing in applicable architectures, this option does not provide the same Compute Engine service-routing control plane and multi-region service-mesh abstraction required by the scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/service-mesh/docs/service-routing/service-routing-overview">docs.cloud.google.com/service-mesh/docs/service-routing/service-routing-overview</a></p><p><a href="http://docs.cloud.google.com/service-mesh/docs/service-routing/advanced-traffic-management">docs.cloud.google.com/service-mesh/docs/service-routing/advanced-traffic-management</a></p><p><a href="http://docs.cloud.google.com/service-mesh/docs/supported-platforms">docs.cloud.google.com/service-mesh/docs/supported-platforms</a></p><p><strong>Q8.</strong></p><p><strong>At Quantivra Systems, you are creating and running containers across different Google Cloud projects. The application you are developing needs to access Google Cloud services from within Google Kubernetes Engine (GKE). What should you do?</strong></p><p><strong>A. Assign an IAM service account to the GKE nodes and let application Pods use the node identity.</strong></p><p><strong>B. Use Workload Identity Federation for GKE so the Pod’s Kubernetes ServiceAccount receives the appropriate Google Cloud permissions.</strong></p><p><strong>C. Store IAM service account credentials as a Kubernetes Secret and mount or expose them to the Pod.</strong></p><p><strong>D. Use an IAM service account together with Kubernetes role-based access control (RBAC) to authorize access to Google Cloud services.</strong></p><p>A is incorrect because the identity attached to a GKE node is shared by workloads that run on that node unless stronger workload identity controls are used. Granting application permissions to the node service account can therefore give multiple Pods broader access than they require and makes least-privilege isolation between workloads difficult. Google recommends giving Pods their own workload identity through Workload Identity Federation for GKE instead of relying on node credentials.</p><blockquote><strong><em>B is correct because Workload Identity Federation for GKE is the recommended mechanism for GKE workloads that need to call Google Cloud APIs. A Kubernetes ServiceAccount can be used as an IAM principal or associated with an IAM service account, allowing the workload to obtain short-lived credentials without storing long-lived service account keys. Permissions can be granted specifically to the workload that needs them, even when resources are located in another project, which supports least privilege and avoids credential distribution.</em></strong></blockquote><p>C is incorrect because a Kubernetes Secret can store a service account key, but the key remains a long-lived credential that must be created, distributed, protected, rotated, and revoked. If the Secret or Pod is compromised, the key can potentially be used outside the cluster. Workload Identity Federation for GKE removes the need to manage downloadable service account keys and is the preferred security model.</p><p>D is incorrect because Kubernetes RBAC governs access to Kubernetes API resources such as Pods, Deployments, and Secrets. It does not grant a workload permission to call Google Cloud APIs such as Cloud Storage, Pub/Sub, or Spanner. Google Cloud access must be authorized through IAM, and Workload Identity Federation for GKE provides the recommended bridge between Kubernetes workload identities and IAM.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity">docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/service-accounts">docs.cloud.google.com/kubernetes-engine/docs/how-to/service-accounts</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/workload-identity">docs.cloud.google.com/kubernetes-engine/docs/how-to/workload-identity</a></p><p><strong>Q9.</strong></p><p><strong>At Veltrixa Platforms, you run a containerized application on Google Kubernetes Engine. Your container images are stored in Artifact Registry, and your team uses an automated CI/CD process. You need to prevent container images with known critical vulnerabilities from being deployed. What should you do?</strong></p><p><strong>A. Use Web Security Scanner to crawl the running web application, review application logs for scan findings, create an attestation stating that the container has no known critical vulnerabilities, and require the attestation with Binary Authorization before deployment.</strong></p><p><strong>B. Use Web Security Scanner to crawl the running web application, manually review the web scan results in the Google Cloud console, create an attestation stating that the container has no known critical vulnerabilities, and require the attestation with Binary Authorization before deployment.</strong></p><p><strong>C. Enable Artifact Analysis automatic vulnerability scanning for the Artifact Registry repository, manually review the image vulnerability results in the Google Cloud console, create an attestation when the image has no disallowed critical vulnerabilities, and require that attestation with Binary Authorization before deployment.</strong></p><p><strong>D. Enable Artifact Analysis automatic vulnerability scanning for the Artifact Registry repository, programmatically evaluate the vulnerability occurrences produced for each image, create an attestation only when the image satisfies the vulnerability policy, and use Binary Authorization to require that attestation before the image can be deployed.</strong></p><p>A is incorrect because Web Security Scanner evaluates deployed web applications for certain web-application vulnerabilities; it does not inspect the operating-system and language packages inside a container image for known CVEs before deployment. Application logs therefore cannot supply the artifact vulnerability evidence required for this software-supply-chain gate.</p><p>B is incorrect because Web Security Scanner is still the wrong scanning mechanism. Its target is a running web application rather than the package inventory of the image stored in Artifact Registry. Manual review also conflicts with the CI/CD requirement because the security gate should be enforceable automatically before deployment.</p><p>C is incorrect because Artifact Analysis is the correct service for automatically scanning container images in Artifact Registry and maintaining vulnerability findings, and Binary Authorization can require an attestation. However, manually reviewing every image in the Google Cloud console does not provide an automated CI/CD gate. A human review step can become a bottleneck and can be skipped accidentally.</p><blockquote><strong><em>D is correct because Artifact Analysis can automatically scan new Artifact Registry container images and continuously update vulnerability metadata as new vulnerability information becomes available. The CI/CD process can programmatically evaluate the image’s vulnerability occurrences and create an attestation only if the image meets the organization’s policy, such as having no prohibited critical vulnerabilities. Binary Authorization then enforces deployment policy at GKE admission time and rejects images that do not have the required attestation. This combines automated detection with an enforceable deployment control.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/artifact-analysis/docs/container-scanning-overview">docs.cloud.google.com/artifact-analysis/docs/container-scanning-overview</a></p><p><a href="http://docs.cloud.google.com/artifact-analysis/docs/package-scan-automatic">docs.cloud.google.com/artifact-analysis/docs/package-scan-automatic</a></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/making-attestations">docs.cloud.google.com/binary-authorization/docs/making-attestations</a></p><p><strong>Q10.</strong></p><p><strong>At Fyntrava Storage, your application stores user files in a Cloud Storage bucket that uses a flat object namespace. You want every user to appear to have an individual empty folder-like prefix in Cloud Storage as soon as the user account is created, even before any files are uploaded. What should you do?</strong></p><p><strong>A. Create a zero-byte object whose name is the desired folder prefix followed by a trailing slash (/).</strong></p><p><strong>B. Create an object with the name of the folder prefix, and then immediately delete the object within that prefix.</strong></p><p><strong>C. Create a zero-byte object with the name of the folder prefix and grant the object WRITER access through an access control list.</strong></p><p><strong>D. Create a zero-byte object with the name of the folder prefix and set its Content-Type metadata to CLOUDSTORAGE_FOLDER.</strong></p><blockquote><strong><em>A is correct because a flat-namespace Cloud Storage bucket does not contain real filesystem directories. Folder-like structures are simulated by object-name prefixes containing slash characters. Google Cloud documents that when an empty folder is created in the Google Cloud console for a flat-namespace bucket, Cloud Storage creates a zero-byte placeholder object whose name ends in a slash, such as user123/. Creating the same type of zero-byte trailing-slash object programmatically provides the folder-like placeholder required by the scenario. If the application instead requires real folder resources and filesystem-like semantics, Cloud Storage also supports hierarchical-namespace buckets, but that is a different bucket model from the flat namespace specified here.</em></strong></blockquote><p>B is incorrect because deleting the only object associated with the prefix leaves no stored object that represents the empty simulated folder in a flat namespace. Since the folder is not an independent filesystem object, removing the placeholder means there is nothing for object-listing tools to discover as an empty prefix.</p><p>C is incorrect because omitting the trailing slash means the object is simply an object whose name happens to match the intended folder label; it does not follow the conventional placeholder-object representation used for an empty simulated folder. In addition, object ACL configuration is unrelated to creating folder semantics and can conflict with modern uniform bucket-level access designs.</p><p>D is incorrect because CLOUDSTORAGE_FOLDER is not a special Cloud Storage Content-Type value that creates a folder resource. In a flat namespace, simulated hierarchy comes from object naming and delimiters, not from a custom MIME type. A zero-byte object ending in / is the documented placeholder representation for an empty simulated folder.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/storage/docs/objects">docs.cloud.google.com/storage/docs/objects</a></p><p><a href="http://docs.cloud.google.com/storage/docs/folders-overview">docs.cloud.google.com/storage/docs/folders-overview</a></p><p><strong>Q11.</strong></p><p><strong>At Heliora Platforms, an application running on Google Kubernetes Engine has intermittent performance problems. Your team uses a third-party logging agent and needs one copy of that agent on every eligible GKE node so it can collect node-local logs. What should you do?</strong></p><p><strong>A. Deploy the third-party logging solution as a DaemonSet.</strong></p><p><strong>B. Modify every application container image to include the logging software.</strong></p><p><strong>C. Connect to each GKE node by using SSH and install the logging software manually.</strong></p><p><strong>D. Deploy the third-party logging solution by using Terraform, but run the logging Pods as a Kubernetes Deployment.</strong></p><blockquote><strong><em>A is correct because a Kubernetes DaemonSet is designed to ensure that all, or a selected subset of, cluster nodes run a copy of a Pod. Node-level log collectors and monitoring agents are standard DaemonSet use cases. When GKE adds a new eligible node, the DaemonSet controller automatically schedules the logging Pod on that node; when a node is removed, the corresponding Pod is cleaned up. This provides the one-agent-per-node behavior required by the scenario without modifying application containers or administering nodes manually.</em></strong></blockquote><p>B is incorrect because embedding the logging agent into every application image couples observability tooling to the application lifecycle and can produce multiple copies of the same node-level collector on one node. It also requires rebuilding and redeploying every application whenever the agent changes. A DaemonSet isolates the logging component from application images and ensures one managed copy per eligible node.</p><p>C is incorrect because manually installing software over SSH does not fit the declarative and ephemeral nature of GKE nodes. Nodes can be automatically upgraded, repaired, replaced, or scaled, causing manual changes to disappear and requiring repeated administrator intervention. Direct node modification also creates configuration drift. A DaemonSet automatically follows node lifecycle changes through Kubernetes.</p><p>D is incorrect because Terraform can be used to create Kubernetes resources, but a Deployment manages a desired number of interchangeable Pod replicas and does not guarantee that exactly one logging Pod runs on each node. Multiple replicas could land on one node while another node has none. The required scheduling behavior comes from the DaemonSet controller, regardless of whether the manifest is applied manually, through Terraform, or through another delivery tool.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/daemonset/">kubernetes.io/docs/concepts/workloads/controllers/daemonset/</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/about-logs">docs.cloud.google.com/kubernetes-engine/docs/concepts/about-logs</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/linux-auditd-logging">docs.cloud.google.com/kubernetes-engine/docs/how-to/linux-auditd-logging</a></p><p><strong>Q12.</strong></p><p><strong>At Nimbrixa Systems, you are planning to deploy hundreds of microservices in a Google Kubernetes Engine cluster. You need to secure service-to-service communication with mutual TLS while using a managed Google Cloud service. What should you do?</strong></p><p><strong>A. Use a global external Application Load Balancer with Google-managed certificates to protect communication between all microservices.</strong></p><p><strong>B. Deploy and operate open-source Istio yourself in the GKE cluster, and enable mutual TLS in that self-managed service mesh.</strong></p><p><strong>C. Install cert-manager in the GKE cluster and use it to issue and renew TLS certificates for each microservice.</strong></p><p><strong>D. Enable Cloud Service Mesh with a managed control plane and enforce mutual TLS for service-to-service communication.</strong></p><p>A is incorrect because a global external Application Load Balancer and Google-managed certificates protect traffic entering an application through the load-balancer frontend. They do not automatically establish workload identities or mutual TLS connections for hundreds of east-west service-to-service calls inside the GKE cluster. The requirement is an internal service-mesh security problem rather than an internet-ingress TLS problem.</p><p>B is incorrect because open-source Istio can provide mutual TLS and service-mesh traffic policy, but deploying and operating Istio yourself does not satisfy the requirement to use a managed service. The team would be responsible for installing, upgrading, securing, and maintaining the service-mesh control plane and associated components.</p><p>C is incorrect because cert-manager automates certificate issuance and renewal in Kubernetes, but certificate lifecycle management alone is not equivalent to a managed service mesh. The team would still need to implement workload identity, certificate trust, mTLS policy, proxy configuration, and service-to-service enforcement across all of the microservices.</p><blockquote><strong><em>D is correct because Cloud Service Mesh is Google Cloud’s managed service-mesh offering for GKE and provides workload identity, traffic management, observability, and service-to-service security based on Istio APIs. With mutual TLS enforced, participating workloads authenticate one another and encrypt service-to-service traffic. Using the managed control plane, and where appropriate the managed data plane, reduces operational responsibility compared with operating open-source Istio directly.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/service-mesh/docs/tutorials/mtls">docs.cloud.google.com/service-mesh/docs/tutorials/mtls</a></p><p><a href="http://docs.cloud.google.com/service-mesh/docs/operate-and-maintain/managed-data-plane">docs.cloud.google.com/service-mesh/docs/operate-and-maintain/managed-data-plane</a></p><p><strong>Q13.</strong></p><p><strong>At Elyntara Applications, you recently developed a new application and want to deploy it on Cloud Run without creating a Dockerfile. Your organization requires container images to be stored in Artifact Registry rather than in an unmanaged local-only image store. Which two approaches can build the application without a Dockerfile and store the resulting image in Artifact Registry? Choose two.</strong></p><p><strong>A. Push the application source code itself to an Artifact Registry Docker repository and treat the source archive as the container image.</strong></p><p><strong>B. Submit an unspecified Cloud Build job whose only instruction is to push an image, without configuring a buildpack image target or another build step.</strong></p><p><strong>C. Use the pack CLI with Google Cloud’s buildpacks and run pack build — publish with an Artifact Registry image URL.</strong></p><p><strong>D. Deploy from source by using gcloud run deploy — source, which uses Google Cloud’s buildpacks and Cloud Build when no Dockerfile is present and stores the built image in Artifact Registry.</strong></p><p><strong>E. Include — platform=kubernetes with gcloud run deploy to make Cloud Run build and deploy the application to Kubernetes.</strong></p><p>A is incorrect because an Artifact Registry Docker repository stores OCI or Docker container images, not arbitrary application source code as if that source were already a runnable image. The source must first be converted into a container image by a Dockerfile-based build, buildpacks, Jib, or another supported build system.</p><p>B is incorrect because Cloud Build can absolutely build source without a Dockerfile when it is explicitly configured to use Google Cloud’s buildpacks, for example with gcloud builds submit — pack image=ARTIFACT_REGISTRY_URL. However, the option as written only says to submit a generic build job to push an image and does not specify any step that actually converts the source into a container image. A push operation alone cannot build an image that does not yet exist.</p><blockquote><strong><em>C is correct because Google Cloud’s buildpacks are compatible with the CNCF pack CLI. The pack build command can detect the supported application language, create an OCI container image without a Dockerfile, and the — publish option can publish that image directly to a properly authenticated Artifact Registry Docker repository. This gives the organization explicit control over the repository URL.</em></strong></blockquote><blockquote><strong><em>D is correct because Cloud Run source deployment is the managed one-command option for this use case. When gcloud run deploy — source is used and no Dockerfile is present, Google Cloud’s buildpacks automatically detect the application language and Cloud Build builds a production-ready container image. The built container is stored in an Artifact Registry repository, normally the regional cloud-run-source-deploy repository, and then deployed to Cloud Run. This removes the requirement for the developer to maintain a Dockerfile or run Docker locally.</em></strong></blockquote><p>E is incorrect because current Cloud Run does not use — platform=kubernetes to deploy a Cloud Run service into Kubernetes. Cloud Run is a fully managed platform, and the old distinction represented by a Kubernetes platform flag is not the current deployment model. The flag also does not describe a no-Dockerfile container build workflow.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/deploying-source-code">docs.cloud.google.com/run/docs/deploying-source-code</a></p><p><a href="http://docs.cloud.google.com/run/docs/building/containers">docs.cloud.google.com/run/docs/building/containers</a></p><p><a href="http://docs.cloud.google.com/build/docs/building/build-containers">docs.cloud.google.com/build/docs/building/build-containers</a></p><p><strong>Q14.</strong></p><p><strong>At Myrelith Platforms, your application runs as a container on Google Kubernetes Engine and needs a sensitive credential. You want to keep the secret outside cluster storage, avoid downloadable service account keys, and use the current Google-recommended GKE integration. What should you do?</strong></p><p><strong>A. Create a native Kubernetes Secret and pass its value to the container as an environment variable.</strong></p><p><strong>B. Enable application-layer Kubernetes Secrets encryption with a Cloud Key Management Service key and continue storing the application credential only as a native Kubernetes Secret.</strong></p><p><strong>C. Store the credential in Cloud Key Management Service, download an IAM service account JSON key for the Pod, mount that JSON key as a volume, and have the application use it to retrieve the credential.</strong></p><p><strong>D. Store the credential in Secret Manager, enable Workload Identity Federation for GKE, grant the workload’s Kubernetes ServiceAccount least-privilege access to the secret, and use the Secret Manager add-on for GKE to mount the secret into the Pod.</strong></p><p>A is incorrect because a native Kubernetes Secret is better than embedding a credential in an image, but the sensitive value remains stored in the cluster’s Kubernetes data store and is governed primarily through Kubernetes access controls. The scenario explicitly asks to keep the credential outside cluster storage and use the current Google-managed integration.</p><p>B is incorrect because application-layer Kubernetes Secrets encryption provides stronger encryption at rest for native Kubernetes Secret objects, but the credential is still stored as a Kubernetes Secret in the cluster. This feature is useful when native Secrets are required, but it does not satisfy the requirement to externalize the secret into a dedicated secret-management service.</p><p>C is incorrect because Cloud Key Management Service manages cryptographic keys; Secret Manager is the purpose-built service for storing application credentials such as passwords and API keys. Downloading and mounting a service account JSON key also introduces a long-lived credential that can be exfiltrated and must be rotated, which Google recommends avoiding when Workload Identity Federation for GKE is available.</p><blockquote><strong><em>D is correct because Secret Manager stores the sensitive credential outside the GKE cluster and provides versioning, IAM authorization, auditing, and managed secret lifecycle features. The Secret Manager add-on for GKE can mount Secret Manager values directly into Pods through the managed CSI integration. The add-on authenticates using the Pod’s Workload Identity Federation for GKE identity, so the workload can receive narrowly scoped access to only the required secret without any downloadable service account key. Current GKE versions can also support automatic rotation of mounted Secret Manager values when that capability is enabled.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/secret-manager-managed-csi-component">docs.cloud.google.com/secret-manager/docs/secret-manager-managed-csi-component</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity">docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/tutorials/workload-identity-secrets">docs.cloud.google.com/kubernetes-engine/docs/tutorials/workload-identity-secrets</a></p><p><strong>Q15.</strong></p><p><strong>At Brivanta Retail, you are the lead developer for a retail application running on Cloud Run with Firestore in Datastore mode. The web interface displays available products and allows users to browse the entire catalog. The original implementation retrieves all product entities in one query. As the catalog has grown, Cloud Run instances sometimes exceed their memory limits during busy periods, and those incidents coincide with spikes in Datastore-mode entity reads. You need to prevent memory exhaustion, reduce unnecessary reads, and optimize browsing performance. What should you do?</strong></p><p><strong>A. Modify the product query to paginate by using integer offsets.</strong></p><p><strong>B. Modify the product query to use a limit, but restart every page from the beginning of the result set.</strong></p><p><strong>C. Increase the Cloud Run memory limit and continue retrieving the full product catalog for each browsing request.</strong></p><p><strong>D. Modify the product query to paginate with a reasonable limit and query cursors.</strong></p><p>A is incorrect because Firestore in Datastore mode supports offsets but explicitly recommends avoiding them. An offset prevents skipped entities from being returned to the application, but the database still reads the skipped entities internally. Those skipped reads increase latency and are billable, so later pages become progressively less efficient as the offset grows.</p><p>B is incorrect because a limit is necessary to bound the amount of data returned in each page, but a limit alone does not provide an efficient way to continue through the complete ordered result set. Reissuing the query from the beginning and otherwise trying to derive later pages can repeat reads or require offsets. A cursor records the position after the last returned result and allows the next page to begin there efficiently.</p><p>C is incorrect because adding memory treats only the immediate symptom. The application would still read and hold the entire catalog for each request, causing unnecessary Firestore reads and making memory requirements grow continuously with the product catalog. Eventually the workload could exceed the new memory allocation as well.</p><blockquote><strong><em>D is correct because cursors allow the application to continue a Datastore-mode query from the position reached by the previous page without paying the read and latency overhead associated with skipped offset entities. Combining a cursor with a sensible page limit constrains the number of entities held in each Cloud Run request, lowers peak memory usage, reduces unnecessary reads, and gives users efficient navigation through the complete product catalog.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/datastore/docs/concepts/queries">docs.cloud.google.com/datastore/docs/concepts/queries</a></p><p><a href="http://docs.cloud.google.com/datastore/docs/best-practices">docs.cloud.google.com/datastore/docs/best-practices</a></p><p><a href="http://cloud.google.com/datastore/pricing">cloud.google.com/datastore/pricing</a></p><p><strong>Q16.</strong></p><p><strong>At Myrovexa Applications, you deployed a Java application to Cloud Run. The application must connect to a Cloud SQL database through the database instance’s private IP address because of regulatory requirements. You want to use the current Google-recommended Cloud Run networking approach. How should you configure connectivity?</strong></p><p><strong>A. Configure only the Cloud Run Cloud SQL connection setting and assume that this alone creates a private VPC network path to the Cloud SQL private IP address.</strong></p><p><strong>B. Configure the Cloud SQL instance with private IP and private services access, and configure the Cloud Run service with Direct VPC egress to a VPC network that can reach the Cloud SQL private IP address.</strong></p><p><strong>C. Use the Cloud SQL Java Connector in the application without configuring Cloud Run with access to the VPC network that contains the private Cloud SQL address.</strong></p><p><strong>D. Run a Cloud SQL Auth Proxy process with the application but do not configure a VPC network path from Cloud Run to the Cloud SQL private IP address.</strong></p><p>A is incorrect because Cloud Run’s Cloud SQL integration can simplify authenticated connectivity, but a private Cloud SQL address is reachable only through an appropriate VPC network path. Merely adding a Cloud SQL connection setting does not replace the VPC connectivity required to route TCP traffic to the instance’s private IP.</p><blockquote><strong><em>B is correct because Cloud SQL private IP uses private services access to make the database reachable from a VPC network. Cloud Run Direct VPC egress is the currently recommended method for sending Cloud Run outbound traffic directly into a VPC without operating Serverless VPC Access connector instances. Once the service uses a VPC network that can reach the Cloud SQL private address, the Java application can connect to the database through that internal address and port. Appropriate database authentication and TLS controls should still be configured independently.</em></strong></blockquote><p>C is incorrect because the Cloud SQL Java Connector can handle Cloud SQL authorization and secure connection behavior, and it can be configured for private-IP connections, but it does not create the underlying network route from Cloud Run to a private address. The Cloud Run service still needs Direct VPC egress or another supported VPC connectivity method.</p><p>D is incorrect because the Cloud SQL Auth Proxy can provide authenticated database connectivity, but a proxy process also cannot reach a private IP address without network connectivity to the VPC. The missing requirement remains the VPC path. In addition, direct private-IP connectivity through Direct VPC egress avoids introducing a proxy solely to solve network reachability.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/configuring/connecting-vpc">docs.cloud.google.com/run/docs/configuring/connecting-vpc</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/connect-run">docs.cloud.google.com/sql/docs/mysql/connect-run</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/private-ip">docs.cloud.google.com/sql/docs/mysql/private-ip</a></p><p><strong>Q17.</strong></p><p><strong>At Daxoria Security, a new government regulation requires your application project to send a duplicate of selected application log entries to a separate Google Cloud project that only the security team can access. The original logs should continue to follow their existing retention path. What should you do?</strong></p><p><strong>A. Create a user-defined Cloud Logging log bucket in the security team’s project and create a log sink in the application project whose filter matches the required application logs and whose destination is that cross-project log bucket.</strong></p><p><strong>B. Create a scheduled job that repeatedly reads matching entries from the _Required log bucket and writes copies into the security team’s log bucket.</strong></p><p><strong>C. Reconfigure the system-created _Default sink so that the selected logs are removed from their current destination and sent only to the security team’s project.</strong></p><p><strong>D. Create a scheduled job that copies only System Event audit logs from the application’s _Required bucket into the security team’s project.</strong></p><blockquote><strong><em>A is correct because Cloud Logging sinks are the native mechanism for routing copies of matching log entries to supported destinations, including log buckets located in another Google Cloud project. The sink filter can select only the regulated application logs, while the application’s existing sinks continue to store the original entries. For a cross-project destination, the sink’s writer identity must receive the appropriate permission on the destination log bucket so Logging can write the routed copies.</em></strong></blockquote><p>B is incorrect because periodically reading logs and writing copies requires custom code, scheduling, checkpointing, retry logic, duplicate handling, and additional IAM permissions. Cloud Logging routing performs this operation continuously and is specifically designed for centralized or cross-project log storage.</p><p>C is incorrect because the requirement is to send a duplicate while preserving the original logging behavior. Replacing or repurposing the normal destination can remove the source copy or alter the application’s existing retention and analysis workflow. A separate user-defined sink should be added for the compliance copy instead.</p><p>D is incorrect because the requirement applies to specified application logs, not only to System Event audit logs. System Event logs are a particular Cloud Audit Logs category and would not capture arbitrary application-generated log entries. A filtered sink can select the exact application logs required by regulation.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/routing/overview">docs.cloud.google.com/logging/docs/routing/overview</a></p><p><a href="http://docs.cloud.google.com/logging/docs/buckets">docs.cloud.google.com/logging/docs/buckets</a></p><p><a href="http://docs.cloud.google.com/logging/docs/export/configure_export_v2">docs.cloud.google.com/logging/docs/export/configure_export_v2</a></p><p><strong>Q18.</strong></p><p><strong>At Draxiva Platforms, your team recently deployed an application on Google Kubernetes Engine. You want an automated notification whenever the average container memory-limit utilization is below 20% or above 80% during the selected evaluation period. How should you configure the alerts?</strong></p><p><strong>A. Create a Cloud Run function that queries the Cloud Monitoring API. Trigger it hourly with Cloud Scheduler and have it send an alert when average memory utilization is outside the range.</strong></p><p><strong>B. In Cloud Monitoring, create an alerting policy using GKE container memory-utilization metrics, with conditions for values below 20% or above 80%.</strong></p><p><strong>C. Create a Cloud Run function that runs on a schedule, executes kubectl top against every workload in the cluster, and sends an email when the calculated average is outside the range.</strong></p><p><strong>D. Write a script that queries memory consumption at the node operating-system level and sends an email when the calculated average is outside the range.</strong></p><p>A is incorrect because a custom scheduled function would duplicate functionality already provided by Cloud Monitoring and would evaluate the metric only at the function’s polling frequency. The team would need to implement metric queries, aggregation, state, threshold evaluation, error handling, and notification delivery itself.</p><blockquote><strong><em>B is correct because GKE automatically exposes container memory metrics to Cloud Monitoring when the relevant observability components are enabled. The kubernetes.io/container/memory/limit_utilization metric represents the fraction of a container’s memory limit in use. An alerting policy can aggregate the relevant time series and define threshold conditions for both the upper and lower boundaries, with notification channels receiving incidents when either condition is violated.</em></strong></blockquote><p>C is incorrect because kubectl top is an interactive operational tool rather than a durable monitoring data source for production alerting. Scheduling a function to run it adds custom code and cluster credentials and provides only point-in-time snapshots instead of using Cloud Monitoring’s time-series evaluation.</p><p>D is incorrect because node-level operating-system memory consumption is not the same as container memory-limit utilization. Multiple Pods and system processes share a node, so host memory data cannot accurately determine whether the application’s containers fall below or exceed the requested utilization thresholds.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/monitoring/alerts">docs.cloud.google.com/monitoring/alerts</a></p><p><a href="http://docs.cloud.google.com/monitoring/api/metrics_kubernetes">docs.cloud.google.com/monitoring/api/metrics_kubernetes</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/introduction-monitoring">docs.cloud.google.com/kubernetes-engine/docs/troubleshooting/introduction-monitoring</a></p><p><strong>Q19.</strong></p><p><strong>At Lumetriq Markets, you developed an application that monitors a large number of stock prices. A high volume of Pub/Sub messages updates one large in-memory database. Each message contains a stock symbol, stock price, and update timestamp, and the consumer uses the timestamp to ignore any update older than the currently stored value. You need to minimize duplicate processing while using a high-throughput subscriber. How should you configure the Pub/Sub subscription?</strong></p><p><strong>A. Create a pull subscription with exactly-once delivery enabled and use StreamingPull in the subscriber.</strong></p><p><strong>B. Create a push subscription with both ordering and exactly-once delivery disabled.</strong></p><p><strong>C. Create a push subscription with exactly-once delivery enabled.</strong></p><p><strong>D. Create a pull subscription with both ordering and exactly-once delivery disabled.</strong></p><blockquote><strong><em>A is correct because Pub/Sub exactly-once delivery is available for pull subscriptions and prevents redelivery after a message has been successfully acknowledged. Google recommends StreamingPull through supported client libraries for high-throughput subscribers, including exactly-once subscribers. Because each stock update contains a timestamp and the application applies only a value newer than the currently stored one, the consumer can safely protect the current-state database from out-of-order stale updates while exactly-once delivery reduces duplicate processing caused by subscriber redelivery.</em></strong></blockquote><p>B is incorrect because push subscriptions do not give the subscriber the same direct StreamingPull flow-control model, and with exactly-once disabled the application can receive redeliveries. The question specifically asks to minimize duplicate processing for a high-volume consumer.</p><p>C is incorrect because Pub/Sub exactly-once delivery is supported only for pull subscriptions, including StreamingPull. Push subscriptions cannot be configured with Pub/Sub’s exactly-once delivery feature, so this configuration is not valid.</p><p>D is incorrect because pull is appropriate for high-throughput subscriber control, but leaving exactly-once delivery disabled means the subscriber must expect at-least-once behavior and handle possible redelivery after acknowledgment failures or deadline expiration. The timestamp can prevent a stale price from replacing a newer value, but it does not by itself eliminate duplicated processing work.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/pubsub/docs/exactly-once-delivery">docs.cloud.google.com/pubsub/docs/exactly-once-delivery</a></p><p><a href="http://docs.cloud.google.com/pubsub/docs/subscriber">docs.cloud.google.com/pubsub/docs/subscriber</a></p><p><a href="http://docs.cloud.google.com/pubsub/docs/ordering">docs.cloud.google.com/pubsub/docs/ordering</a></p><p><strong>Q20.</strong></p><p><strong>At Nimbrixa Labs, you are developing an application based on a microservices architecture that includes Cloud Run, Bigtable, and Pub/Sub. You want to create a minimum viable product and make the development, testing, and debugging cycle as fast and inexpensive as possible. What should you do?</strong></p><p><strong>A. Use Cloud Shell Editor and Cloud Shell to deploy the application, and test all functionality by using live Google Cloud resources in the project.</strong></p><p><strong>B. Use local emulators for supported services such as Bigtable and Pub/Sub, test the Cloud Run container locally, and deploy the code to your Google Cloud project after local validation.</strong></p><p><strong>C. Use Cloud Build to create a pipeline with a unit-testing stage and a manual-approval stage, and deploy every development iteration to your Google Cloud project.</strong></p><p><strong>D. Use Cloud Code to develop, deploy, and test all microservice resources against live Google Cloud services, and use Cloud Logging to review the resource logs.</strong></p><p>A is incorrect because Cloud Shell is a convenient managed development environment, but repeatedly deploying every development iteration against live Bigtable, Pub/Sub, and Cloud Run resources introduces network latency, cloud-resource setup, and potential service costs. It does not take advantage of the local emulators and local container-testing workflows provided for rapid development.</p><blockquote><strong><em>B is correct because Bigtable and Pub/Sub both provide local emulators intended for development and testing, allowing most integration work to occur without provisioning or consuming the production services. Cloud Run containers can also be built and exercised locally through supported local development tooling before deployment. This shortens the edit-test-debug loop, keeps tests isolated, and minimizes the cost of creating a minimum viable product. Final integration behavior that depends on production-only features should still be validated in a Google Cloud test environment before release.</em></strong></blockquote><p>C is incorrect because Cloud Build is appropriate for automated CI/CD, but a full pipeline with manual approval for every early development iteration adds latency and cloud build activity when the goal is the fastest and least expensive MVP development loop. CI should be added for repeatable validation, but local emulation is more efficient for frequent developer experimentation.</p><p>D is incorrect because Cloud Code is useful for cloud-native development and local Cloud Run testing, but this option unnecessarily assumes that the microservices should be deployed to live Google Cloud resources and debugged primarily through Cloud Logging. Using local Bigtable and Pub/Sub emulators avoids much of that deployment overhead and service consumption during early development.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigtable/docs/emulator">docs.cloud.google.com/bigtable/docs/emulator</a></p><p><a href="http://docs.cloud.google.com/pubsub/docs/emulator">docs.cloud.google.com/pubsub/docs/emulator</a></p><p><a href="http://docs.cloud.google.com/run/docs/testing/local">docs.cloud.google.com/run/docs/testing/local</a></p><p><strong>Q21.</strong></p><p><strong>At Veltrixa Software, you manage an application deployed to Google Kubernetes Engine clusters across multiple environments. Cloud Build runs automated user-acceptance tests, Artifact Analysis is integrated into the software-supply-chain workflow, and Binary Authorization is enabled in the Google Cloud projects that host your environments. Only container images that successfully pass the required automated UAT tests may be deployed to the production GKE cluster. You have already created a Binary Authorization attestor. What should you do next?</strong></p><p><strong>A. After the UAT phase, sign the attestation with a private key stored as a Kubernetes Secret. Add a cluster-specific Binary Authorization rule in the UAT project’s policy.</strong></p><p><strong>B. After the UAT phase, sign the attestation with a private key stored as a Kubernetes Secret. Add a cluster-specific Binary Authorization rule in the production project’s policy.</strong></p><p><strong>C. After the UAT phase, sign the attestation with an asymmetric signing key stored in Cloud Key Management Service. Add a default Binary Authorization rule in the UAT project’s policy.</strong></p><p><strong>D. After the UAT phase, sign the attestation with an asymmetric signing key stored in Cloud Key Management Service. Add a cluster-specific rule to the production project’s Binary Authorization policy that requires the attestor.</strong></p><p>A is incorrect because a Kubernetes Secret is not the preferred location for the private signing key used to establish trusted software-supply-chain attestations. More importantly, configuring the enforcement rule only in the UAT project’s policy does not prevent an unattested image from being deployed to the production cluster.</p><p>B is incorrect because placing the rule in the production policy is correct, but keeping an attestation-signing private key as a Kubernetes Secret creates unnecessary exposure of a high-value signing credential to cluster access. Cloud KMS provides managed protection, IAM controls, auditing, and cryptographic operations for asymmetric signing keys.</p><p>C is incorrect because using a Cloud KMS asymmetric signing key is the secure choice, but a rule in the UAT project’s Binary Authorization policy governs deployments in that project rather than enforcing the gate on the production GKE cluster. The deployment control must be applied where production admission occurs.</p><blockquote><strong><em>D is correct because the UAT pipeline should create a Binary Authorization attestation only after the image passes the required tests. Cloud KMS securely holds the asymmetric private key used to sign that attestation without exposing the private key as a Kubernetes secret. The production project’s Binary Authorization policy can then contain a cluster-specific admission rule requiring the attestor for the production GKE cluster. Images without the required UAT attestation are blocked at production admission time.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/making-attestations">docs.cloud.google.com/binary-authorization/docs/making-attestations</a></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/configuring-policy-cli">docs.cloud.google.com/binary-authorization/docs/configuring-policy-cli</a></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/creating-attestors-cli">docs.cloud.google.com/binary-authorization/docs/creating-attestors-cli</a></p><p><strong>Q22.</strong></p><p><strong>At Quorvexa Media, your team uses Cloud Storage for a video and image application that was recently migrated to Google Cloud. Following a viral surge, users report application instability that coincides with a tenfold increase in HTTP 429 responses from the Cloud Storage APIs. You need to resolve the errors and make the application more resilient if load increases again. What should you do?</strong></p><p><strong>A. Optimize the application code to reduce unnecessary calls to the Cloud Storage APIs to prevent HTTP 429 errors.</strong></p><p><strong>B. Compress the video and image files to reduce their size and bandwidth usage. Implement a custom throttling mechanism that limits the number of concurrent Cloud Storage API calls.</strong></p><p><strong>C. Migrate all image and video data to Firestore in Native mode. Replace the Cloud Storage APIs in the application code with Firestore APIs.</strong></p><p><strong>D. Implement a retry strategy with exponential backoff for requests that receive HTTP 429 responses.</strong></p><p>A is incorrect because reducing unnecessary requests is a useful optimization and can lower pressure on Cloud Storage, but it does not provide the required behavior when legitimate traffic still produces transient HTTP 429 responses. The application needs a retry policy that allows Cloud Storage time to scale and recover rather than failing the user request immediately.</p><p>B is incorrect because compressing media can reduce object size and network transfer, but Cloud Storage request-rate behavior is driven by request operations rather than only by file size. A custom concurrency limiter can help shape traffic in some designs, but a fixed custom throttle is not the primary Google-recommended response to retryable 429 errors and can unnecessarily reduce throughput when the service is healthy.</p><p>C is incorrect because Firestore is a document database and is not a replacement for Cloud Storage’s durable object-storage model for large image and video files. Migrating binary media to Firestore would require substantial application changes and would use a service that is poorly matched to the workload.</p><blockquote><strong><em>D is correct because Cloud Storage explicitly recommends retrying HTTP 429 responses by using exponential backoff. Increasing the delay between repeated attempts reduces pressure on the service and avoids synchronized retry storms while Cloud Storage scales to the higher request rate. For long-term stability, the team should also ramp traffic gradually where practical, distribute request load appropriately, and remove unnecessary operations, but exponential backoff is the required application behavior among these choices.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/storage/docs/request-rate">docs.cloud.google.com/storage/docs/request-rate</a></p><p><a href="http://docs.cloud.google.com/storage/docs/retry-strategy">docs.cloud.google.com/storage/docs/retry-strategy</a></p><p><strong>Q23.</strong></p><p><strong>At Orvessa Software, you need to deploy a new feature to production on Cloud Run. Your SRE team requires gradual deployments so that defects in a new revision affect only a small percentage of traffic initially. You want to configure the rollout with minimal effort. What should you do?</strong></p><p><strong>A. Configure the application’s frontend load balancer to toggle manually between the old and new Cloud Run revisions.</strong></p><p><strong>B. Modify the application code so that it sends a small percentage of users to the newly deployed revision.</strong></p><p><strong>C. Deploy the new revision without immediately serving all traffic, configure it to receive a small percentage of traffic, monitor for errors, and gradually increase its traffic percentage when the revision is healthy.</strong></p><p><strong>D. Deploy the new revision so that it immediately receives all traffic. If errors occur, roll back to the previous revision and repeat until the deployment is verified.</strong></p><p>A is incorrect because Cloud Run already provides revision-aware traffic management. Adding manual revision-switching logic to a frontend load balancer creates unnecessary configuration and operational steps for behavior the Cloud Run service can manage directly.</p><p>B is incorrect because implementing traffic splitting in application code couples release-management behavior to the application and requires extra code, testing, and state. Cloud Run supports native percentage-based traffic allocation between revisions, so a custom routing implementation is unnecessary.</p><blockquote><strong><em>C is correct because Cloud Run supports deploying a revision with no traffic initially and then splitting traffic across old and new revisions by percentage. The team can send a small share of requests to the new revision, observe errors and performance, then increase the percentage incrementally without redeploying. If a problem occurs, traffic can be returned to the previous revision quickly. This is the built-in Cloud Run gradual-rollout workflow and requires minimal additional infrastructure.</em></strong></blockquote><p>D is incorrect because immediately routing 100% of production traffic to the new revision maximizes the potential blast radius of a defect. Although Cloud Run supports fast rollback, the requirement is to reduce the probability of large downtime by limiting initial exposure, which requires gradual traffic migration rather than an all-at-once deployment.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/rollouts-rollbacks-traffic-migration">docs.cloud.google.com/run/docs/rollouts-rollbacks-traffic-migration</a></p><p><a href="http://docs.cloud.google.com/run/docs/managing/revisions">docs.cloud.google.com/run/docs/managing/revisions</a></p><p><strong>Q24.</strong></p><p><strong>At Nimbrixa Environmental, you are developing an air-quality monitoring platform for a large city. Thousands of devices send and receive small REST payloads every minute by using curl. The backend runs on Compute Engine in a single Google Cloud region and uses Premium Tier networking. You need to expose the backend while minimizing average client Time to First Byte latency. How should you build the service?</strong></p><p><strong>A. Run the backends on individual Compute Engine VMs. Create a weighted round-robin routing policy in Cloud DNS, and configure the devices to use that DNS name.</strong></p><p><strong>B. Run the backends on individual Compute Engine VMs. Create a round-robin routing policy in Cloud DNS for the VM addresses, and configure the devices to use that DNS name.</strong></p><p><strong>C. Run the backends in a managed instance group. Create an external passthrough Network Load Balancer in front of the managed instance group, and configure the devices to connect to that load balancer.</strong></p><p><strong>D. Run the backends in a managed instance group. Create a global external Application Load Balancer in front of the managed instance group, and configure the devices to connect to that load balancer.</strong></p><p>A is incorrect because DNS-based weighted routing returns endpoint information to clients but does not provide the global Google Front End termination and connection optimizations of an Application Load Balancer. Client DNS caching can also make traffic distribution less responsive to backend conditions. The application is HTTP-based, so an HTTP-aware global load balancer is a better fit.</p><p>B is incorrect because simple DNS round-robin has the same limitations as weighted DNS routing and also lacks application-layer health-aware proxying, connection reuse, and Google’s globally distributed frontend infrastructure. It does not provide the strongest latency characteristics for a globally or widely distributed set of HTTP clients reaching one regional backend.</p><p>C is incorrect because an external passthrough Network Load Balancer operates at Layer 4 and preserves direct client-to-backend transport characteristics. It is appropriate when applications require pass-through TCP or UDP behavior, but it does not provide the HTTP-aware edge proxying and global frontend behavior designed for minimizing HTTP request latency.</p><blockquote><strong><em>D is correct because a global external Application Load Balancer accepts HTTP or HTTPS requests at Google’s distributed edge and carries traffic across Google’s Premium Tier network toward the regional managed instance group. Google documents this architecture as the appropriate global HTTP load-balancing model and uses curl Time to First Byte when evaluating user-perceived latency. The managed instance group also provides scalable and health-checked backend capacity rather than exposing individual VMs directly.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/application-load-balancer">docs.cloud.google.com/load-balancing/docs/application-load-balancer</a></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/tutorials/optimize-app-latency">docs.cloud.google.com/load-balancing/docs/tutorials/optimize-app-latency</a></p><p><strong>Q25.</strong></p><p><strong>At Brivanta Forums, you are developing a discussion portal on Cloud Run. Incoming requests pass through several microservices before a response is returned, and some services call databases. You need to run a load test and identify which part of the distributed request path becomes a bottleneck under load. You want to follow Google-recommended practices. What should you do?</strong></p><p><strong>A. Modify every application response to contain a time series showing elapsed time for each service, and use Cloud Logging observability analytics to create a heatmap.</strong></p><p><strong>B. Use Cloud Trace for the load-test requests, instrumenting downstream operations where needed, and review request and span timings in Cloud Trace.</strong></p><p><strong>C. Expose one latency metric per service for every request, configure Google Cloud Managed Service for Prometheus, and use only those aggregated metrics to locate the bottleneck.</strong></p><p><strong>D. Add log statements that record elapsed time in every service and export the logs to BigQuery for manual correlation.</strong></p><p>A is incorrect because adding an internal timing time series to every user response couples diagnostic information to the public application contract and adds unnecessary response payload. Cloud Logging analytics can analyze log data, but distributed tracing is purpose-built to represent the hierarchical timing of one request across multiple services.</p><blockquote><strong><em>B is correct because Cloud Trace tracks end-to-end request latency and represents individual operations as spans. Cloud Run automatically creates trace context for incoming requests, and application instrumentation such as OpenTelemetry can add spans for service calls, database operations, and business logic. During a load test, the team can inspect waterfall-style traces and latency distributions to identify which service or downstream dependency accounts for the delay.</em></strong></blockquote><p>C is incorrect because Prometheus latency metrics are valuable for aggregate service-level trends, alerting, and dashboards, but aggregated metrics do not naturally reconstruct one distributed request through all of its service calls. They can show that a service is slow, but Cloud Trace more directly identifies the specific span causing end-to-end latency.</p><p>D is incorrect because logs can be manually correlated when they contain consistent trace identifiers and timing information, but building a custom BigQuery correlation workflow is more complex than using distributed tracing. It also requires the team to implement its own request hierarchy and timing analysis.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/trace/docs/overview">docs.cloud.google.com/trace/docs/overview</a></p><p><a href="http://docs.cloud.google.com/run/docs/trace">docs.cloud.google.com/run/docs/trace</a></p><p><a href="http://docs.cloud.google.com/trace/docs/setup">docs.cloud.google.com/trace/docs/setup</a></p><p><strong>Q26.</strong></p><p><strong>At Talveriq Data, you are using a current supported Python 3 runtime to develop an API that stores application data in Cloud SQL. You need to perform database CRUD operations securely and reliably with minimal connection-management effort. What should you do?</strong></p><p><strong>A. Use Managed Service for Apache Airflow to manage the connection from the Python application to Cloud SQL, and grant the application service account workflow-worker permissions.</strong></p><p><strong>B. Use the Cloud SQL Admin API as the application’s SQL query interface, and grant the application service account cloudsql.instances.login.</strong></p><p><strong>C. Use the Cloud SQL Python Connector together with the appropriate Python database driver, and grant the application identity the Cloud SQL Client role containing the Cloud SQL instance connection permissions.</strong></p><p><strong>D. Use a Cloud SQL emulator from Cloud Shell to connect to the production database, and grant the user cloudsql.instances.login.</strong></p><p>A is incorrect because Managed Service for Apache Airflow is a workflow-orchestration service and is not a connection layer for application CRUD requests to Cloud SQL. Adding an Airflow environment would create unnecessary infrastructure and latency for a normal application database connection.</p><p>B is incorrect because the Cloud SQL Admin API manages Cloud SQL resources such as instances, databases, backups, and configuration. It is not the SQL data plane used to execute application INSERT, SELECT, UPDATE, and DELETE statements. The application must connect with a normal database driver through an appropriate Cloud SQL connection mechanism.</p><blockquote><strong><em>C is correct because the Cloud SQL Python Connector is a supported language connector that works with standard Python database drivers and provides secure Cloud SQL connection handling without requiring the developer to manage SSL certificates or authorized public IP lists manually. It uses Application Default Credentials for connection authorization. The Cloud SQL Client role provides cloudsql.instances.connect and cloudsql.instances.get, which are the core permissions needed by connectors to establish Cloud SQL connectivity. Database-level users and privileges still separately determine which SQL CRUD operations the application may perform.</em></strong></blockquote><p>D is incorrect because Cloud SQL does not provide a local emulator that transparently substitutes for a production database in this manner. Cloud Shell can be used as an administrative client environment, but the production API should connect through a supported database driver and Cloud SQL connector from its actual runtime.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/connect-connectors">docs.cloud.google.com/sql/docs/mysql/connect-connectors</a></p><p><a href="http://docs.cloud.google.com/iam/docs/roles-permissions/cloudsql">docs.cloud.google.com/iam/docs/roles-permissions/cloudsql</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/roles-and-permissions">docs.cloud.google.com/sql/docs/mysql/roles-and-permissions</a></p><p><strong>Q27.</strong></p><p><strong>At Kelvanta Security, internal applications run on Cloud Run and use Secret Manager to obtain credentials for internal databases. Each running application instance can cache a credential for up to 15 minutes. You need to rotate the database credentials without causing application downtime. Which approach best supports a zero-downtime rotation among the choices below?</strong></p><p><strong>A. Introduce a new database username and password as a new credential while keeping the previous username and password valid for at least the 15-minute cache-overlap period. Store the new credential as a new secret version, allow applications to resolve the updated credential, and retire the old credential only after all caches can have expired.</strong></p><p><strong>B. Make applications resolve the latest secret only when an instance starts, and keep using that credential for the full lifetime of the instance.</strong></p><p><strong>C. Change only the password for the existing database username, make the new password the latest secret value immediately, and let applications cache the value for up to 15 minutes.</strong></p><p><strong>D. Store the new password as a new secret version, change the application code to reference that version, and require a maintenance-window deployment before the credential can be used.</strong></p><blockquote><strong><em>A is correct because zero-downtime secret rotation requires an overlap period in which both the old and new credentials remain valid while running workloads migrate to the new value. Google Cloud’s Secret Manager rotation guidance explicitly recommends creating the new credential in the external system without invalidating the credential already used by running workloads, then adding a new secret version, rolling it out, and cleaning up the old credential only after applications have moved away from it. Because instances can cache credentials for 15 minutes, keeping the old username-password pair valid for at least that overlap prevents cached instances from losing database access. In a production implementation, Google also recommends pinning and gradually rolling out specific secret version numbers rather than relying blindly on the latest alias.</em></strong></blockquote><p>B is incorrect because Cloud Run instances can remain alive for different lengths of time and new instances can appear during scaling. If the credential is fetched only at startup and retained indefinitely, there is no bounded period after which all instances are guaranteed to move to the rotated credential. That complicates safe retirement of the old database credential.</p><p>C is incorrect because changing the password for the same database identity usually invalidates the old password immediately. Cloud Run instances that still have the previous password in their 15-minute caches can then fail authentication until they refresh the secret. The cache window therefore turns an immediate single-password replacement into a potential availability incident.</p><p>D is incorrect because version pinning and controlled rollout are good Secret Manager practices, but requiring a maintenance window is unnecessary when the application and database support overlapping credentials and a gradual rollout. The objective is explicitly to avoid downtime, so the stronger strategy is to keep both credentials valid temporarily, migrate workloads, verify success, and then disable the previous credential.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/rotation-recommendations">docs.cloud.google.com/secret-manager/docs/rotation-recommendations</a></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/best-practices">docs.cloud.google.com/secret-manager/docs/best-practices</a></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/secret-rotation">docs.cloud.google.com/secret-manager/docs/secret-rotation</a></p><p><strong>Q28.</strong></p><p><strong>At Pyralith Services, three teams are developing an ecommerce application in the same Google Cloud project. Team A builds REST APIs that expose core application functionality. Teams B and C call those APIs from downstream processes running as Cloud Run services. You need to expose the APIs securely while minimizing credential-management overhead. How should you design the solution?</strong></p><p><strong>A. Team A uses service accounts to authorize API Gateway. Teams B and C each create a service account, export a private key for that service account, and store the key in Secret Manager. Teams B and C use the key to create authentication tokens for API Gateway.</strong></p><p><strong>B. Team A uses Apigee hybrid to create one API key and shares that key with Teams B and C. Teams B and C store the API key in Secret Manager and use the shared key to access the API.</strong></p><p><strong>C. Team A uses an API key to authorize API Gateway and shares the same API key with Teams B and C. Teams B and C store the API key in Secret Manager and use it to access the API endpoint.</strong></p><p><strong>D. Team A configures API Gateway to authenticate the calling service identities. Teams B and C each use a dedicated service account attached to their Cloud Run service and use short-lived Google-signed identity credentials from that workload identity when calling the API.</strong></p><p>A is incorrect because service-account authentication can be used for service-to-service API calls, but exporting service account private keys creates long-lived credentials that must be stored, rotated, audited, and protected from exfiltration. Cloud Run already supports attached service identities and can obtain short-lived credentials without downloading a private key, so the key-based design adds unnecessary security risk and operational overhead.</p><p>B is incorrect because one shared API key does not provide strong service identity. If the key is copied or leaked, it can be used outside the intended Cloud Run workloads, and API Gateway or Apigee cannot distinguish Team B from Team C based solely on the same shared secret. Apigee hybrid also introduces substantially more platform-management complexity than the scenario requires.</p><p>C is incorrect because API keys are useful for identifying calling applications for quota, monitoring, and basic access-management scenarios, but they are not a strong replacement for workload authentication. A shared key can be copied and does not provide the same cryptographic principal identity and short-lived credential model as service accounts.</p><blockquote><strong><em>D is correct because each Cloud Run service can run with a dedicated service account as its service identity and obtain short-lived Google-signed identity credentials without a downloadable key. API Gateway can validate JWT-based authentication according to its API configuration, allowing Team A to authenticate callers before forwarding requests to the backend. Dedicated workload identities also let the organization distinguish Teams B and C and apply least-privilege access without maintaining private key files in Secret Manager.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/api-gateway/docs/authenticating-users-jwt">docs.cloud.google.com/api-gateway/docs/authenticating-users-jwt</a></p><p><a href="http://docs.cloud.google.com/api-gateway/docs/authenticate-service-account">docs.cloud.google.com/api-gateway/docs/authenticate-service-account</a></p><p><a href="http://docs.cloud.google.com/run/docs/authenticating/service-to-service">docs.cloud.google.com/run/docs/authenticating/service-to-service</a></p><p><strong>Q29.</strong></p><p><strong>At Ceralith Software, you maintain a CI/CD pipeline for an application running on GKE. Cloud Build creates container images and stores them in Artifact Registry. Your pipeline currently deploys the latest tag. You recently needed to roll back a deployment after 24 hours, but the latest tag had already been moved to a newer image, making the rollback difficult. You need to prevent this problem efficiently. What should you do?</strong></p><p><strong>A. Rebuild the container image separately for every environment and tag it with the environment name.</strong></p><p><strong>B. Build one container image for each new application version and assign it a unique version identifier rather than relying on latest; where appropriate, enforce immutable tags or deploy by image digest.</strong></p><p><strong>C. Implement Helm charts to manage the Kubernetes deployments while continuing to use the latest image tag.</strong></p><p><strong>D. Create a separate Artifact Registry repository for every application version.</strong></p><p>A is incorrect because rebuilding an image for each environment means development, QA, staging, and production are not necessarily running the exact same artifact. Even when the source revision is the same, build dependencies or environment changes can produce different image digests, weakening reproducibility and making release promotion harder to audit.</p><blockquote><strong><em>B is correct because each release should identify an immutable artifact. Artifact Registry stores each container version by a unique digest, and the organization can additionally use unique human-readable version tags such as v2.4.1. Artifact Registry supports immutable tags so a version tag cannot later be moved to another digest. GKE deployments can also reference the digest directly, which guarantees that a rollback selects the exact previously tested image even if other tags have changed.</em></strong></blockquote><p>C is incorrect because Helm can package and parameterize Kubernetes manifests, but it does not make the latest tag immutable. If the chart still references latest and that tag is moved to another digest, the same rollback ambiguity remains.</p><p>D is incorrect because Artifact Registry repositories can store many versions of the same image efficiently. Creating an entire repository for each version adds unnecessary administration, IAM, cleanup, and naming complexity. Unique immutable tags or image digests provide version identity within one repository.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/docker/names">docs.cloud.google.com/artifact-registry/docs/docker/names</a></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/repositories/update-repo-settings">docs.cloud.google.com/artifact-registry/docs/repositories/update-repo-settings</a></p><p><a href="http://docs.cloud.google.com/build/docs/deploying-builds/deploy-gke">docs.cloud.google.com/build/docs/deploying-builds/deploy-gke</a></p><p><strong>Q30.</strong></p><p><strong>At Calveriq Tickets, customers use your website to purchase tickets for local concert events. When a customer selects a seat, the website must provide immediate confirmation only after that seat has actually been reserved. How should you design the ticket-ordering process?</strong></p><p><strong>A. Upload the seat-reservation request to a Cloud Storage bucket, which triggers an event to the backend service that processes the seat reservation.</strong></p><p><strong>B. Submit the seat-reservation request in an HTTP POST request to an Application Load Balancer, and configure the load balancer to route the request to the backend service that processes the reservation and returns the result.</strong></p><p><strong>C. Add the seat-reservation request to a Cloud Tasks queue, which triggers Workflows to process the seat reservation.</strong></p><p><strong>D. Publish the seat-reservation request to a Pub/Sub topic, and configure the backend service to subscribe to the topic to process the reservation.</strong></p><p>A is incorrect because Cloud Storage events are asynchronous. Successfully writing a request object proves only that the request was stored, not that the selected seat was successfully reserved. The browser would require an additional polling or callback mechanism before it could provide authoritative confirmation.</p><blockquote><strong><em>B is correct because seat selection is a synchronous transactional operation: the customer needs to know whether the reservation succeeded before receiving confirmation. An HTTP POST routed through an Application Load Balancer to the reservation backend allows that backend to perform the availability check and transactional reservation and then return success or failure in the same request-response interaction. The backend must use appropriate database concurrency controls to prevent two users from successfully reserving the same seat.</em></strong></blockquote><p>C is incorrect because Cloud Tasks provides asynchronous task delivery. Enqueuing the reservation request confirms that the task was accepted, but it does not confirm that the seat was actually available and reserved. Workflows can orchestrate asynchronous processing, but the user still needs a separate mechanism to receive the final outcome.</p><p>D is incorrect because Pub/Sub provides durable asynchronous messaging and is excellent for decoupled background processing, but publishing a message does not return the result of the reservation transaction. The website could not immediately tell the customer whether the selected seat had been successfully acquired without additional correlation and response logic.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/application-load-balancer">docs.cloud.google.com/load-balancing/docs/application-load-balancer</a></p><p><a href="http://docs.cloud.google.com/tasks/docs/comp-pub-sub">docs.cloud.google.com/tasks/docs/comp-pub-sub</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e06de2cfd18f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Digital Leader Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-digital-leader-practice-questions-e8c40223d977?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8c40223d977</guid>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 12:02:35 GMT</pubDate>
            <atom:updated>2026-10-07T12:02:35.349Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dcYYtXoiK4TGK-eXyu1zzQ.jpeg" /><figcaption>KloudExams GCP Google Cloud Digital Leader Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Digital Leader certification</strong>? Knowing what cloud services do is a starting point — understanding how they support business goals and digital transformation is what helps you make informed decisions.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise connecting Google Cloud capabilities with specific business needs. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through official Google Cloud learning resources and real-world business examples.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>At the fictional company Kelmora Technologies, the operating systems of some virtual machines may contain security vulnerabilities. How can the organization most effectively identify all affected virtual machines that do not have the latest relevant security updates?</strong></p><p><strong>A. View Security Command Center to identify virtual machines with operating-system vulnerability findings.</strong></p><p><strong>B. View Compliance Reports Manager to identify and download a recent PCI DSS audit report.</strong></p><p><strong>C. View Security Command Center to identify virtual machines that were started more than two weeks ago.</strong></p><p><strong>D. View Compliance Reports Manager to identify and download a recent SOC 1 audit report.</strong></p><blockquote><strong><em>A is correct because Security Command Center can receive operating-system vulnerability findings from VM Manager vulnerability reports. These findings identify vulnerabilities, including CVEs, in installed operating-system packages on supported Compute Engine VMs and provide organization-level visibility when the relevant capabilities are enabled. That directly addresses the need to find VMs whose operating systems require security remediation or patching.</em></strong></blockquote><p>B is incorrect because Compliance Reports Manager provides access to compliance documentation such as PCI DSS reports, certifications, and audit materials about Google Cloud’s compliance posture. A PCI DSS audit report does not inspect the patch level of the organization’s individual VM operating systems and therefore cannot identify which VMs need operating-system security updates.</p><p>C is incorrect because the age or start time of a VM is not a reliable indicator of patch status. A VM started yesterday could have been created from an outdated image or could contain vulnerable packages, while a VM running for months could have been continuously patched. Vulnerability and patch information must be based on the software actually installed, not simply on how long the VM has been running.</p><p>D is incorrect because a SOC 1 report evaluates controls relevant to financial reporting for the service organization. It is useful for assurance and compliance purposes, but it does not inventory the customer’s Compute Engine VMs or determine which guest operating systems are missing security updates.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/security-command-center/docs/concepts-vulnerabilities-findings">docs.cloud.google.com/security-command-center/docs/concepts-vulnerabilities-findings</a></p><p><a href="http://cloud.google.com/security/compliance/compliance-reports-manager/">cloud.google.com/security/compliance/compliance-reports-manager/</a></p><p><strong>Q2.</strong></p><p><strong>At the fictional company Marovex Global, your multinational organization has servers running mission-critical workloads on premises around the world. You want to manage these workloads consistently and centrally, and you want to stop managing the underlying physical infrastructure. What should the organization do?</strong></p><p><strong>A. Migrate the workloads to a public cloud.</strong></p><p><strong>B. Migrate the workloads to a central office building.</strong></p><p><strong>C. Migrate the workloads to multiple local colocation facilities.</strong></p><p><strong>D. Migrate the workloads to multiple local private clouds.</strong></p><blockquote><strong><em>A is correct because a public cloud transfers responsibility for owning, operating, refreshing, and physically maintaining the underlying data-center infrastructure to the cloud service provider while giving the customer centralized management interfaces, APIs, policies, and automation. A global public cloud also provides regions around the world, allowing workloads to remain geographically distributed while being governed through a consistent cloud operating model.</em></strong></blockquote><p>B is incorrect because moving global workloads into one central office does not remove infrastructure-management responsibility. The organization would still own or operate servers, power, cooling, networking, hardware lifecycle, and physical security, while also introducing geographic concentration and a potentially serious single-site availability risk.</p><p>C is incorrect because colocation facilities provide space, power, cooling, and connectivity, but the organization generally continues to own or manage the servers and much of the infrastructure stack. Using multiple local colocation sites also preserves the operational complexity of managing distributed physical environments instead of eliminating it.</p><p>D is incorrect because multiple private clouds can provide cloud-like automation and control but normally leave the organization responsible for substantially more of the underlying infrastructure and lifecycle management than a public-cloud model. Operating several local private clouds would also make the goal of simplified, centralized, consistent management harder to achieve.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/learn/what-is-cloud-computing">cloud.google.com/learn/what-is-cloud-computing</a></p><p><a href="http://cloud.google.com/learn/what-is-a-cloud-service-provider">cloud.google.com/learn/what-is-a-cloud-service-provider</a></p><p><strong>Q3.</strong></p><p><strong>At the fictional company Nexoryn Digital, the organization is increasing its software release velocity, but upgrades to a VM-based application take a long time because rolling updates must wait for operating systems to boot. You need to make application deployments faster. What should the organization do?</strong></p><p><strong>A. Migrate the VMs to the cloud and add more resources to them.</strong></p><p><strong>B. Convert the applications into containers.</strong></p><p><strong>C. Increase the resources allocated to the VMs.</strong></p><p><strong>D. Automate the VM upgrade rollouts.</strong></p><p>A is incorrect because moving the same VM-based architecture to the cloud and increasing VM resources does not remove the fundamental delay caused by booting a complete guest operating system for each replacement VM. The bottleneck described in the question is the heavyweight VM deployment unit, not insufficient CPU or memory.</p><blockquote><strong><em>B is correct because containers package the application and its dependencies while sharing the host operating-system kernel instead of booting a complete guest OS for every application instance. They are therefore much lighter than VMs and generally start much faster. Containerizing the application directly addresses the OS boot-time delay and supports faster, more repeatable deployments and rolling updates.</em></strong></blockquote><p>C is incorrect because assigning additional CPU or memory to each VM might improve application runtime performance, but it does not fundamentally eliminate the need to boot a guest operating system during VM replacement. The deployment process would remain VM-based and would retain the boot-time overhead described in the question.</p><p>D is incorrect because automation can reduce manual effort and make rolling upgrades more consistent, but it cannot remove the time required for each VM operating system to boot. If boot time is the primary limiting factor, changing the application packaging and deployment unit to containers addresses the root cause more directly than automating the existing VM process.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/discover/containers-vs-vms">cloud.google.com/discover/containers-vs-vms</a></p><p><a href="http://cloud.google.com/learn/what-are-containers">cloud.google.com/learn/what-are-containers</a></p><p><strong>Q4.</strong></p><p><strong>At the fictional company Avenquill Streams, the organization needs to build streaming data pipelines. It does not want to manage the individual servers that perform the data processing. Instead, it wants a managed service that automatically scales with the amount of data to be processed. Which Google Cloud product or feature should the organization choose?</strong></p><p><strong>A. Pub/Sub</strong></p><p><strong>B. Dataflow</strong></p><p><strong>C. Knowledge Catalog</strong></p><p><strong>D. BigQuery data preparation</strong></p><p>A is incorrect because Pub/Sub is a managed asynchronous messaging service used to ingest and distribute event streams between producers and consumers. It is commonly used as a source for streaming pipelines, but it does not itself perform the general-purpose transformations, windowing, aggregation, and processing logic required by the question.</p><blockquote><strong><em>B is correct because Dataflow is Google Cloud’s fully managed service for batch and streaming data processing. Dataflow executes Apache Beam pipelines on managed worker infrastructure and can automatically scale the worker resources according to the amount of data being processed. This allows the organization to focus on pipeline logic instead of provisioning and managing individual processing servers.</em></strong></blockquote><p>C is incorrect because Knowledge Catalog is Google Cloud’s data discovery, metadata, and governance service. It helps users discover and understand data assets and their business context, but it is not a streaming data-processing engine and does not run autoscaling transformation pipelines.</p><p>D is incorrect because BigQuery data preparation helps users clean, transform, enrich, and prepare data in BigQuery with AI-assisted suggestions and scheduled workflows. It is useful for data preparation, but it is not the general-purpose continuous streaming pipeline service described in the question. Dataflow is purpose-built for scalable streaming processing.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/dataflow/docs/overview">docs.cloud.google.com/dataflow/docs/overview</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines">docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/data-prep-introduction">docs.cloud.google.com/bigquery/docs/data-prep-introduction</a></p><p><strong>Q5.</strong></p><p><strong>Avenquill Analytics, a fictional company, wants to migrate its data management solution to Google Cloud. It needs a database that can scale compute capacity up or down, support relational SQL and ACID transactions at scale, and allow consistent reads of earlier versions of data within the configured version-retention window. Which Google Cloud product or service should it use?</strong></p><p><strong>A. BigQuery</strong></p><p><strong>B. Bigtable</strong></p><p><strong>C. Pub/Sub</strong></p><p><strong>D. Spanner</strong></p><p>A is incorrect because BigQuery is a serverless analytical data warehouse optimized for large-scale analytics. BigQuery does support SQL and multi-statement transactions, but it is primarily designed for analytical workloads rather than serving as a horizontally scalable operational relational database for transactional application traffic. The requirement for relational ACID transactions at operational scale points more directly to Spanner.</p><p>B is incorrect because Bigtable is a massively scalable NoSQL wide-column database designed for high-throughput, low-latency key-value and time-series workloads. It does not provide the relational SQL model and general multi-row, multi-table ACID transaction semantics that are central to this question.</p><p>C is incorrect because Pub/Sub is a messaging and event-ingestion service used to decouple systems and distribute event streams. It is not a relational database, does not execute transactional SQL queries over stored historical records, and therefore cannot satisfy the data-management requirements.</p><blockquote><strong><em>D is correct because Spanner is a fully managed, horizontally scalable database that provides relational SQL support and strongly consistent ACID transactions at global scale. It can scale compute capacity as workload needs change and supports reads at earlier timestamps through its timestamp-bound and version-retention capabilities. These characteristics make Spanner the best match for a transactional SQL database that must operate at very large scale while supporting consistent point-in-time reads.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs">docs.cloud.google.com/spanner/docs</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/transactions">docs.cloud.google.com/spanner/docs/transactions</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/timestamp-bounds">docs.cloud.google.com/spanner/docs/timestamp-bounds</a></p><p><strong>Q6.</strong></p><p><strong>At the fictional company Brivexa Imaging, the organization needs to categorize objects in a large group of static images by using machine learning. Which Google Cloud product or service should the organization use?</strong></p><p><strong>A. BigQuery ML</strong></p><p><strong>B. Video Intelligence API</strong></p><p><strong>C. Cloud Vision API</strong></p><p><strong>D. AutoML tabular training in Gemini Enterprise Agent Platform</strong></p><p>A is incorrect because BigQuery ML lets users create and use machine learning models with SQL over data in BigQuery. It is well suited to analytical prediction problems involving tabular and warehouse data, but it is not the direct pre-trained image-analysis service for categorizing objects in static images.</p><p>B is incorrect because the Video Intelligence API analyzes video content rather than a collection of static images. Video-oriented capabilities such as shot analysis, labels, and other video annotations do not make it the appropriate service for this image-only requirement.</p><blockquote><strong><em>C is correct because the Cloud Vision API provides pre-trained machine learning capabilities for analyzing images. Its label-detection functionality can identify and categorize objects, concepts, activities, and other content in static images without requiring the organization to train its own custom model.</em></strong></blockquote><p>D is incorrect because AutoML tabular training in Gemini Enterprise Agent Platform is designed for tabular classification, regression, and forecasting problems. The input in this question consists of images rather than rows and columns of tabular features, so a tabular AutoML workflow is not appropriate.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/vision/docs/detect-labels-image-api">docs.cloud.google.com/vision/docs/detect-labels-image-api</a></p><p><a href="http://docs.cloud.google.com/vision/docs">docs.cloud.google.com/vision/docs</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/automl-training-overview">docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/automl-training-overview</a></p><p><strong>Q7.</strong></p><p><strong>At the fictional company Ceralune Analytics, the organization needs to analyze data to gather insights into its daily operations. It wants a data analytics warehouse where it can pay for the data it stores and the queries it performs. Which Google Cloud product should the organization choose?</strong></p><p><strong>A. Cloud SQL</strong></p><p><strong>B. Managed Service for Apache Spark</strong></p><p><strong>C. Spanner</strong></p><p><strong>D. BigQuery</strong></p><p>A is incorrect because Cloud SQL is a managed relational database service for MySQL, PostgreSQL, and SQL Server. It is designed primarily for operational and transactional application workloads rather than as a serverless enterprise data warehouse optimized for large-scale analytics.</p><p>B is incorrect because Managed Service for Apache Spark is a managed platform for Spark and related open-source data-processing workloads. It is useful for batch, streaming, and data-engineering jobs, but it is not the Google Cloud analytical data warehouse described in the question.</p><p>C is incorrect because Spanner is a globally scalable, strongly consistent operational database for mission-critical transactional workloads. Although Spanner supports SQL and can integrate with analytical systems, it is not the usage-based analytical data warehouse intended for large-scale ad-hoc analysis.</p><blockquote><strong><em>D is correct because BigQuery is Google Cloud’s fully managed, serverless analytical data platform and data warehouse. It supports consumption-based pricing models in which organizations can pay for stored data and query processing without provisioning database servers. BigQuery is specifically designed for large-scale analytical queries and business insights.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/introduction">docs.cloud.google.com/bigquery/docs/introduction</a></p><p><a href="http://cloud.google.com/bigquery/pricing">cloud.google.com/bigquery/pricing</a></p><p><a href="http://docs.cloud.google.com/managed-spark/docs">docs.cloud.google.com/managed-spark/docs</a></p><p><strong>Q8.</strong></p><p><strong>Cyrellis Retail wants to predict the behavior of visitors to its public website and plans to build a machine learning model. The team has strong database and SQL skills but only basic machine learning skills and wants to use its existing database-oriented expertise. Which Google Cloud product or feature should Cyrellis Retail choose?</strong></p><p><strong>A. BigQuery ML</strong></p><p><strong>B. LookML</strong></p><p><strong>C. TensorFlow</strong></p><p><strong>D. Cloud SQL</strong></p><blockquote><strong><em>A is correct because BigQuery ML lets data analysts create, train, evaluate, and use machine learning models with GoogleSQL directly in BigQuery. This allows a team that already understands SQL and database analysis to perform predictive modeling without first becoming proficient in a general-purpose machine learning framework or moving large datasets into a separate training environment. BigQuery ML is specifically designed to make machine learning accessible to SQL practitioners, which directly matches the team’s skills and the website-behavior prediction requirement.</em></strong></blockquote><p>B is incorrect because LookML is the modeling language used by Looker to define semantic data models, business metrics, dimensions, and relationships for business intelligence and analytics. It helps organizations create governed analytical experiences, but it is not itself the primary feature for training predictive machine learning models from SQL. It therefore does not directly satisfy the requirement to build the predictive model.</p><p>C is incorrect because TensorFlow is a powerful open-source machine learning framework for developing and training models, but it generally requires considerably more machine learning and programming expertise than BigQuery ML. A team with only basic ML skills that specifically wants to leverage existing SQL and database knowledge would face a steeper development and operational learning curve with TensorFlow. BigQuery ML is the more efficient fit for the stated skill profile.</p><p>D is incorrect because Cloud SQL is a managed relational database service rather than a machine learning development feature. It can store the application’s data and can be queried with SQL, but it does not provide the integrated SQL-based model creation and prediction capabilities that BigQuery ML provides. Database familiarity alone does not make Cloud SQL a machine learning solution.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/bigquery/docs/bqml-introduction">cloud.google.com/bigquery/docs/bqml-introduction</a></p><p><a href="http://cloud.google.com/bigquery/docs/bqml-introduction#advantages_of_bigquery_ml">cloud.google.com/bigquery/docs/bqml-introduction#advantages_of_bigquery_ml</a></p><p><a href="http://cloud.google.com/bigquery/docs/introduction">cloud.google.com/bigquery/docs/introduction</a></p><p><strong>Q9.</strong></p><p><strong>At the fictional company Dovaryn Systems, the organization is running all of its workloads in a private cloud based on VMware vSphere. It has decided to move to Google Cloud as quickly as possible, with minimal changes to the current environment while using the maximum amount of managed services Google offers. What should the organization do?</strong></p><p><strong>A. Migrate the workloads to Google Cloud VMware Engine</strong></p><p><strong>B. Migrate the workloads to Compute Engine</strong></p><p><strong>C. Migrate the workloads to Bare Metal Solution</strong></p><p><strong>D. Migrate the workloads to Google Kubernetes Engine</strong></p><blockquote><strong><em>A is correct because Google Cloud VMware Engine is a fully managed service that runs the VMware platform in Google Cloud and includes technologies such as vSphere, vCenter, vSAN, NSX, and HCX. It lets an organization retain familiar VMware tools, skills, and operational processes while Google manages the underlying cloud infrastructure, making it the best fit for a rapid migration with minimal environmental changes.</em></strong></blockquote><p>B is incorrect because migrating the VMware workloads to Compute Engine can be a valid modernization path, but it changes the infrastructure model from VMware to Google Cloud virtual machines. That generally requires more migration work and operational change than moving the existing VMware environment into Google Cloud VMware Engine.</p><p>C is incorrect because Bare Metal Solution provides dedicated physical infrastructure for specialized workloads with particular hardware, licensing, or legacy requirements. It is not a managed VMware software-defined data center and therefore does not preserve the existing VMware operating model as directly as Google Cloud VMware Engine.</p><p>D is incorrect because Google Kubernetes Engine is a managed Kubernetes platform for containerized workloads. Moving existing VMware virtual machine workloads to GKE would require containerization and application modernization, which conflicts with the requirement to move quickly with minimal changes.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/vmware-engine/docs">docs.cloud.google.com/vmware-engine/docs</a></p><p><a href="http://docs.cloud.google.com/vmware-engine/docs/overview">docs.cloud.google.com/vmware-engine/docs/overview</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/kubernetes-engine-overview">docs.cloud.google.com/kubernetes-engine/docs/concepts/kubernetes-engine-overview</a></p><p><strong>Q10.</strong></p><p><strong>The fictional company Veluntra Compute wants to optimize its use of Google Cloud discounts for virtual machine workloads. It expects to use 200 vCPUs continuously for the next three years and forecasts spikes of up to 300 vCPUs approximately 30% of the time. What should it choose?</strong></p><p><strong>A. 1-year committed use discount for 200 vCPUs</strong></p><p><strong>B. 3-year committed use discount for 300 vCPUs</strong></p><p><strong>C. 3-year committed use discount for 200 vCPUs</strong></p><p><strong>D. Regular pay-as-you-go pricing</strong></p><p>A is incorrect because the company expects the 200-vCPU baseline to continue for three years. A one-year commitment would provide a shorter commitment period and generally a lower discount than a three-year commitment for an equally predictable long-term baseline. Since the requirement explicitly states three years of constant use, the longer commitment better aligns with the forecast.</p><p>B is incorrect because only 200 vCPUs are required continuously. The additional 100 vCPUs are needed only during spikes that occur around 30% of the time. Committing to all 300 vCPUs for three years would risk paying for 100 committed vCPUs during the majority of periods when they are not needed.</p><blockquote><strong><em>C is correct because the 200-vCPU requirement is stable and predictable for the entire three-year period, which is an ideal pattern for a three-year committed use discount. Three-year commitments provide higher discount rates than comparable one-year commitments. Usage above the committed baseline, such as the occasional additional 100 vCPUs, can be billed as overage at applicable on-demand pricing instead of creating an unnecessarily large long-term commitment.</em></strong></blockquote><p>D is incorrect because relying entirely on pay-as-you-go pricing would forgo the savings available from committing to the known 200-vCPU baseline. On-demand pricing is useful for uncertain or burst capacity, but it is not the most cost-efficient choice for resources that the organization knows it will use continuously for three years.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview">docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview</a></p><p><a href="http://docs.cloud.google.com/compute/docs/instances/signing-up-committed-use-discounts">docs.cloud.google.com/compute/docs/instances/signing-up-committed-use-discounts</a></p><p><a href="http://docs.cloud.google.com/compute/docs/instances/upgrade-commitment-term">docs.cloud.google.com/compute/docs/instances/upgrade-commitment-term</a></p><p><strong>Q11.</strong></p><p><strong>The fictional company Doventra Systems wants to migrate an on-premises environment to Google Cloud. The environment includes container images and virtual machine disk files. Which Google Cloud products can be used to store the container images and stage the virtual machine disk files for migration?</strong></p><p><strong>A. Compute Engine and Filestore</strong></p><p><strong>B. Artifact Registry and Cloud Storage</strong></p><p><strong>C. Dataflow and BigQuery</strong></p><p><strong>D. Pub/Sub and Cloud Storage</strong></p><p>A is incorrect because Compute Engine is the destination compute service for virtual machines and Filestore provides managed NFS file storage. Neither pairing is the appropriate combination for storing container images and staging virtual disk image files for Google Cloud migration.</p><blockquote><strong><em>B is correct because Artifact Registry is Google Cloud’s current managed repository service for Docker and OCI container images, while Cloud Storage can store virtual disk image files such as VMDK or other supported formats before they are imported with Migrate to Virtual Machines. This combination directly addresses the two artifact types in the question.</em></strong></blockquote><p>C is incorrect because Dataflow is a managed data-processing service and BigQuery is an analytics data warehouse. They are designed for data processing and analytics rather than storing container images and staging virtual machine disk images for migration.</p><p>D is incorrect because Pub/Sub is a messaging service for asynchronous event and data delivery. Although Cloud Storage is appropriate for staging virtual disk files, Pub/Sub is not a container image repository, so the pair does not satisfy both requirements.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/docker">docs.cloud.google.com/artifact-registry/docs/docker</a></p><p><a href="http://docs.cloud.google.com/migrate/virtual-machines/docs/5.0/migrate/image_import">docs.cloud.google.com/migrate/virtual-machines/docs/5.0/migrate/image_import</a></p><p><a href="http://docs.cloud.google.com/storage/docs/uploading-objects">docs.cloud.google.com/storage/docs/uploading-objects</a></p><p><strong>Q12.</strong></p><p><strong>At the fictional company Lumicrest Security, the security team manages access to production systems through a group in an LDAP directory. After the directory identities and groups are synchronized into Cloud Identity or Google Workspace, how should access be managed in the Google Cloud production project?</strong></p><p><strong>A. Assign the proper role to the service account in the project’s IAM allow policy</strong></p><p><strong>B. Grant each user the roles/iam.serviceAccountUser role on a service account that exists in the Google Group.</strong></p><p><strong>C. Assign the proper role to the Google Group in the project’s IAM allow policy.</strong></p><p><strong>D. Create the project in a folder with the same name as the LDAP directory group.</strong></p><p>A is incorrect because a service account represents a workload identity, not the LDAP-based group of human users described in the question. Granting the production role to a service account would authorize workloads using that service account rather than automatically granting the required access to members of the synchronized directory group.</p><p>B is incorrect because roles/iam.serviceAccountUser allows a principal to act as or attach a service account in supported scenarios; it is not the general mechanism for granting production-resource permissions to members of a directory group. In addition, a service account being associated with a group does not replace normal IAM role bindings for the group.</p><blockquote><strong><em>C is correct because Google groups are supported IAM principals. After LDAP or Active Directory group membership is synchronized into Cloud Identity or Google Workspace, the organization can grant the required IAM role directly to the corresponding Google Group in the project’s allow policy. Users then inherit that Google Cloud access through group membership, which is more scalable and easier to administer than granting roles to individual users.</em></strong></blockquote><p>D is incorrect because a folder’s display name has no automatic relationship to an LDAP group and does not grant group members access. Folders are resource-hierarchy containers used for organization and policy inheritance; access still requires an IAM role binding to an appropriate principal such as a Google Group.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/iam/docs/groups-in-cloud-console">docs.cloud.google.com/iam/docs/groups-in-cloud-console</a></p><p><a href="http://docs.cloud.google.com/iam/docs/granting-changing-revoking-access">docs.cloud.google.com/iam/docs/granting-changing-revoking-access</a></p><p><a href="http://docs.cloud.google.com/identity/docs/how-to/create-cloud-identity-user-accounts">docs.cloud.google.com/identity/docs/how-to/create-cloud-identity-user-accounts</a></p><p><strong>Q13.</strong></p><p><strong>The fictional company Quenlith Industries needs to ensure that the Google Cloud resources of each department are segregated from those of other departments. Each department has separate development, testing, and production environments. Which resource-hierarchy strategy should the company choose?</strong></p><p><strong>A. Create a project per department, and create a folder per environment in each project.</strong></p><p><strong>B. Create a folder per department, and create a project per environment in each folder.</strong></p><p><strong>C. Create a Cloud Identity domain per department, and create a project per environment in each domain.</strong></p><p><strong>D. Create a Cloud Identity domain per environment, and create a project per department in each domain.</strong></p><p>A is incorrect because Google Cloud folders cannot be children of projects. In the resource hierarchy, folders sit above projects, not below them. In addition, putting all department environments into one project would provide less isolation between development, testing, and production.</p><blockquote><strong><em>B is correct because folders can represent departments and provide administrative and policy boundaries, while separate projects beneath each department folder can represent development, testing, and production. Projects are strong isolation, IAM, quota, and billing boundaries, so using one project per environment helps segregate resources while the department folder keeps related projects organized under a common policy hierarchy.</em></strong></blockquote><p>C is incorrect because Cloud Identity domains are identity-management boundaries, not the standard resource-hierarchy mechanism for dividing an organization’s Google Cloud resources by department. An organization normally uses folders and projects to create those resource and policy boundaries.</p><p>D is incorrect because creating a separate Cloud Identity domain for every environment is not the appropriate way to organize development, testing, and production resources. Cloud Identity manages users, groups, and organizational identity, whereas folders and projects are the intended Google Cloud resource-hierarchy constructs for this scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy">docs.cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy</a></p><p><a href="http://docs.cloud.google.com/resource-manager/docs/creating-managing-folders">docs.cloud.google.com/resource-manager/docs/creating-managing-folders</a></p><p><a href="http://docs.cloud.google.com/resource-manager/docs/creating-managing-projects">docs.cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><strong>Q14.</strong></p><p><strong>The fictional company Solvenza Media has a production job running on Compute Engine virtual machines that are members of a managed instance group. The job needs access to a BigQuery dataset. What should be included as a principal in the IAM allow policy on the BigQuery dataset?</strong></p><p><strong>A. The Compute Engine instance group</strong></p><p><strong>B. The project that owns the Compute Engine instance</strong></p><p><strong>C. The Compute Engine service account</strong></p><p><strong>D. The Compute Engine instance</strong></p><p>A is incorrect because a Compute Engine instance group is a resource used to manage a collection of VM instances; it is not the workload identity that authenticates to BigQuery. IAM access to the dataset should be granted to a supported principal representing the job’s identity.</p><p>B is incorrect because the project containing the VM is a resource container, not the specific workload principal that should receive dataset access. Granting access based on the workload’s service account provides a more precise and least-privilege authorization model.</p><blockquote><strong><em>C is correct because applications running on Compute Engine can authenticate to Google Cloud APIs through the service account attached to the VM instances. The BigQuery dataset’s IAM allow policy can grant the required BigQuery role to that service account, allowing every VM in the managed instance group that runs with the service account to access the dataset according to the granted permissions. A user-managed service account with only the required permissions is the recommended design.</em></strong></blockquote><p>D is incorrect because an individual Compute Engine VM instance is not normally the IAM principal used for calling BigQuery. The VM’s attached service account supplies the workload identity and credentials used for authorized API requests. Granting access to the service account also works cleanly with managed instance groups where instances can be recreated or replaced.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/access/create-enable-service-accounts-for-instances">docs.cloud.google.com/compute/docs/access/create-enable-service-accounts-for-instances</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/access-control">docs.cloud.google.com/bigquery/docs/access-control</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/samples/bigquery-create-dataset-iam">docs.cloud.google.com/bigquery/docs/samples/bigquery-create-dataset-iam</a></p><p><strong>Q15.</strong></p><p><strong>The fictional company Braventa Systems needs to segment Google Cloud resources used by each team from those used by other teams. The teams’ efforts change frequently, and the company needs to reduce operational risk while maintaining cost visibility. Which approach does Google recommend?</strong></p><p><strong>A. One project per team.</strong></p><p><strong>B. One organization per team.</strong></p><p><strong>C. One project that contains all of each team’s resources.</strong></p><p><strong>D. One top-level folder per team.</strong></p><p>A is incorrect because a project is an important trust, policy, quota, and billing boundary, but limiting an entire team to a single project is unnecessarily rigid when that team’s applications and initiatives change frequently. A team may need multiple projects for separate applications, environments, or lifecycle boundaries. Placing those projects under a team folder provides the required team-level grouping while preserving the flexibility to create and retire projects independently.</p><p>B is incorrect because Google recommends using a single organization resource in most enterprises unless there is a specific legal, ownership, or isolation reason to maintain separate organizations. Creating an organization resource for every team would fragment centralized governance, increase administrative overhead, and make organization-wide policy management more difficult.</p><p>C is incorrect because concentrating all changing workloads associated with a team into one long-lived project reduces the ability to isolate applications or environments by trust boundary and lifecycle. Projects are useful units for cost attribution and isolation, and a folder can group multiple projects belonging to the same team without forcing unrelated efforts into one project.</p><blockquote><strong><em>D is correct because folders are designed to model organizational structures such as departments and teams and can contain multiple projects. IAM and organization policies applied to the team folder can be inherited by the projects beneath it, which reduces repeated policy administration and operational risk. The team can create separate projects as its work changes while keeping them grouped under one stable top-level folder, preserving visibility and governance at the team level.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy">docs.cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy</a></p><p><a href="http://docs.cloud.google.com/resource-manager/docs/creating-managing-projects">docs.cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><a href="http://docs.cloud.google.com/iam/docs/resource-hierarchy-access-control">docs.cloud.google.com/iam/docs/resource-hierarchy-access-control</a></p><p><strong>Q16.</strong></p><p><strong>The fictional company Meridianyx Bank is developing an application to manage payments and online bank accounts in multiple regions. Each transaction must be handled consistently in the database, and the organization anticipates extremely large growth in stored data. Which Google Cloud product should it choose?</strong></p><p><strong>A. Cloud SQL.</strong></p><p><strong>B. Spanner.</strong></p><p><strong>C. Cloud Storage.</strong></p><p><strong>D. BigQuery.</strong></p><p>A is incorrect because Cloud SQL is a fully managed relational database for MySQL, PostgreSQL, and SQL Server and is appropriate for many transactional applications, but it does not provide the same globally distributed horizontal scaling model as Spanner. The question combines global multi-region operation, transactional consistency, and extremely large growth, which are defining strengths of Spanner.</p><blockquote><strong><em>B is correct because Spanner is a fully managed, mission-critical database that provides transactional consistency at global scale, synchronous replication for high availability, and horizontal scalability. It is specifically suited to globally distributed transactional applications that require strong consistency while scaling to very large datasets and transaction volumes. Those characteristics align directly with payment and online banking workloads that cannot tolerate inconsistent transaction state.</em></strong></blockquote><p>C is incorrect because Cloud Storage is object storage, not a relational transactional database. It is appropriate for files, backups, media, data lakes, and other object workloads, but it does not provide the relational transaction semantics required for maintaining consistent bank-account and payment records.</p><p>D is incorrect because BigQuery is an analytical data platform optimized for large-scale data analysis rather than operational transaction processing. It can analyze massive datasets, but an online banking system requires a transactional database designed for consistent reads and writes and ongoing application transactions. Spanner is the appropriate database for that operational requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs">docs.cloud.google.com/spanner/docs</a></p><p><a href="http://cloud.google.com/products/databases">cloud.google.com/products/databases</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/true-time-external-consistency">docs.cloud.google.com/spanner/docs/true-time-external-consistency</a></p><p><strong>Q17.</strong></p><p><strong>The fictional company Marovelle Hotels is considering storing customer reservation transactions in a managed cloud database rather than relying on locally managed hotel infrastructure. What is an important benefit for a large hotel chain?</strong></p><p><strong>A. On-premises hardware access to transaction data</strong></p><p><strong>B. Real-time data transformation at scale within an on-premises database</strong></p><p><strong>C. Real-time business transaction accuracy at scale</strong></p><p><strong>D. Physical hardware access during peak demand</strong></p><p>A is incorrect because moving reservation data to a managed cloud database reduces dependence on direct access to on-premises hardware. Applications access the database through supported network interfaces and APIs rather than requiring employees to physically access storage or database servers.</p><p>B is incorrect because this option describes processing within an on-premises database, which contradicts the scenario of benefiting from cloud-based data infrastructure. Cloud databases are used partly to avoid the limitations and operational burden associated with scaling local database hardware.</p><blockquote><strong><em>C is correct because managed cloud databases can support scalable, highly available transactional applications while maintaining appropriate consistency guarantees. For a hotel chain processing reservations from many properties and digital channels, accurate transaction processing at scale helps prevent conflicting or inconsistent reservation state while allowing the platform to grow with demand. Google Cloud provides database services, including Spanner for globally scalable, strongly consistent transactional workloads.</em></strong></blockquote><p>D is incorrect because physical access to hardware is not a benefit of cloud computing and is unnecessary for handling peak demand. Cloud platforms abstract the underlying physical infrastructure and provide scalable resources without requiring the customer to own or physically operate the servers.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/products/databases">cloud.google.com/products/databases</a></p><p><a href="http://docs.cloud.google.com/spanner/docs">docs.cloud.google.com/spanner/docs</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/true-time-external-consistency">docs.cloud.google.com/spanner/docs/true-time-external-consistency</a></p><p><strong>Q18.</strong></p><p><strong>At the fictional company Junivex Software, developers need a current Google Cloud observability tool that continuously analyzes how production application code consumes CPU and memory while imposing very low overhead. Which capability best matches Cloud Profiler?</strong></p><p><strong>A. Implement code updates in real time without affecting the service level objective (SLO).</strong></p><p><strong>B. Continuously analyze CPU and memory behavior of production applications with low overhead.</strong></p><p><strong>C. Manage source code and accelerate application development.</strong></p><p><strong>D. Analyze source code only during user downtime.</strong></p><p>A is incorrect because Cloud Profiler is an observability and performance-analysis service, not a code deployment mechanism. Releasing code updates requires source control and deployment or CI/CD tooling, and SLO protection is managed through release and reliability practices rather than Cloud Profiler.</p><blockquote><strong><em>B is correct because Cloud Profiler is a statistical, low-overhead profiler that continuously gathers CPU usage and memory-allocation information from production applications and attributes that resource consumption to application code. This lets developers find performance hot spots under real production workloads without requiring the application to be stopped.</em></strong></blockquote><p>C is incorrect because source-code management is handled by version-control and software-development tools. Cloud Profiler analyzes the runtime performance characteristics of application code; it does not manage repositories, branches, commits, or the development lifecycle.</p><p>D is incorrect because Cloud Profiler is specifically designed for continuous profiling in production. Its low overhead allows profiling while the application remains available to users, so deliberately waiting for user downtime is unnecessary.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/profiler/docs">docs.cloud.google.com/profiler/docs</a></p><p><a href="http://docs.cloud.google.com/profiler/docs/about-profiler">docs.cloud.google.com/profiler/docs/about-profiler</a></p><p><a href="http://docs.cloud.google.com/profiler/docs/concepts-profiling">docs.cloud.google.com/profiler/docs/concepts-profiling</a></p><p><strong>Q19.</strong></p><p><strong>The fictional company Keralune Labs wants to build entirely new infrastructure and new applications directly in the cloud rather than migrate an existing application. Which application modernization approach should the company use?</strong></p><p><strong>A. Move the application to the cloud, and then change it.</strong></p><p><strong>B. Change their application, and then move it to the cloud.</strong></p><p><strong>C. Invent in greenfield.</strong></p><p><strong>D. Invent in brownfield.</strong></p><p>A is incorrect because moving an existing application first and modernizing it afterward is a migration and modernization strategy for an existing workload. The scenario states that the organization wants to build entirely new infrastructure and applications, so there is no existing application that first needs to be moved.</p><p>B is incorrect because changing an existing application before migrating it is another modernization path for an existing workload. It can be appropriate when an organization wants to refactor or replatform before moving, but the question explicitly describes net-new application development rather than transformation of a current application.</p><blockquote><strong><em>C is correct because greenfield development means building a new application without the constraints of an existing legacy implementation. A company creating new infrastructure and applications directly in the cloud can adopt cloud-native architectures, managed services, automation, modern DevOps practices, and serverless or container-based platforms from the beginning. This is the approach most aligned with invention and transformation for net-new workloads.</em></strong></blockquote><p>D is incorrect because brownfield work involves an existing environment, application, or technology estate that must be modified, integrated, migrated, or modernized. Since the organization is building entirely new applications and infrastructure rather than working around an existing system, greenfield is the appropriate term.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/solutions/application-modernization">cloud.google.com/solutions/application-modernization</a></p><p><a href="http://cloud.google.com/blog/products/cloud-migration/real-world-cloud-migration-strategies">cloud.google.com/blog/products/cloud-migration/real-world-cloud-migration-strategies</a></p><p><a href="http://docs.cloud.google.com/architecture/migration-to-gcp-getting-started">docs.cloud.google.com/architecture/migration-to-gcp-getting-started</a></p><p><strong>Q20.</strong></p><p><strong>At the fictional company Meridianova Services, employees access a business application through a software as a service (SaaS) model.</strong></p><p><strong>What is the organization responsible for among the following choices?</strong></p><p><strong>A. Maintaining overall system operability</strong></p><p><strong>B. Maintaining customer-facing content</strong></p><p><strong>C. Monitoring data center servers</strong></p><p><strong>D. Monitoring physical computer networks</strong></p><p>A is incorrect because in a SaaS model the provider operates the application service and the underlying platform and infrastructure. The customer still has responsibilities, especially for its identities, access controls, configuration, and data, but maintaining the overall operability of the SaaS platform itself is primarily the provider’s responsibility.</p><blockquote><strong><em>B is correct because customers remain responsible for the data and content that they choose to place into and manage through a SaaS application, as well as for appropriate access to that information. Google Cloud’s shared-responsibility guidance states that in SaaS the provider owns most security responsibilities while the customer remains responsible for access controls and the data stored in the application. Customer-facing business content therefore remains within the customer’s area of responsibility.</em></strong></blockquote><p>C is incorrect because the SaaS provider operates and monitors the underlying data-center servers that support the service. One of the principal benefits of SaaS is that the customer consumes a complete application without having to manage the physical servers, operating systems, and platform layers beneath it.</p><p>D is incorrect because monitoring the provider’s physical computer networks is part of the cloud provider’s infrastructure responsibility. The customer can monitor its own application usage, access, data, and relevant connectivity, but it does not operate the physical network fabric of the SaaS provider’s data centers.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/architecture/framework/security/shared-responsibility-shared-fate">docs.cloud.google.com/architecture/framework/security/shared-responsibility-shared-fate</a></p><p><a href="http://cloud.google.com/learn/what-is-cloud-security">cloud.google.com/learn/what-is-cloud-security</a></p><p><strong>Q21.</strong></p><p><strong>The fictional company Brivanta Security suffers a major data leak only six months after upgrading its security system.</strong></p><p><strong>What should the organization do to ensure long-term security?</strong></p><p><strong>A. Extend data retention policy lengths to at least seven years.</strong></p><p><strong>B. Pay more to get the best security system available on the market.</strong></p><p><strong>C. Hire cybersecurity experts to further develop its data security plan.</strong></p><p><strong>D. Wait for a more advanced security system to be released.</strong></p><p>A is incorrect because increasing the retention period does not inherently improve security and can increase the amount of sensitive information exposed if a breach occurs. Retention periods should be based on business, legal, regulatory, and data-governance requirements. Long-term security requires ongoing risk assessment, access control, monitoring, incident response, vulnerability management, defense in depth, and continuous improvement rather than simply retaining data for a longer period.</p><p>B is incorrect because security cannot be guaranteed by purchasing the most expensive product. Security is an ongoing organizational and technical discipline involving people, processes, architecture, configuration, identity, data protection, monitoring, detection, response, and governance. Even sophisticated security products can be misconfigured or become ineffective against changing threats if the organization does not maintain a comprehensive security program.</p><blockquote><strong><em>C is correct because a major breach shortly after a technology upgrade indicates that security needs to be treated as a continuing risk-management program rather than as a one-time product purchase. Qualified cybersecurity professionals can investigate the breach, identify weaknesses, improve the security architecture, establish defense-in-depth controls, refine incident-response processes, and continuously reassess emerging threats. Google Cloud security guidance emphasizes designing and operating security as an ongoing discipline with layered preventive and detective controls.</em></strong></blockquote><p>D is incorrect because waiting for future technology leaves known risks unresolved and assumes that a newer product will automatically solve organizational security problems. Threats, vulnerabilities, configurations, and business requirements continuously change. The organization should address its current weaknesses immediately and maintain an evolving security program instead of postponing action.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/architecture/framework/security">docs.cloud.google.com/architecture/framework/security</a></p><p><a href="http://docs.cloud.google.com/architecture/blueprints/security-foundations">docs.cloud.google.com/architecture/blueprints/security-foundations</a></p><p><a href="http://cloud.google.com/security/shared-fate">cloud.google.com/security/shared-fate</a></p><p><strong>Q22.</strong></p><p><strong>The fictional company Noralume Development wants a solution that lets developers build applications using their preferred programming languages and tools while minimizing time spent on infrastructure management. What solution should the organization use?</strong></p><p><strong>A. Cloud Run</strong></p><p><strong>B. Managed Service for Apache Spark</strong></p><p><strong>C. Pub/Sub</strong></p><p><strong>D. Bare Metal Solution</strong></p><blockquote><strong><em>A is correct because Cloud Run is a fully managed application platform that runs containers, functions, and application code on scalable Google Cloud infrastructure. Developers can use virtually any programming language when they package the application appropriately, while Google manages server provisioning, much of the runtime infrastructure, and automatic scaling.</em></strong></blockquote><p>B is incorrect because Managed Service for Apache Spark is designed for Apache Spark and related big-data processing workloads. It can run Spark code in supported languages, but it is not the general-purpose managed application platform described in the question.</p><p>C is incorrect because Pub/Sub is an asynchronous messaging service that connects producers and consumers through topics and subscriptions. It is an important application-integration component but does not itself host arbitrary application code.</p><p>D is incorrect because Bare Metal Solution provides dedicated physical servers for specialized enterprise workloads with particular hardware, licensing, or legacy requirements. It involves a much more infrastructure-oriented operating model than the fully managed developer experience requested.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/overview/what-is-cloud-run">docs.cloud.google.com/run/docs/overview/what-is-cloud-run</a></p><p><a href="http://docs.cloud.google.com/run/docs/quickstarts/build-and-deploy/deploy-service-other-languages">docs.cloud.google.com/run/docs/quickstarts/build-and-deploy/deploy-service-other-languages</a></p><p><a href="http://docs.cloud.google.com/managed-spark/docs">docs.cloud.google.com/managed-spark/docs</a></p><p><strong>Q23.</strong></p><p><strong>At the fictional company Orlivena Apps, developers use Google Kubernetes Engine (GKE) and want to resolve application errors quickly while developing and releasing new versions.</strong></p><p><strong>What makes GKE an effective solution for this work?</strong></p><p><strong>A. It reduces the time needed to iterate on various solutions.</strong></p><p><strong>B. It provides intelligent recommendations to optimize application development.</strong></p><p><strong>C. It removes the need for containerizing applications.</strong></p><p><strong>D. It stores application code on multiple servers to facilitate collaboration across silos.</strong></p><blockquote><strong><em>A is correct because GKE supports modern container-based development and CI/CD workflows in which teams can build a new image, deploy it to a development or staging environment, test it, and promote or roll it back through repeatable automation. Kubernetes deployments and modern delivery tooling enable developers to test fixes and application changes rapidly, reducing the time needed to iterate when troubleshooting defects or improving features.</em></strong></blockquote><p>B is incorrect because Google Cloud provides recommendations in several operational and optimization products, but automated development recommendations are not the defining reason GKE helps developers resolve application errors. GKE’s primary value here comes from repeatable container deployment, orchestration, controlled rollouts, observability integrations, and fast iteration.</p><p>C is incorrect because GKE is a Kubernetes service and Kubernetes fundamentally orchestrates containerized workloads. Developers normally package application components in container images before deploying them as Kubernetes workloads. GKE manages the orchestration platform; it does not eliminate the container model.</p><p>D is incorrect because source-code collaboration is handled by version-control and source-repository systems, not by GKE. GKE runs containerized application workloads built from source code and images; it is not a distributed source-code repository whose purpose is to store code on multiple servers for collaboration.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/tutorials/modern-cicd-gke-developer-workflow">docs.cloud.google.com/kubernetes-engine/docs/tutorials/modern-cicd-gke-developer-workflow</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/stateless-apps">docs.cloud.google.com/kubernetes-engine/docs/how-to/stateless-apps</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/kubernetes-engine-overview">docs.cloud.google.com/kubernetes-engine/docs/concepts/kubernetes-engine-overview</a></p><p><strong>Q24.</strong></p><p><strong>The fictional company Ceralune Archives stores long-term backup files in Cloud Storage. The files are expected to be accessed only during an annual disaster recovery test. Which storage class is the most cost-effective for this access pattern?</strong></p><p><strong>A. Coldline class</strong></p><p><strong>B. Standard class</strong></p><p><strong>C. Nearline class</strong></p><p><strong>D. Archive class</strong></p><p>A is incorrect because Coldline storage is designed for infrequently accessed data and has lower at-rest storage cost than Standard or Nearline storage, but it is generally aimed at data that might be read or modified about once per quarter or less. For long-term backup data expected to be accessed only around once per year, Archive storage provides a lower at-rest storage price and is specifically designed for archival and disaster recovery use cases.</p><p>B is incorrect because Standard storage is intended for frequently accessed or short-lived data and does not have a minimum storage duration or retrieval fees. Those characteristics are valuable for active workloads, but its higher at-rest storage cost makes it unnecessarily expensive for long-term backup files that are almost never read.</p><p>C is incorrect because Nearline storage is intended for infrequently accessed data that is typically read about once per month or less. It has a 30-day minimum storage duration and retrieval charges, but its at-rest storage cost is higher than Coldline and Archive. An annual disaster-recovery test represents a much colder access pattern.</p><blockquote><strong><em>D is correct because Archive storage is Cloud Storage’s lowest-cost primary storage class for long-term archival data and disaster recovery. It has a 365-day minimum storage duration and higher retrieval and operation costs, so it is best suited to data that remains stored for long periods and is accessed very rarely. Long-term backup files retained for annual disaster-recovery testing fit that pattern better than the warmer storage classes.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/storage/docs/storage-classes">docs.cloud.google.com/storage/docs/storage-classes</a></p><p><a href="http://docs.cloud.google.com/storage/docs/lifecycle">docs.cloud.google.com/storage/docs/lifecycle</a></p><p><strong>Q25.</strong></p><p><strong>At the fictional company Evernexa AI, the organization wants to use an open-source library with a flexible ecosystem of tools to create and train its own machine learning models. Which product or solution should it use?</strong></p><p><strong>A. Cloud Run functions</strong></p><p><strong>B. Apache Beam</strong></p><p><strong>C. Dataflow</strong></p><p><strong>D. TensorFlow</strong></p><p>A is incorrect because Cloud Run functions is a serverless compute capability for running event-driven or HTTP functions. It can execute application logic that calls a machine learning model, but it is not an open-source machine learning framework for designing and training models.</p><p>B is incorrect because Apache Beam is an open-source programming model for defining batch and streaming data-processing pipelines. It can prepare or transform machine learning data, but it is not primarily a framework for constructing and training machine learning models.</p><p>C is incorrect because Dataflow is Google Cloud’s fully managed service for running batch and streaming processing pipelines, including pipelines written with Apache Beam. It can be part of an ML data pipeline, but it is not the open-source model-building framework requested.</p><blockquote><strong><em>D is correct because TensorFlow is an end-to-end machine learning platform with a broad ecosystem for building, training, evaluating, and deploying machine learning models. It provides multiple levels of abstraction, including high-level model-building APIs, and supports experimentation, distributed training, and production deployment. It therefore directly fits the requirement for a flexible machine learning framework.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://tensorflow.org/about">tensorflow.org/about</a></p><p><a href="http://tensorflow.org/guide">tensorflow.org/guide</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/overview">docs.cloud.google.com/dataflow/docs/overview</a></p><p><strong>Q26.</strong></p><p><strong>At the fictional company Gildenex Operations, the organization is running critical production workloads and requires a Google Cloud Customer Care offering with the fastest response times and a dedicated Technical Account Manager. Which service level should it choose?</strong></p><p><strong>A. Enhanced</strong></p><p><strong>B. Basic</strong></p><p><strong>C. Premium</strong></p><p><strong>D. Standard</strong></p><p>A is incorrect because Enhanced Support is designed for production workloads and provides faster response times than Standard Support, but a named Technical Account Manager is a defining Premium Support capability. Enhanced Support therefore does not satisfy all of the requirements in the question.</p><p>B is incorrect because Basic Support provides limited support resources and does not provide the rapid technical response and named Technical Account Manager required for critical enterprise production workloads.</p><blockquote><strong><em>C is correct because Premium Support is designed for enterprises running critical workloads. It offers Google’s fastest standard support response targets, including a 15-minute initial response target for qualifying business-critical P1 cases, and provides proactive guidance from a named Technical Account Manager. Those capabilities directly match both requirements in the scenario.</em></strong></blockquote><p>D is incorrect because Standard Support provides technical support suitable for workloads under development and offers slower initial response targets than Premium Support. It also does not include the named Technical Account Manager capability required by the scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/support/premium">cloud.google.com/support/premium</a></p><p><a href="http://cloud.google.com/support/enhanced">cloud.google.com/support/enhanced</a></p><p><a href="http://cloud.google.com/support">cloud.google.com/support</a></p><p><strong>Q27.</strong></p><p><strong>At the fictional company Iridexa Labs, what is generative AI?</strong></p><p><strong>A. A machine learning model for language translation</strong></p><p><strong>B. A specific field of AI that can be used to create content</strong></p><p><strong>C. An umbrella term that includes all types of machine learning algorithms</strong></p><p><strong>D. An AI model for labeling images</strong></p><p>A is incorrect because language translation can be performed by AI and machine learning systems, including generative models, but translation is only one possible application. It does not define the broader concept of generative AI.</p><blockquote><strong><em>B is correct because generative AI is a field of artificial intelligence and machine learning in which models generate new content based on patterns learned from training data. Depending on the model, generated content can include text, software code, images, audio, music, or video. The defining characteristic is creation or generation of new output rather than only classification or prediction.</em></strong></blockquote><p>C is incorrect because machine learning is broader than generative AI. Many machine learning algorithms perform classification, regression, clustering, forecasting, ranking, anomaly detection, and other tasks without generating new content. Generative AI is therefore a subset or field within the broader AI and machine learning landscape, not an umbrella term for all ML methods.</p><p>D is incorrect because image labeling is primarily a classification or recognition task. A model examines an existing image and identifies the objects or concepts it contains. A generative image model instead creates or modifies image content, so labeling images does not define generative AI.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/docs/generative-ai/glossary">docs.cloud.google.com/docs/generative-ai/glossary</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/beginners-guide">docs.cloud.google.com/gemini-enterprise-agent-platform/models/beginners-guide</a></p><p><a href="http://cloud.google.com/learn/what-is-artificial-intelligence">cloud.google.com/learn/what-is-artificial-intelligence</a></p><p><strong>Q28.</strong></p><p><strong>The fictional company Junavex Commerce needs to store daily transactional information such as customer records and purchase history. The information follows a consistent schema and records must be cross-referenced through relationships. Which type of service should the organization use?</strong></p><p><strong>A. Non-relational database</strong></p><p><strong>B. Data warehouse</strong></p><p><strong>C. Relational database</strong></p><p><strong>D. Data lake</strong></p><p>A is incorrect because non-relational databases are useful when an application needs flexible document, key-value, wide-column, or graph data models and might not require traditional relational schemas and joins. Although some NoSQL databases support transactions and relationships in various ways, the stated requirement for a consistent schema and strongly cross-referenced records maps more naturally to a relational database.</p><p>B is incorrect because a data warehouse is optimized for analytical queries, reporting, aggregation, and large-scale analysis of historical data. It can contain customer and purchase data, but it is generally not the primary operational system for daily transaction processing where records are continually created, updated, and related during business operations.</p><blockquote><strong><em>C is correct because relational databases organize structured information into tables with rows, columns, keys, and defined relationships. They are well suited to transactional application data such as customers, orders, products, and purchase history, where referential relationships and transaction consistency matter. SQL also makes it straightforward to query and join related records.</em></strong></blockquote><p>D is incorrect because a data lake is designed to store large quantities of raw structured, semi-structured, and unstructured data for future processing and analysis. It provides flexibility but does not inherently provide the relational schema, referential relationships, and transactional database behavior requested by this application.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/learn/what-is-a-relational-database">cloud.google.com/learn/what-is-a-relational-database</a></p><p><a href="http://docs.cloud.google.com/sql/docs/introduction">docs.cloud.google.com/sql/docs/introduction</a></p><p><a href="http://cloud.google.com/learn/what-are-transactional-databases">cloud.google.com/learn/what-are-transactional-databases</a></p><p><strong>Q29.</strong></p><p><strong>At the fictional company Brivanta Retail, the organization processes monthly batches of sales data to analyze trends and derive business insights. It wants more current information so it can improve decision-making and respond closer to real time. What should the organization do?</strong></p><p><strong>A. Change from a relational database to a NoSQL database.</strong></p><p><strong>B. Filter the data so reports are generated faster.</strong></p><p><strong>C. Switch from batch processing to stream processing.</strong></p><p><strong>D. Process batch reports weekly instead of monthly.</strong></p><p>A is incorrect because changing the database data model does not inherently make the analytics pipeline real time. Relational and NoSQL databases can both participate in batch or streaming architectures. The central limitation is that the data is processed only at scheduled intervals rather than continuously as new events arrive.</p><p>B is incorrect because filtering can reduce the volume of data processed and might improve the speed of a particular report, but it does not remove the latency created by waiting for the monthly batch. The organization needs a different processing model if it wants insights based on continuously arriving data.</p><blockquote><strong><em>C is correct because stream processing continuously processes data records as they arrive instead of waiting for a large scheduled batch. Google Cloud describes streaming analytics as appropriate for near-real-time aggregation, correlation, filtering, and decision-making. Moving the sales pipeline to a streaming model can therefore make current transaction information available for analysis much sooner and support faster business decisions.</em></strong></blockquote><p>D is incorrect because moving from monthly to weekly batches reduces the delay but still leaves the organization working with stale data for as long as a week. It is an improvement over monthly processing but does not satisfy the stated goal of near-real-time analysis as effectively as stream processing.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/learn/what-is-streaming-analytics">cloud.google.com/learn/what-is-streaming-analytics</a></p><p><a href="http://cloud.google.com/products/dataflow">cloud.google.com/products/dataflow</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines">docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines</a></p><p><strong>Q30.</strong></p><p><strong>At the fictional company Farlivex Compute, the organization uses Compute Engine and wants to receive sustained-use discounts on eligible resources. What should the organization do?</strong></p><p><strong>A. Choose Spot VMs when creating virtual machines.</strong></p><p><strong>B. Choose low-powered virtual machines.</strong></p><p><strong>C. Nothing. Sustained-use discounts are automatically applied.</strong></p><p><strong>D. Commit to virtual machine usage for a one- or three-year period.</strong></p><p>A is incorrect because Spot VMs are a separate discounted Compute Engine purchasing model for fault-tolerant workloads that can tolerate preemption. Spot pricing can substantially reduce compute cost, but selecting Spot VMs is not the action required to receive sustained-use discounts and Spot usage follows its own pricing model.</p><p>B is incorrect because sustained-use discounts are based on how long eligible Compute Engine resources are used during the billing month, not simply on choosing a lower-powered machine type. Some eligible machine series receive sustained-use discounts when their qualifying usage exceeds the relevant monthly threshold.</p><blockquote><strong><em>C is correct because Compute Engine automatically calculates and applies sustained-use discounts to eligible resource usage. When eligible resources are used for more than 25% of a billing month and are not covered by another incompatible discount, the applicable sustained-use discount is applied automatically. The customer does not need to purchase or manually activate the discount.</em></strong></blockquote><p>D is incorrect because committing to one or three years is how an organization obtains committed use discounts, which are a separate discount mechanism for predictable usage. Committed use discounts require a commitment, while sustained-use discounts require no commitment and are automatically calculated for eligible usage.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/sustained-use-discounts">docs.cloud.google.com/compute/docs/sustained-use-discounts</a></p><p><a href="http://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview">docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview</a></p><p><a href="http://docs.cloud.google.com/compute/docs/instances/spot">docs.cloud.google.com/compute/docs/instances/spot</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8c40223d977" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Generative AI Leader Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-generative-ai-leader-practice-questions-888830d2ef94?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/888830d2ef94</guid>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 12:02:24 GMT</pubDate>
            <atom:updated>2026-10-07T12:02:24.312Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FtoRBXWmKhKoslGMGs39LQ.jpeg" /><figcaption>KloudExams GCP Google Cloud Generative AI Leader Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Generative AI Leader certification</strong>? Knowing what generative AI can do is a starting point — understanding when to use it, how to recognise its limitations, and how to apply it responsibly is what helps you identify meaningful business opportunities.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise evaluating generative AI approaches against specific business needs. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through official Google Cloud learning resources and practical generative AI use cases.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>The fictitious highly regulated financial institution Solvexa Capital wants to use Gemini as the core decision engine for a loan approval system that will deterministically approve or reject loan applications based on a strict set of predefined criteria. Why is this an inappropriate use case for Gemini?</strong></p><p><strong>A. Gemini deployment for this scenario would be too expensive and complex.</strong></p><p><strong>B. Gemini is designed for flexible content generation and inference, not rigid rule-based decisions.</strong></p><p><strong>C. Gemini is not equipped to handle structured numerical data for financial assessments.</strong></p><p><strong>D. Gemini cannot integrate with required financial databases.</strong></p><p>A is incorrect because cost and deployment complexity are not the fundamental reason the proposed architecture is inappropriate. Gemini models on Google Cloud can be consumed through managed APIs, so an organization does not necessarily need to operate model-serving infrastructure itself. Cost, latency, and operational complexity are legitimate design considerations, but they do not address the core requirement in the question: the loan decision must follow strict predefined criteria and be deterministic.</p><blockquote><strong><em>B is correct because Gemini is a generative model family intended for flexible reasoning, understanding, and content generation, and generative output is not guaranteed to be perfectly deterministic. Even when generation parameters are configured for lower variability or a fixed seed is used, Google Cloud documentation explicitly notes that deterministic output is not guaranteed. A loan approval policy based on fixed regulatory or business rules should therefore use deterministic rules, policy logic, or a state-machine-style decision process for the approval decision itself. Gemini could still assist with non-deterministic supporting tasks, such as summarizing documents or extracting information, provided appropriate controls are used.</em></strong></blockquote><p>C is incorrect because Gemini can process and reason over numerical information represented in supported inputs, and the inability to handle structured numerical data is not the issue described. A model can be provided with financial values and can generate analyses or explanations based on them. The concern is that a probabilistic generative model should not replace strict deterministic decision logic when the outcome must exactly follow predefined approval rules every time.</p><p>D is incorrect because Gemini-based applications can be integrated with external systems through application code, APIs, and tool or function-calling patterns. Databases and financial systems can be exposed to an application through controlled interfaces, subject to authentication, authorization, governance, and security requirements. Integration capability therefore does not make this use case inappropriate. The mismatch lies between probabilistic generative inference and the requirement for a rigid, deterministic approval engine.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/inference">docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/inference</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values</a></p><p><a href="http://developers.google.com/machine-learning/glossary/agent">developers.google.com/machine-learning/glossary/agent</a></p><p><strong>Q2.</strong></p><p><strong>A user at the fictitious company Cinderleaf Media asks a generative AI model about the scientific accuracy of a popular science fiction movie. The model confidently states that humans can travel faster than light, references specific but entirely fictional theories, and provides made-up explanations of how this is achieved according to the movie’s “established science.” The model presents this information as factual without indicating that it originates from a fictional work. What type of model limitation is this?</strong></p><p><strong>A. Bias</strong></p><p><strong>B. Hallucination</strong></p><p><strong>C. Knowledge cutoff</strong></p><p><strong>D. Data dependency</strong></p><p>A is incorrect because bias refers to systematic tendencies that can produce unfair, skewed, or prejudiced outcomes, often because of issues in training data, data collection, labeling, model design, or interaction patterns. The scenario does not primarily show systematic favoritism or unfair treatment toward a group. Instead, the model invents false theories and presents them confidently as real facts, which is the hallmark of hallucination.</p><blockquote><strong><em>B is correct because a hallucination occurs when a generative model produces plausible-sounding but factually incorrect, fabricated, irrelevant, or otherwise unsupported information and presents it as though it were valid. Here, the model invents scientific theories and explanations and fails to distinguish fictional material from real-world scientific knowledge. The confidence of the response does not make it reliable; confident fabrication is a common form of model hallucination.</em></strong></blockquote><p>C is incorrect because a knowledge cutoff concerns information that was not available to a model because it arose after the period represented in its training or knowledge. Nothing in the scenario depends on a recent event that occurred after a cutoff date. The model is not merely missing up-to-date information; it is creating false scientific claims and fictional supporting explanations. That makes hallucination the more precise limitation.</p><p>D is incorrect because data dependency is a broad concept describing how model behavior and quality depend on the data used for training, tuning, grounding, or prompting. Poor data can contribute to many downstream problems, including bias and incorrect outputs, but the observed behavior in this question has a specific name: the model fabricated factual-sounding information. The most direct classification of that failure is hallucination.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini/docs/discover/responsible-ai">docs.cloud.google.com/gemini/docs/discover/responsible-ai</a></p><p><a href="http://developers.google.com/machine-learning/glossary">developers.google.com/machine-learning/glossary</a></p><p><a href="http://developers.google.com/machine-learning/glossary/responsible-ai">developers.google.com/machine-learning/glossary/responsible-ai</a></p><p><strong>Q3.</strong></p><p><strong>The fictitious organization Kestrava Services is collecting data to train a generative AI model for customer service. They want to ensure security throughout the machine learning lifecycle. What is a critical consideration at this stage?</strong></p><p><strong>A. Applying the latest software patches to the AI model on a regular basis.</strong></p><p><strong>B. Monitoring the AI model’s performance for unexpected outputs and potential errors.</strong></p><p><strong>C. Establishing ethical guidelines for AI model responses to ensure fairness and avoid harm.</strong></p><p><strong>D. Implementing access controls and protecting sensitive information within the training data.</strong></p><p>A is incorrect because software patching is an important operational security practice for the infrastructure, operating systems, libraries, containers, and other components that host or support an AI workload, but an AI model itself is not typically “patched” in the same sense as an operating system or vulnerable software package. More importantly, the stem focuses on the data-collection and training stage. At that stage, controlling access to training data and protecting sensitive information are more directly relevant security priorities.</p><p>B is incorrect because monitoring model behavior is essential during evaluation and production operation, where teams need to identify unexpected outputs, drift, safety issues, quality regressions, and other failures. However, the question asks for a critical consideration while collecting data for training. Protecting the confidentiality and integrity of the training data through access controls and sensitive-data protections is therefore more immediate at this stage than runtime output monitoring.</p><p>C is incorrect because ethical guidelines and fairness policies are important elements of responsible AI governance, but they do not directly address the security risk highlighted by the collection of potentially sensitive training information. Responsible AI and security overlap, yet they are not interchangeable. A model-response policy can reduce harmful behavior, while access controls, data classification, de-identification, encryption, and related controls protect the underlying data from unauthorized access or exposure.</p><blockquote><strong><em>D is correct because training data is a core asset in the AI lifecycle and may contain personal, confidential, regulated, proprietary, or otherwise sensitive information. Organizations should control which identities and services can access that data, apply least-privilege permissions, protect data in storage and transit, and use appropriate security boundaries and sensitive-data controls. Securing training data early reduces the risk of unauthorized disclosure, data exfiltration, inappropriate model exposure, and compliance failures later in the lifecycle.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/security/securing-ai">cloud.google.com/security/securing-ai</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/security-controls">docs.cloud.google.com/gemini-enterprise-agent-platform/models/security-controls</a></p><p><a href="http://cloud.google.com/security/vpc-service-controls">cloud.google.com/security/vpc-service-controls</a></p><p><strong>Q4.</strong></p><p><strong>The fictitious company Vantessa Commerce wants to build a model to classify customer reviews as positive, negative, or neutral. They have collected a dataset of thousands of customer reviews, and each review has been manually tagged with the corresponding sentiment: positive, negative, or neutral. What type of machine learning should the company use?</strong></p><p><strong>A. Reinforcement learning</strong></p><p><strong>B. Deep learning</strong></p><p><strong>C. Unsupervised learning</strong></p><p><strong>D. Supervised learning</strong></p><p>A is incorrect because reinforcement learning is designed for sequential decision-making in which an agent interacts with an environment and receives rewards or penalties for its actions. The customer-review dataset does not describe an environment, actions, or a reward feedback loop. Instead, it contains input examples paired with known correct sentiment labels, which is the standard setup for supervised classification.</p><p>B is incorrect because deep learning identifies a family of model architectures rather than the fundamental learning setup. A deep neural network could certainly be used to implement a sentiment classifier, especially for large text datasets, but it would still be trained in a supervised manner because each review has a known target sentiment. The question asks which type of machine learning fits the labeled-data situation, so supervised learning is more precise.</p><p>C is incorrect because unsupervised learning is used when the training data does not provide target labels and the objective is to discover patterns, clusters, associations, or latent structure. In this scenario, every customer review is already manually labeled as positive, negative, or neutral. Ignoring those labels and using an unsupervised method would discard exactly the supervisory signal that makes direct sentiment classification possible.</p><blockquote><strong><em>D is correct because supervised learning trains a model using examples for which the correct output is known. Here, the review text is the input and the manually assigned sentiment is the label. The model can learn relationships between textual patterns and the three sentiment categories and then apply that learned mapping to previously unseen reviews. Because there are three mutually exclusive target classes, this is a supervised multi-class classification problem.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/discover/what-is-supervised-learning">cloud.google.com/discover/what-is-supervised-learning</a></p><p><a href="http://developers.google.com/machine-learning/intro-to-ml/what-is-ml">developers.google.com/machine-learning/intro-to-ml/what-is-ml</a></p><p><a href="http://developers.google.com/machine-learning/crash-course/classification">developers.google.com/machine-learning/crash-course/classification</a></p><p><strong>Q5.</strong></p><p><strong>The fictitious company Caldris Surveyworks is developing a generative AI application to analyze customer feedback collected through online surveys. Stakeholders are concerned about potential privacy risks associated with this data because the feedback contains personally identifiable information. They need to mitigate these risks before using the data to train the AI model. What action should the company prioritize?</strong></p><p><strong>A. Focusing on collecting only quantitative feedback data in future surveys.</strong></p><p><strong>B. Implementing strong access controls to limit which teams can view the raw survey data.</strong></p><p><strong>C. Ensuring that the AI model is trained on a large and diverse dataset.</strong></p><p><strong>D. Applying data anonymization techniques to remove or obscure sensitive data.</strong></p><p>A is incorrect because collecting only quantitative feedback in future surveys does not adequately address the personally identifiable information already present in the dataset intended for training. It can also unnecessarily reduce the richness of customer feedback without guaranteeing privacy, because quantitative data can still contain identifying or quasi-identifying attributes. The immediate requirement is to reduce privacy risk in the data that will actually be used for model training.</p><p>B is incorrect because strong access controls are an important privacy and security safeguard and should be implemented, but access restriction alone does not remove personally identifiable information from the training dataset. Authorized users, training pipelines, or downstream systems could still process raw PII. Since the question asks what should be prioritized before using the data to train the model, de-identifying or anonymizing the sensitive fields directly reduces the amount of identifiable information exposed to the training process.</p><p>C is incorrect because dataset size and diversity can improve model quality, representativeness, and fairness, but those properties do not solve a privacy problem. A large, diverse dataset can still contain names, email addresses, phone numbers, account identifiers, or other sensitive information. Increasing the amount of data without first protecting PII could actually increase the amount of sensitive information placed at risk.</p><blockquote><strong><em>D is correct because anonymization or de-identification techniques can remove, redact, mask, tokenize, generalize, or otherwise transform sensitive information before it enters the model-training workflow. This directly addresses the stated privacy risk by reducing the ability to associate training records with identifiable individuals. Access controls, governance, retention policies, and other safeguards should still be used as complementary controls, but de-identification is the most direct priority among the listed choices when PII must be mitigated before training.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sensitive-data-protection/docs/classification-redaction">docs.cloud.google.com/sensitive-data-protection/docs/classification-redaction</a></p><p><a href="http://docs.cloud.google.com/sensitive-data-protection/docs/inspect-sensitive-text-de-identify">docs.cloud.google.com/sensitive-data-protection/docs/inspect-sensitive-text-de-identify</a></p><p><a href="http://cloud.google.com/security/securing-ai">cloud.google.com/security/securing-ai</a></p><p><strong>Q6.</strong></p><p><strong>The fictitious company Velunex Products collects customer feedback through open-ended survey questions where customers can write detailed responses in their own words, such as “The product was easy to use, and the customer support was excellent, but the delivery took longer than expected.” What type of data is this?</strong></p><p><strong>A. Quantitative data</strong></p><p><strong>B. Structured data</strong></p><p><strong>C. Labeled data</strong></p><p><strong>D. Unstructured data</strong></p><p>A is incorrect because quantitative data consists primarily of measurable numeric values that can be directly counted or analyzed mathematically, such as a rating from 1 to 5, purchase amount, response time, or number of support cases. The example is a free-form natural-language response that expresses multiple ideas and sentiments rather than a predefined numeric measurement. Numerical features could later be extracted from it, but the original survey response is not quantitative data.</p><p>B is incorrect because structured data follows a predefined schema with well-defined fields, types, rows, columns, or other consistently organized elements. An open-ended written response can vary greatly in length, vocabulary, topics, grammar, and meaning and does not naturally conform to a fixed tabular schema. Metadata associated with the response, such as customer ID, timestamp, or survey question number, may be structured, but the free-form text itself is unstructured.</p><p>C is incorrect because labeled data means that an example has been assigned a target label or annotation for a specific machine learning task. For example, if the survey response were manually tagged as “mixed sentiment” or “delivery complaint,” then the response-label pair could form labeled training data. The question only states that customers provide free-form responses; it does not state that those responses have been assigned labels.</p><blockquote><strong><em>D is correct because free-form natural-language text is a standard example of unstructured data. It does not follow a rigid predefined data model and can contain varying subjects, opinions, sentence structures, and levels of detail. Natural language processing and generative AI techniques are commonly used to extract information such as sentiment, themes, entities, or summaries from this kind of unstructured text.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/discover/unstructured-data">cloud.google.com/discover/unstructured-data</a></p><p><a href="http://cloud.google.com/discover/data-lake-vs-data-warehouse">cloud.google.com/discover/data-lake-vs-data-warehouse</a></p><p><a href="http://developers.google.com/machine-learning/decision-forests/appropriate-data">developers.google.com/machine-learning/decision-forests/appropriate-data</a></p><p><strong>Q7.</strong></p><p><strong>At the fictional company Orvexa Holdings, the office of the CISO wants to use generative AI to help automate tasks such as summarizing case information, researching threats, and taking actions such as creating detection rules. What agent should they use?</strong></p><p><strong>A. Data agent</strong></p><p><strong>B. Security agent</strong></p><p><strong>C. Customer service agent</strong></p><p><strong>D. Code agent</strong></p><p>A is incorrect because a data agent is oriented toward working with data, such as querying datasets, generating insights, or assisting with analytics workflows. The tasks in the scenario are specifically security-operations activities: incident investigation, threat research, case summarization, and detection engineering.</p><blockquote><strong><em>B is correct because a security agent is designed for cybersecurity and security-operations workflows. Google Security Operations uses Gemini capabilities to help analysts summarize cases, investigate threats, recommend response actions, search security telemetry with natural language, and create detections or playbooks. Those capabilities align directly with the CISO organization’s stated requirements.</em></strong></blockquote><p>C is incorrect because a customer service agent is designed to assist customers or contact-center personnel with support interactions. Its typical responsibilities include answering customer questions, recommending responses, retrieving knowledge, or automating service conversations, rather than investigating cyber threats or authoring detection rules.</p><p>D is incorrect because a code agent primarily assists with software-development tasks such as generating, explaining, reviewing, testing, or modifying code. Although code-oriented tools can help with some security-related programming, they are not the best role-specific agent for case investigation, threat intelligence, security recommendations, and detection engineering across a security operations environment.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/security/products/security-operations">cloud.google.com/security/products/security-operations</a></p><p><a href="http://cloud.google.com/security/products/security-operations/investigate">cloud.google.com/security/products/security-operations/investigate</a></p><p><a href="http://cloud.google.com/security/products/security-information-event-management">cloud.google.com/security/products/security-information-event-management</a></p><p><strong>Q8.</strong></p><p><strong>A software developer at the fictional company Draxelle Software needs a highly efficient, open large language model that can be fine-tuned on a local machine for rapid prototyping of a chatbot application. The model should provide strong natural-language understanding and generation while remaining lightweight enough to run on limited hardware. Which Google-developed family of models should the developer use?</strong></p><p><strong>A. Gemma</strong></p><p><strong>B. Gemini 3.1 Flash Image (Nano Banana 2)</strong></p><p><strong>C. Gemini</strong></p><p><strong>D. Veo</strong></p><blockquote><strong><em>A is correct because Gemma is Google’s family of lightweight open models intended for developers who want flexible deployment and customization. Gemma models are available in multiple sizes, can be tuned for specialized tasks, and are designed to run in environments ranging from laptops and developer workstations to cloud infrastructure. These properties make Gemma well suited to local prototyping and fine-tuning for a chatbot.</em></strong></blockquote><p>B is incorrect because Gemini 3.1 Flash Image, also known as Nano Banana 2, is optimized for image generation and editing. Its core purpose is visual content creation from text or image inputs rather than serving as a lightweight general-purpose open language model for local chatbot fine-tuning.</p><p>C is incorrect because Gemini is Google’s flagship family of powerful multimodal foundation models and is commonly accessed through managed Gemini services and APIs. It is a strong choice for many enterprise generative AI applications, but the requirement emphasizes a lightweight open model that the developer can fine-tune and run locally, which aligns more directly with Gemma.</p><p>D is incorrect because Veo is Google’s video-generation model family. It is intended for generating and working with video content rather than providing a lightweight, locally fine-tunable large language model for conversational natural-language applications.</p><p><strong>Reference links:</strong></p><p><a href="https://docs.google.com/document/u/0/d/1eRu7tmB7ckRC3NrncUtGKJ4vPKoYLU0L0mPStoOiKfQ/edit">ai.google.dev/gemma/docs</a></p><p><a href="https://docs.google.com/document/u/0/d/1eRu7tmB7ckRC3NrncUtGKJ4vPKoYLU0L0mPStoOiKfQ/edit">ai.google.dev/gemma/docs/get_started</a></p><p><a href="https://docs.google.com/document/u/0/d/1eRu7tmB7ckRC3NrncUtGKJ4vPKoYLU0L0mPStoOiKfQ/edit">ai.google.dev/gemma/docs/core/model_card</a></p><p><strong>Q9.</strong></p><p><strong>The fictional company Eduquanta Labs wants to create an AI-powered educational solution that provides personalized learning experiences for students. The platform will assess a student’s knowledge, recommend relevant learning materials, generate personalized exercises, provide lesson structure, and track progress. What type of AI solution should the company use?</strong></p><p><strong>A. A learning management system (LMS)</strong></p><p><strong>B. A customized learning agent</strong></p><p><strong>C. An AI-powered recommendation system for learning resources</strong></p><p><strong>D. A large language model fine-tuned on educational content</strong></p><p>A is incorrect because a learning management system is primarily designed to organize courses, deliver content, manage enrollment, and track learner progress. An LMS can be an important component of the overall application, but by itself it does not necessarily reason about an individual student’s knowledge, dynamically select interventions, and generate personalized exercises as an autonomous AI solution.</p><blockquote><strong><em>B is correct because a customized learning agent can combine multiple capabilities and pursue the broader educational goal across an ongoing interaction. It can assess the learner’s current state, retrieve or recommend appropriate materials, generate exercises, adapt subsequent instruction based on responses, and coordinate progress-related actions. This agentic approach best matches the multi-step, personalized behavior required by the scenario.</em></strong></blockquote><p>C is incorrect because an AI-powered recommendation system focuses on selecting relevant resources based on user characteristics, behavior, or preferences. Recommendations would satisfy only one part of the requirement. Such a system would not inherently generate personalized exercises, manage a learning conversation, assess mastery across steps, or coordinate the complete learning journey.</p><p>D is incorrect because a fine-tuned large language model could generate educational explanations and exercises, but a model alone does not provide the orchestration, state management, progress tracking, tool use, or multi-step decision-making needed for the full solution. A customized agent can use an appropriate language model while also coordinating these additional capabilities.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/agents">docs.cloud.google.com/gemini-enterprise-agent-platform/agents</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/agent-studio/design-agents">docs.cloud.google.com/gemini-enterprise-agent-platform/agent-studio/design-agents</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/overview">docs.cloud.google.com/gemini-enterprise-agent-platform/overview</a></p><p><strong>Q10.</strong></p><p><strong>The development team at the fictional company Novyrix Digital is eager to build generative AI solutions with Google Cloud but has limited AI-development experience and needs to launch its solution quickly. What Google Cloud benefit would most directly help the company achieve this goal?</strong></p><p><strong>A. Google Cloud’s pre-trained models and low- and no-code AI tools and services.</strong></p><p><strong>B. Google Cloud’s focus on continuous improvement provides access to the latest AI tools, features, and best practices.</strong></p><p><strong>C. Google Cloud’s comprehensive training materials and tutorials help developers.</strong></p><p><strong>D. Google Cloud’s collaborative AI community and support forums connect developers with AI experts.</strong></p><blockquote><strong><em>A is correct because pre-trained foundation models and low-code or no-code development experiences reduce the amount of specialized model-development work required before a team can create a useful application. Gemini Enterprise Agent Platform provides access to foundation models through Model Garden, while Agent Studio supports low-code interaction and agent design. These capabilities are the most direct way for a team with limited AI expertise to prototype and launch quickly.</em></strong></blockquote><p>B is incorrect because continuous product improvement is valuable over the long term and gives teams access to newer capabilities, but it does not directly remove the immediate development barriers faced by an inexperienced team. The need in the scenario is rapid solution creation with minimal specialized AI work, which is better addressed by pre-trained models and simplified development tools.</p><p>C is incorrect because training materials and tutorials can improve the team’s knowledge and are useful for skill development, but learning takes time and does not by itself provide an immediately deployable model or simplified development environment. It is therefore less direct than using ready-to-use models and low-code tools.</p><p>D is incorrect because communities and support resources can help developers solve problems and learn from experts, but they are supplementary resources. They do not reduce implementation effort as directly as pre-trained models, managed APIs, Model Garden, and low-code agent-building capabilities.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform">cloud.google.com/products/gemini-enterprise-agent-platform</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/overview">docs.cloud.google.com/gemini-enterprise-agent-platform/overview</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models">docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models</a></p><p><strong>Q11.</strong></p><p><strong>The marketing team at the fictional company Marquelyt Media wants to use a Google foundation model for social media and advertising campaigns. The team needs to create written articles and images from text, lacks deep AI expertise, and wants a versatile solution. Which Google foundation model family should it use?</strong></p><p><strong>A. Gemini</strong></p><p><strong>B. Gemma</strong></p><p><strong>C. Veo</strong></p><p><strong>D. Gemini 3.1 Flash Image (Nano Banana 2)</strong></p><blockquote><strong><em>A is correct because Gemini is Google’s broad, multimodal foundation-model family and is the best fit when a team needs versatile generative capabilities rather than a narrowly specialized media model. Gemini supports rich text generation, and current Gemini image models can generate images from text prompts. This gives a marketing team a unified model family for producing written content and visual campaign assets while using managed Google AI interfaces rather than building models from scratch.</em></strong></blockquote><p>B is incorrect because Gemma is Google’s family of open models intended for flexible deployment, customization, experimentation, and developer-controlled use. Gemma is valuable when open weights and fine-tuning are priorities, but the scenario emphasizes a team with limited AI expertise that wants a broadly capable managed solution for campaign content.</p><p>C is incorrect because Veo is specialized for video generation. It is appropriate when the primary requirement is creating video content, but it is not the best general foundation-model choice for producing both written articles and still-image marketing assets.</p><p>D is incorrect because Gemini 3.1 Flash Image, also known as Nano Banana 2, is specifically optimized for image generation and editing. It is an excellent choice for visual assets, but the requirement spans substantial written content as well as images. The broader Gemini family is therefore the more complete answer for this mixed-content marketing use case.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models">docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/image-generation">docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/image-generation</a></p><p><a href="http://cloud.google.com/solutions/ai">cloud.google.com/solutions/ai</a></p><p><strong>Q12.</strong></p><p><strong>The data science team at the fictional company Datavexa Analytics needs a centralized and organized location to store multiple model versions, track model metadata, and easily deploy selected versions to applications. What Google Cloud service should the team use?</strong></p><p><strong>A. Cloud Storage</strong></p><p><strong>B. Agent Platform Pipelines</strong></p><p><strong>C. BigQuery</strong></p><p><strong>D. Model Registry</strong></p><p>A is incorrect because Cloud Storage is an object-storage service that can hold model artifact files, datasets, and other binary objects, but it does not by itself provide the model-centric lifecycle features requested. It lacks the specialized registry experience for organizing models and versions, tracking model resources, and deploying selected registered models to endpoints.</p><p>B is incorrect because Agent Platform Pipelines is designed to orchestrate and automate repeatable machine learning workflows such as data preparation, training, evaluation, and deployment. Pipelines can produce and register models as part of an MLOps workflow, but they are not the centralized model catalog used to manage model versions and their deployment lifecycle.</p><p>C is incorrect because BigQuery is an analytics data warehouse and can also create and manage BigQuery ML models. However, the requirement is for a centralized registry that organizes different model versions, associated metadata, and deployment actions across the machine learning lifecycle, which is the role of Model Registry.</p><blockquote><strong><em>D is correct because Model Registry in Gemini Enterprise Agent Platform is the centralized repository for managing the lifecycle of machine learning models. It provides an organized view of models and model versions, supports model versioning and associated metadata, and allows a selected model version to be deployed to an endpoint. Those capabilities directly match the team’s requirements.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/introduction">docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/introduction</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/versioning">docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/versioning</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/pipelines/introduction">docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/pipelines/introduction</a></p><p><strong>Q13.</strong></p><p><strong>The fictional company Auralith Systems is evaluating Google Cloud for a new generative AI initiative. What is a characteristic of Google Cloud as a generative AI company?</strong></p><p><strong>A. Google Cloud relies on proprietary, closed-source AI technologies for maximum security benefits.</strong></p><p><strong>B. Google Cloud has an AI-first focus that enables innovation, with continuous updates and broad integration across its platform.</strong></p><p><strong>C. Google Cloud ensures that all generative AI models and data are completely secured and isolated from external networks.</strong></p><p><strong>D. Google Cloud provides fully autonomous AI agents that require zero configuration or management overhead.</strong></p><p>A is incorrect because Google Cloud does not rely exclusively on proprietary closed technologies. In addition to Google models such as Gemini, Gemini Enterprise Agent Platform provides access to Google, partner, and open models through Model Garden and supports open frameworks for building AI applications and agents. Security is achieved through platform controls, governance, identity, networking, and other protections rather than by requiring every model to be closed.</p><blockquote><strong><em>B is correct because Google Cloud integrates generative AI broadly across its platform and continually adds models, agent-development capabilities, MLOps tools, security features, data integrations, and AI-assisted experiences. Gemini Enterprise Agent Platform provides a unified environment for models and agents, while other Google Cloud products integrate Gemini capabilities for data, security, customer experience, development, and productivity. This continuing AI-centered innovation is the characteristic that best matches the choices.</em></strong></blockquote><p>C is incorrect because the statement makes an absolute guarantee that all generative AI models and data are completely isolated from external networks. Google Cloud provides strong and configurable security controls, but connectivity, data access, grounding, integrations, APIs, and network architecture depend on the customer’s design and chosen services. Complete network isolation is not an inherent universal property of every generative AI workload.</p><p>D is incorrect because AI agents still require purposeful design, instructions, tools, permissions, data connections, evaluation, security controls, deployment decisions, monitoring, and governance. Google Cloud provides managed services that reduce operational effort, but enterprise agents are not universally zero-configuration or free of management responsibilities.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/solutions/ai">cloud.google.com/solutions/ai</a></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform">cloud.google.com/products/gemini-enterprise-agent-platform</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models">docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models</a></p><p><strong>Q14.</strong></p><p><strong>Fictional company Quendrix Systems wants to adopt generative AI but is concerned about vendor lock-in and wants flexibility in the technologies and models it can use. Which Google Cloud strength most directly helps address this concern?</strong></p><p><strong>A. Google Cloud strictly adheres to proprietary technologies, ensuring consistent performance across its ecosystem.</strong></p><p><strong>B. Google Cloud provides pre-packaged solutions for easy deployment, eliminating the need for customization and integration.</strong></p><p><strong>C. Google Cloud AI solutions use an open approach that supports customer choice across offerings.</strong></p><p><strong>D. Google Cloud focuses on automation with the goal of replacing human jobs with AI.</strong></p><p>A is incorrect because an exclusively proprietary approach would increase, not reduce, concerns about vendor lock-in. Google Cloud supports its own models as well as open and partner models, multiple serving approaches, common development frameworks, and an ecosystem intended to give organizations architectural choices rather than forcing every workload into one proprietary model family.</p><p>B is incorrect because managed and prebuilt services can simplify deployment, but Google Cloud does not eliminate the need for customization or integration in every generative AI project. More importantly, pre-packaging alone does not address vendor lock-in. A company concerned about long-term flexibility needs model and technology choice, portability options, and the ability to select different approaches as requirements evolve.</p><blockquote><strong><em>C is correct because Google Cloud emphasizes an open and flexible AI ecosystem. Through Gemini Enterprise Agent Platform and Model Garden, customers can work with Google models as well as supported open and partner models, and they can choose among managed and self-deployed serving approaches where appropriate. This breadth of choice makes it easier to select technology based on workload needs and reduces dependence on a single model provider or one rigid implementation path.</em></strong></blockquote><p>D is incorrect because the purpose of Google Cloud’s generative AI offerings is to augment business processes, improve productivity, and enable new applications, not to define success as replacing people. Human oversight, responsible AI practices, and workflow design remain important. This option also has no connection to the company’s stated concern about technology choice and vendor lock-in.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform">cloud.google.com/products/gemini-enterprise-agent-platform</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models">docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/choose-serving-option">docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/choose-serving-option</a></p><p><strong>Q15.</strong></p><p><strong>Fictional company Solvanta Sales finds that its sales team spends too much time researching leads and manually entering information into its CRM. The company wants to improve efficiency so representatives can focus on customer relationships and closing deals. What should the company do?</strong></p><p><strong>A. Develop a custom AI solution using the Cloud Natural Language API to analyze lead communications and automatically update the CRM.</strong></p><p><strong>B. Implement Gemini Enterprise for Customer Experience to qualify leads and route them to the appropriate sales representatives.</strong></p><p><strong>C. Implement the Gemini Enterprise app with unified enterprise search and a CRM-connected agent to automate lead research and data entry.</strong></p><p><strong>D. Integrate the CRM with a popular sales intelligence platform to provide representatives with additional lead information.</strong></p><p>A is incorrect because the Cloud Natural Language API can analyze text for capabilities such as sentiment, entities, syntax, and content classification, but it is not by itself an end-to-end sales productivity solution. Solvanta Sales would still need to design, build, secure, and maintain custom orchestration for searching enterprise sources and writing updates back to the CRM. That adds development effort when an enterprise agent and connector approach more directly addresses the workflow.</p><p>B is incorrect because Gemini Enterprise for Customer Experience is focused on customer-service and contact-center experiences, including conversational agents, agent assistance, and interaction analytics. Although some concepts such as routing and conversational automation are relevant to customer engagement, this scenario is specifically about sales representatives researching leads and maintaining CRM records, making a CRM-connected enterprise productivity solution a better fit.</p><blockquote><strong><em>C is correct because the Gemini Enterprise app combines enterprise search, AI assistance, agents, and connectors to business systems. Current connectors include CRM platforms and can support natural-language interaction with sales data and workflows. A CRM-connected agent can help representatives retrieve lead context, summarize relevant information, and automate supported actions across systems, reducing repetitive research and data-entry work while keeping the sales team focused on higher-value interactions.</em></strong></blockquote><p>D is incorrect because a sales intelligence platform may enrich lead information and improve prospect research, but it does not necessarily solve the manual work of searching multiple internal sources and entering or updating CRM data. The option adds another information source rather than directly providing the cross-system AI search and agentic automation requested.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/gemini-enterprise">cloud.google.com/gemini-enterprise</a></p><p><a href="http://cloud.google.com/gemini-enterprise/connectors">cloud.google.com/gemini-enterprise/connectors</a></p><p><a href="http://docs.cloud.google.com/natural-language/docs/basics">docs.cloud.google.com/natural-language/docs/basics</a></p><p><strong>Q16.</strong></p><p><strong>Fictional company Trivexa Logistics wants a generative AI agent to automatically check real-time inventory across warehouses and adjust delivery schedules. The agent must use live data from the company’s internal inventory systems. What is the most cost-effective approach?</strong></p><p><strong>A. Build a custom API instead of using a generative AI agent.</strong></p><p><strong>B. Use pre-built generative AI chatbots to answer inventory questions.</strong></p><p><strong>C. Use Agent Studio to fine-tune a model with sample inventory data.</strong></p><p><strong>D. Use Google Cloud databases and Gemini Enterprise Agent Platform so the agent can access live inventory data.</strong></p><p>A is incorrect because a custom API can expose inventory data, and an API may be one component of the final architecture, but replacing the requested agent with a custom API alone does not provide the reasoning and workflow automation needed to interpret inventory state and adjust delivery schedules. Building the entire solution from lower-level components would also generally require more custom application logic and maintenance.</p><p>B is incorrect because a generic pre-built chatbot is optimized for conversational responses and would not automatically gain secure access to Trivexa Logistics’ live warehouse systems. Without tools, connectors, or functions that retrieve current data and perform actions, its responses could be stale and it could not reliably update delivery schedules.</p><p>C is incorrect because fine-tuning changes a model’s behavior based on training examples; it does not create a continuously updated connection to operational inventory. Training on sample inventory data would capture a snapshot and could quickly become obsolete. Real-time operational decisions require retrieval from live systems at request time rather than embedding changing inventory values into model parameters.</p><blockquote><strong><em>D is correct because the agent can use Gemini Enterprise Agent Platform together with Google Cloud data services and function calling or other tool integrations to retrieve current inventory and invoke business operations. Function calling is specifically designed to bridge a model with external systems so it can fetch up-to-date information and take actions. This keeps fast-changing inventory in the operational data layer instead of repeatedly retraining the model, making the architecture more practical and cost-effective.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tools/function-calling">docs.cloud.google.com/gemini-enterprise-agent-platform/models/tools/function-calling</a></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform">cloud.google.com/products/gemini-enterprise-agent-platform</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/agent-studio">docs.cloud.google.com/gemini-enterprise-agent-platform/agent-studio</a></p><p><strong>Q17.</strong></p><p><strong>Fictional company Belqora Retail has a vast, frequently updated ecommerce catalog. Customers struggle to find the right products, and support agents spend too much time locating product details. The company wants to improve search accuracy and efficiency for both customers and employees. Which Google Cloud approach should it use?</strong></p><p><strong>A. Agent Conversation</strong></p><p><strong>B. Cloud Natural Language API</strong></p><p><strong>C. Prebuilt retrieval-augmented generation (RAG) with Agent Search</strong></p><p><strong>D. Model Garden</strong></p><p>A is incorrect because Agent Conversation is intended for building conversational experiences, but a conversational interface alone does not solve the core requirement of accurately retrieving information from a large, frequently changing product catalog. The company needs strong indexing, retrieval, grounding, and search capabilities over its own data; conversation can be layered on top but is not the most direct answer.</p><p>B is incorrect because the Cloud Natural Language API analyzes text for capabilities such as entities, sentiment, syntax, and classification. Those capabilities can enrich text-processing workflows, but the service is not an enterprise search and retrieval system for indexing a product catalog and returning the most relevant, grounded product information to users.</p><blockquote><strong><em>C is correct because Agent Search provides Google-quality retrieval and answer generation for websites and structured or unstructured enterprise data and can serve as a grounding system for generative AI applications. A prebuilt RAG approach keeps generated responses connected to Belqora Retail’s current catalog rather than relying only on model memory. As the catalog changes, the retrieval layer can surface relevant current information, improving both customer product discovery and support-agent efficiency.</em></strong></blockquote><p>D is incorrect because Model Garden is used to discover, evaluate, customize, and deploy available Google, partner, and open models. Selecting a model is only one part of the solution and does not by itself provide the search index, information retrieval, or grounding over the company’s continuously changing catalog that the scenario requires.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform/agent-search">cloud.google.com/products/gemini-enterprise-agent-platform/agent-search</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/rag-overview">docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/rag-overview</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models">docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models</a></p><p><strong>Q18.</strong></p><p><strong>Fictional company Finvoria Advisory wants its finance team to use Gemma for routine daily work so analysts can spend more time on higher-value activities. Which business problem can Gemma most efficiently address?</strong></p><p><strong>A. Extract key financial figures and insights from varied document formats for quick reporting.</strong></p><p><strong>B. Build and deploy sophisticated internal knowledge bases that always answer finance questions with accurate and up-to-date information.</strong></p><p><strong>C. Analyze large financial transaction and market datasets to detect anomalies and predict future performance.</strong></p><p><strong>D. Efficiently produce high-quality written summaries and initial drafts of financial communications.</strong></p><p>A is incorrect because extracting reliable structured figures from many heterogeneous document formats is primarily a document-understanding and data-extraction problem. A multimodal Gemma model can reason over supported inputs and may contribute to such a workflow, but production-grade extraction from varied financial documents usually requires ingestion, parsing, validation, schema handling, and often specialized document-processing controls. It is more complex than the direct language-generation use case in option D.</p><p>B is incorrect because an accurate and continually up-to-date internal knowledge base requires more than a foundation model. It needs data ingestion, indexing, retrieval, permissions, freshness controls, and typically a grounded retrieval architecture. Gemma can be part of the generation layer, but using Gemma alone does not create or maintain such a knowledge system or guarantee current, accurate answers.</p><p>C is incorrect because large-scale anomaly detection and financial forecasting are analytical and predictive machine learning problems that depend on structured datasets, suitable statistical or machine learning methods, feature engineering, validation, and quantitative evaluation. A generative text model is not automatically the most efficient or reliable tool for this specialized numerical prediction task.</p><blockquote><strong><em>D is correct because Gemma is a family of open models designed for generative tasks including text generation, reasoning, and summarization. Drafting financial communications and summarizing existing material are direct language-generation tasks that can reduce repetitive writing effort while leaving analysts responsible for reviewing factual accuracy, tone, regulatory requirements, and final approval. This is therefore the most natural and efficient use of Gemma among the listed options.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://ai.google.dev/gemma/docs">ai.google.dev/gemma/docs</a></p><p><a href="http://ai.google.dev/gemma/docs/core/model_card_4">ai.google.dev/gemma/docs/core/model_card_4</a></p><p><a href="http://ai.google.dev/gemma/docs/capabilities/text/basic">ai.google.dev/gemma/docs/capabilities/text/basic</a></p><p><strong>Q19.</strong></p><p><strong>Fictional company Brinova Learning has a learning and development team that wants to quickly create a new-hire training video with a custom avatar and voiceover that matches the company’s branding and key messaging. The team did not receive any money to spend on production. What should it do?</strong></p><p><strong>A. Generate the video frames with Gemini 3.1 Flash Image.</strong></p><p><strong>B. Create a video with Google Vids.</strong></p><p><strong>C. Train a model with Gemini Enterprise Agent Platform and produce a video.</strong></p><p><strong>D. Prompt the Gemini app to create a video.</strong></p><p>A is incorrect because Gemini 3.1 Flash Image is an image-generation and image-editing model, not a complete workplace video-production application. Generating individual visual frames would still leave the team responsible for assembling those frames into a video, synchronizing narration, creating or maintaining an avatar, adding captions, applying branding, and editing the final sequence. That introduces unnecessary manual production work compared with a tool designed specifically for creating business videos.</p><blockquote><strong><em>B is correct because Google Vids is an AI-powered video creation application for work and directly supports the requirements in the scenario. Vids provides AI avatars, including customizable avatars, built-in voices, AI-generated voiceovers, captions, templates, stock assets, and editing tools in a single video-production workflow. Google’s current Vids product information states that users with a Google account can try preset and custom avatars at no cost, while feature availability, regional rollout, and generation limits can vary by account and plan. For a learning and development team that needs a branded training video without a separate traditional production budget, Google Vids is the most purpose-built and efficient option.</em></strong></blockquote><p>C is incorrect because Gemini Enterprise Agent Platform is intended for building, deploying, scaling, governing, and optimizing AI models and agents. Training or customizing a model merely to produce a new-hire training video would require substantially more technical effort, infrastructure, configuration, and potentially billable resources than using a ready-made video creation application. The team does not need to develop a new model to satisfy a straightforward business-video requirement.</p><p>D is incorrect because the Gemini app can generate video on supported accounts, but it is oriented toward conversational generation rather than the complete workplace video-authoring workflow described here. Google Vids is specifically designed for business video creation and brings avatars, voiceovers, scenes, branding assets, captions, and editing together in one production environment, making it the better fit for this training use case.</p><p><strong>Reference links:</strong></p><p><a href="http://workspace.google.com/products/vids/ai-avatars/">workspace.google.com/products/vids/ai-avatars/</a></p><p><a href="http://support.google.com/docs/answer/16334946">support.google.com/docs/answer/16334946</a></p><p><a href="http://support.google.com/docs/answer/15070345">support.google.com/docs/answer/15070345</a></p><p><strong>Q20.</strong></p><p><strong>Fictional company Praxilium Systems is evaluating an agentic AI architecture for workflows that require several specialized AI capabilities. What is a primary benefit of using a multi-agent system?</strong></p><p><strong>A. To manage complex tasks that demand coordinated AI functions.</strong></p><p><strong>B. To serve as a platform for hosting traditional, non-AI applications.</strong></p><p><strong>C. To simplify the most basic and repetitive rule-based tasks.</strong></p><p><strong>D. To consolidate all unique AI functions into a single, undifferentiated model.</strong></p><blockquote><strong><em>A is correct because a multi-agent system lets multiple specialized agents cooperate on a larger objective. Individual agents can be assigned different responsibilities, tools, data access, or areas of expertise, while orchestration coordinates how work is delegated and how intermediate results are combined. This architecture is particularly useful for complex enterprise tasks that naturally decompose into specialized subtasks and require coordinated reasoning or actions across multiple systems.</em></strong></blockquote><p>B is incorrect because hosting traditional applications is the role of general application infrastructure such as Cloud Run, Google Kubernetes Engine, or Compute Engine. A multi-agent architecture is specifically concerned with coordinating autonomous or semi-autonomous AI agents and their tools; using it merely as a hosting platform for ordinary non-AI applications would not provide the primary benefit of the architecture.</p><p>C is incorrect because simple, deterministic, repetitive tasks are often handled more efficiently with conventional automation, scripts, workflow engines, or a single agent. Introducing multiple agents adds orchestration, communication, observability, security, and coordination overhead. Multi-agent systems are most valuable when task complexity or specialization justifies that additional architecture.</p><p>D is incorrect because a multi-agent system does the opposite of collapsing every capability into one undifferentiated model. Its value comes from decomposing a complex workflow into distinct agents or functions that can specialize and cooperate. This modularity can improve maintainability, separation of responsibilities, and the ability to apply the appropriate tools or models to different subtasks.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/agents">docs.cloud.google.com/gemini-enterprise-agent-platform/agents</a></p><p><a href="http://docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system">docs.cloud.google.com/architecture/choose-design-pattern-agentic-ai-system</a></p><p><a href="http://cloud.google.com/products/gemini-enterprise-agent-platform">cloud.google.com/products/gemini-enterprise-agent-platform</a></p><p><strong>Q21.</strong></p><p><strong>Fictional company Tervexa Services is developing a conversational AI chatbot. The company needs the chatbot to engage in human-like conversations and provide accurate information. What should it do to enhance the chatbot’s ability to understand and respond effectively to user prompts?</strong></p><p><strong>A. Lower the model temperature setting to produce more consistent and predictable responses.</strong></p><p><strong>B. Use strict keyword matching to ensure that the chatbot only responds to specific commands.</strong></p><p><strong>C. Limit the chatbot’s training data to prevent it from learning irrelevant information.</strong></p><p><strong>D. Use prompt engineering techniques such as few-shot prompting to provide the chatbot with examples of successful interactions.</strong></p><p>A is incorrect because lowering temperature can make generation more deterministic and consistent, which can be useful when predictable wording is important. However, temperature does not teach the model what a successful conversational response should look like, how to handle representative user intents, or what interaction pattern the application expects. It therefore does not address the requirement as directly as example-based prompting.</p><p>B is incorrect because strict keyword matching is a traditional rule-based approach that limits flexibility. It can fail when users paraphrase, use synonyms, provide additional context, or phrase a request conversationally. Such rigidity works against the goal of human-like interaction and does not take advantage of the language understanding capabilities of a generative model.</p><p>C is incorrect because deliberately restricting training data is not a general technique for improving conversational understanding and can reduce the model’s coverage of language patterns and relevant knowledge. Application developers also typically do not retrain a foundation model simply to improve prompt interpretation. They use prompting, grounding, evaluation, and, when needed, targeted customization.</p><blockquote><strong><em>D is correct because few-shot prompting provides representative examples of inputs and desired responses. The model can infer patterns from those demonstrations, including tone, structure, scope, conversational behavior, and the type of information expected in a good answer. Google Cloud recommends few-shot examples as a prompt-design technique for making outputs more reliable and effective, making this the strongest option for improving the chatbot’s response behavior among the listed choices.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/few-shot-examples">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/few-shot-examples</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values</a></p><p><strong>Q22.</strong></p><p><strong>Fictional company Clyvora Commerce is using a language model to solve complex customer service inquiries. For a particular issue, the prompt includes the following instructions:</strong></p><p><strong>“To address this customer’s problem, we should first identify the core issue they are experiencing. Then we need to check if there are any known solutions or workarounds in our knowledge base. If a solution exists, we should clearly explain it to the customer. If not, we might need to escalate the issue to a specialist. Following these steps will help us provide a comprehensive and helpful response. Now, given the customer’s message: ‘My order hasn’t arrived, and the tracking number shows no updates for a week,’ what should be the next step in resolving this?”</strong></p><p><strong>What type of prompting is this?</strong></p><p><strong>A. Few-shot</strong></p><p><strong>B. Chain-of-thought</strong></p><p><strong>C. Role-based</strong></p><p><strong>D. Zero-shot</strong></p><p>A is incorrect because few-shot prompting provides multiple example input-output pairs that demonstrate how the model should perform a task. The prompt in this question does not give several solved customer-service examples. Instead, it explicitly lays out a sequence of reasoning steps that the model should follow before reaching a conclusion.</p><blockquote><strong><em>B is correct because chain-of-thought prompting encourages a model to work through a problem using intermediate reasoning steps before producing the final answer. The prompt explicitly says to identify the core issue, check for known solutions, determine whether a solution exists, and otherwise consider escalation. That structured, step-by-step reasoning pattern is the defining characteristic of chain-of-thought prompting. Modern Gemini thinking models can perform internal reasoning without always requiring explicit step-by-step instructions, but the prompt shown in the question is still a clear example of the chain-of-thought prompting technique.</em></strong></blockquote><p>C is incorrect because role-based prompting would primarily instruct the model to adopt a persona or professional role, such as “act as an experienced ecommerce support specialist.” The prompt may concern customer service, but it does not chiefly define a persona. Its defining feature is the ordered reasoning process the model is told to follow.</p><p>D is incorrect because zero-shot prompting asks the model to perform a task without providing examples, but simply lacking examples does not make zero-shot the best description when another more specific technique is clearly present. Here, the model is given explicit intermediate reasoning steps, so chain-of-thought is the more precise classification.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/glossary">docs.cloud.google.com/gemini-enterprise-agent-platform/resources/glossary</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies</a></p><p><a href="http://cloud.google.com/discover/what-is-prompt-engineering">cloud.google.com/discover/what-is-prompt-engineering</a></p><p><strong>Q23.</strong></p><p><strong>Fictional company Aerivon Goods has a marketing team that wants to use a generative AI model to create product descriptions for a new line of eco-friendly water bottles. The team provides the prompt, “Write a product description for our new water bottle.” The model produces a generic description that is factually accurate but lacks engaging language and does not highlight the environmental benefits that are central to the brand. What should the team do?</strong></p><p><strong>A. Lower the temperature setting of the model to produce more consistent results.</strong></p><p><strong>B. Increase the token count for the model to allow for longer descriptions.</strong></p><p><strong>C. Add details to the prompt about the audience, tone, and keywords.</strong></p><p><strong>D. Train the model on a dataset of marketing materials from other eco-friendly brands.</strong></p><p>A is incorrect because lowering temperature makes responses more deterministic and less varied, but it does not supply the missing marketing context. The current problem is not excessive randomness; it is an underspecified prompt. A more predictable generic description would still fail to emphasize the target audience, desired tone, and environmental differentiators.</p><p>B is incorrect because allowing more output tokens only permits the model to generate a longer response. Length does not guarantee better relevance or stronger brand alignment. Without more specific instructions, the model could simply generate a longer generic description that still misses the environmental messaging.</p><blockquote><strong><em>C is correct because effective prompt engineering requires clear context, objectives, constraints, and audience information. The team should specify who the description is for, the desired tone, the environmental benefits to emphasize, important product facts, keywords, brand style, and the intended output format or length. Giving the model this information reduces ambiguity and makes it much more likely that the output will be engaging and aligned with the company’s marketing goals.</em></strong></blockquote><p>D is incorrect because model training or fine-tuning is a substantially heavier intervention than the scenario requires. The team has not established that the model lacks the capability to produce good marketing copy; the initial prompt is simply too vague. Prompt refinement should be tried first because it is faster, cheaper, easier to evaluate, and avoids the legal, brand, and quality concerns associated with collecting and training on other companies’ marketing material.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/system-instruction-introduction">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/system-instruction-introduction</a></p><p><a href="http://cloud.google.com/discover/what-is-prompt-engineering">cloud.google.com/discover/what-is-prompt-engineering</a></p><p><strong>Q24.</strong></p><p><strong>Fictional company Novelyra News uses a large language model to automatically generate summaries of news articles for its website. A summary of an international summit is generally accurate, but the model hallucinates a detail that did not occur. How should the company overcome this hallucination?</strong></p><p><strong>A. Fine-tune the model on a larger dataset of news articles.</strong></p><p><strong>B. Use grounding to base the model output on the source articles.</strong></p><p><strong>C. Implement stricter safety settings to filter out potentially controversial topics.</strong></p><p><strong>D. Increase the temperature setting of the model to encourage more diverse outputs.</strong></p><p>A is incorrect because fine-tuning can adapt a model’s behavior, style, or performance for particular task patterns, but training on more news articles does not guarantee that a specific summary will stay faithful to the exact source article being summarized. The application already has an authoritative source for the facts: the article itself. Connecting generation to that source is more direct and easier to keep current than changing model parameters.</p><blockquote><strong><em>B is correct because grounding ties the generated response to authoritative source information supplied or retrieved for the request. Novelyra News should ground the summary in the source article so the model bases factual claims on the actual content instead of relying on unsupported assumptions or general model knowledge. Grounding does not eliminate every possible error, but it materially reduces hallucination risk and makes factual verification easier.</em></strong></blockquote><p>C is incorrect because safety settings are designed to detect or block categories of potentially harmful content. A fabricated factual detail can be completely noncontroversial and still be wrong. Increasing safety filtering therefore does not address the underlying problem of factual faithfulness to the source article.</p><p>D is incorrect because increasing temperature introduces greater randomness into generation. That can be useful for brainstorming or creative writing, but a factual news summary should prioritize accuracy and consistency. Higher randomness could make unsupported details more likely rather than correcting the hallucination.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/grounding">docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/grounding</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/ground-responses-using-rag">docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/ground-responses-using-rag</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values">docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/adjust-parameter-values</a></p><p><strong>Q25.</strong></p><p><strong>Fictional company Virelix Retail has a large online catalog and wants to improve customer experience and drive sales by implementing multimodal search capabilities using images, voice, and text. What is a primary business benefit of this capability?</strong></p><p><strong>A. Reduced dependency on keyword optimization for product listings and improved search engine rankings.</strong></p><p><strong>B. Streamlined inventory management processes and more accurate demand forecasting for popular items.</strong></p><p><strong>C. Lowered operational costs associated with managing and updating product information across different platforms and channels.</strong></p><p><strong>D. Improved customer engagement and product discovery leading to increased satisfaction and potential sales.</strong></p><p>A is incorrect because multimodal search can reduce a shopper’s dependence on exact keywords inside the retailer’s own search experience, but it does not primarily exist to improve external search engine optimization or public search rankings. The major value is helping shoppers discover relevant products more naturally through different input modes.</p><p>B is incorrect because inventory management and demand forecasting are separate retail analytics and supply-chain use cases. Search interaction data may eventually contribute to business insights, but multimodal product search does not inherently forecast demand or streamline warehouse and inventory operations.</p><p>C is incorrect because a managed search service may reduce some technical implementation burden, but multimodal search does not primarily solve the operational problem of maintaining product information across every channel and platform. Catalog governance and content synchronization require their own data-management processes.</p><blockquote><strong><em>D is correct because multimodal search makes product discovery more intuitive by allowing shoppers to express intent in the mode that is most convenient, including natural-language text, images, or voice. A customer can search using a photo, describe a desired product conversationally, or use spoken input rather than guessing the retailer’s exact keywords. Better intent understanding and more relevant discovery can reduce search friction, improve engagement and satisfaction, and increase the likelihood of clicks, conversions, and sales.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/retail/docs/what-is-it">docs.cloud.google.com/retail/docs/what-is-it</a></p><p><a href="http://docs.cloud.google.com/retail/docs/how-it-works">docs.cloud.google.com/retail/docs/how-it-works</a></p><p><a href="http://cloud.google.com/use-cases/ai-in-retail">cloud.google.com/use-cases/ai-in-retail</a></p><p><strong>Q26.</strong></p><p><strong>Fictional company Quoralis Operations is identifying tasks that could be automated with generative AI. According to Google-recommended practices, when should generative AI be used to automate tasks?</strong></p><p><strong>A. When tasks are highly creative and require original thought.</strong></p><p><strong>B. When tasks are complex and require strategic decision-making.</strong></p><p><strong>C. When tasks are repetitive and rule-based.</strong></p><p><strong>D. When tasks involve sensitive information or require human oversight.</strong></p><p>A is incorrect because generative AI can assist with creative work, such as brainstorming or drafting, but highly creative activities that depend on original judgment, brand direction, cultural nuance, or final artistic responsibility often benefit from human leadership. Such work is usually better augmented by AI than completely automated.</p><p>B is incorrect because complex strategic decisions require business context, accountability, tradeoff analysis, judgment, and often human responsibility for consequences. Generative AI can summarize information, explore scenarios, and support decision-makers, but a task that fundamentally requires strategic judgment is not the clearest candidate for end-to-end automation.</p><blockquote><strong><em>C is correct because repetitive and well-defined tasks are strong candidates for automation. When a workflow has recurring inputs, predictable patterns, clear rules, measurable outcomes, and manageable risk, generative AI or other automation techniques can reduce manual effort and allow employees to spend more time on higher-value work. Organizations should still evaluate whether generative AI is actually necessary or whether deterministic automation would be simpler for a strictly rule-based process.</em></strong></blockquote><p>D is incorrect because sensitive tasks or tasks requiring human oversight demand additional governance rather than automatic removal of human involvement. Generative AI can assist with such work, but privacy, security, compliance, explainability, and human review requirements may make full automation inappropriate. The presence of sensitive information is therefore not, by itself, a reason to automate a task.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/docs/ai-ml/generative-ai/evaluate-define-generative-ai-use-case">docs.cloud.google.com/docs/ai-ml/generative-ai/evaluate-define-generative-ai-use-case</a></p><p><a href="http://cloud.google.com/gemini-enterprise/business">cloud.google.com/gemini-enterprise/business</a></p><p><a href="http://ai.google/principles/">ai.google/principles/</a></p><p><strong>Q27.</strong></p><p><strong>Fictional company Lyradon Group wants to choose a generative AI use case that is likely to succeed and create the greatest business impact. According to Google Cloud-recommended practices, what key factor should the company determine first?</strong></p><p><strong>A. The number of employees who will be trained to use the new generative AI tools.</strong></p><p><strong>B. The availability of pre-trained models offered on various cloud computing platforms.</strong></p><p><strong>C. The frequency of updates to the underlying foundation models used by different generative AI platforms.</strong></p><p><strong>D. The specific business problems the company aims to solve and the desired outcomes.</strong></p><p>A is incorrect because employee training and change management are important for successful adoption, but the organization first needs to know what business problem it is trying to solve. Training people on a tool without a defined business objective can create activity without measurable value and makes it difficult to determine whether the initiative succeeded.</p><p>B is incorrect because model availability is a technology-selection consideration that comes after the business problem and success criteria have been defined. Starting with a catalog of available models can encourage technology-first experimentation that is disconnected from measurable organizational value.</p><p>C is incorrect because the release cadence of foundation models can affect long-term platform planning and evaluation, but it should not be the first factor in selecting a use case. A company should not choose a business initiative simply because a model family is updated frequently; it should choose a problem whose solution can generate meaningful and measurable outcomes.</p><blockquote><strong><em>D is correct because Google Cloud recommends beginning with clear, measurable business goals or needs and working backward from desired outcomes. Lyradon Group should identify the specific problem, affected users, expected business value, success criteria, risks, constraints, and feasibility before selecting technology. This business-value-first approach makes it possible to prioritize high-impact use cases and measure whether generative AI actually improves efficiency, customer experience, revenue, cost, or another desired result.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/docs/ai-ml/generative-ai/evaluate-define-generative-ai-use-case">docs.cloud.google.com/docs/ai-ml/generative-ai/evaluate-define-generative-ai-use-case</a></p><p><a href="http://cloud.google.com/discover/what-is-enterprise-ai">cloud.google.com/discover/what-is-enterprise-ai</a></p><p><a href="http://cloud.google.com/transform/how-to-build-an-effective-ai-strategy">cloud.google.com/transform/how-to-build-an-effective-ai-strategy</a></p><p><strong>Q28.</strong></p><p><strong>Fictional company Norevixa Sales wants to create dynamic and personalized video pitches for potential clients. The sales team receives client information in various formats and needs an AI model that can turn the information into engaging video content tailored to each client’s needs and challenges. Which current Google model should the company use?</strong></p><p><strong>A. Gemma</strong></p><p><strong>B. Gemini 3.5 Flash</strong></p><p><strong>C. Gemini 3.1 Flash Image</strong></p><p><strong>D. Veo 3.1</strong></p><p>A is incorrect because Gemma is a family of open models intended for flexible developer use across supported language, reasoning, coding, and multimodal scenarios, but it is not Google’s dedicated video-generation model line. The required output is a generated video pitch, so a model explicitly designed for video generation is a better fit.</p><p>B is incorrect because Gemini 3.5 Flash is a broad Gemini model suitable for reasoning, understanding multimodal information, and many application tasks, but it is not the dedicated video-generation model listed here. Gemini could help analyze the supplied customer context or draft a video concept, but the question asks for the model that should create the actual video content.</p><p>C is incorrect because Gemini 3.1 Flash Image is optimized for image understanding and generation. It can create still images from prompts and support image-editing workflows, but the requested output is a personalized moving video rather than a still visual.</p><blockquote><strong><em>D is correct because Veo 3.1 is Google’s current video-generation model line and supports text-to-video and image-to-video generation, including high-quality video output and audio generation. A workflow can use client information to construct an appropriate prompt or supporting assets and then use Veo to generate tailored video content. This directly matches the requirement to create dynamic video pitches rather than static images or text.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate">docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3–1-generate</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-text">docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-text</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions">docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions</a></p><p><strong>Q29.</strong></p><p><strong>Fictional company Braventa Retail has fragmented phone, email, and basic website-chat support. It needs a unified cloud solution with integrated communication channels, consistent customer experiences, scalable support, and enterprise security and privacy. Which Google Cloud offering should it use?</strong></p><p><strong>A. Gemini Enterprise Agent Platform</strong></p><p><strong>B. Google Cloud Contact Center as a Service (CCaaS)</strong></p><p><strong>C. Conversational AI</strong></p><p><strong>D. Agent Search on Gemini Enterprise Agent Platform</strong></p><p>A is incorrect because Gemini Enterprise Agent Platform is the developer platform for building, deploying, governing, and optimizing AI models and agents. It can power sophisticated customer-experience solutions and integrate with other systems, but it does not by itself provide the complete contact-center operational layer for handling and routing customer interactions across voice and digital channels.</p><blockquote><strong><em>B is correct because Google Cloud Contact Center as a Service is a cloud-native, omnichannel contact-center platform designed to queue and route customer interactions across voice and digital channels, integrate with CRM systems, support human and virtual agents, and provide a consistent customer-service environment. It is specifically intended to provide the unified operational contact-center foundation described in the question and is part of Google Cloud’s current Gemini Enterprise for Customer Experience portfolio.</em></strong></blockquote><p>C is incorrect because conversational AI is a broad technology category rather than the complete Google Cloud contact-center platform required by this scenario. Conversational agents can automate customer interactions and answer questions, but a complete contact center also requires capabilities such as channel integration, queuing, routing, agent operations, escalation, reporting, and enterprise administration.</p><p>D is incorrect because Agent Search provides enterprise information retrieval, search, answer generation, and grounding over configured data sources. It could improve self-service or help a support agent find information, but it does not provide omnichannel telephony, email and chat routing, contact-center queues, agent management, or the full operational infrastructure needed here.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/gemini-enterprise-cx">cloud.google.com/gemini-enterprise-cx</a></p><p><a href="http://cloud.google.com/solutions/contact-center-as-a-service">cloud.google.com/solutions/contact-center-as-a-service</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-cx/cx-agent-studio/deploy/ccaas">docs.cloud.google.com/gemini-enterprise-cx/cx-agent-studio/deploy/ccaas</a></p><p><strong>Q30.</strong></p><p><strong>Fictional company Cendralis Support is developing a system to automatically categorize customer support emails. It has thousands of past emails, and each email was manually reviewed and assigned a category such as “Billing Inquiry,” “Technical Support,” or “Feature Request.” What type of data is this?</strong></p><p><strong>A. Unlabeled data</strong></p><p><strong>B. Labeled data</strong></p><p><strong>C. Structured data</strong></p><p><strong>D. Raw data</strong></p><p>A is incorrect because unlabeled data contains input information without a known target or category attached to each example. In this scenario, every email already has a human-assigned category indicating the desired output, so the dataset is not unlabeled.</p><blockquote><strong><em>B is correct because labeled data consists of examples that include both the input and the expected label or target. Here, the email text is the input and categories such as “Billing Inquiry,” “Technical Support,” and “Feature Request” are the labels. These input-label pairs are exactly the kind of data commonly used to train and evaluate supervised classification models.</em></strong></blockquote><p>C is incorrect because structured data refers to information organized according to a predefined schema, commonly in rows, columns, and typed fields. The category field itself may be structured, but the email body is typically unstructured text. More importantly, the question asks about the fact that each email has been assigned a known category, which specifically makes the examples labeled data.</p><p>D is incorrect because raw data generally refers to data in its original or minimally processed state. These emails have already undergone human review and annotation, and each has been assigned a target category. That annotation step adds semantic information that turns the examples into labeled data for a supervised learning task.</p><p><strong>Reference links:</strong></p><p><a href="http://developers.google.com/machine-learning/glossary/fundamentals">developers.google.com/machine-learning/glossary/fundamentals</a></p><p><a href="http://developers.google.com/machine-learning/glossary">developers.google.com/machine-learning/glossary</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=888830d2ef94" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Associate Cloud Engineer Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-associate-cloud-engineer-practice-questions-f3cd266f7116?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3cd266f7116</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[gcp]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 02:41:06 GMT</pubDate>
            <atom:updated>2026-10-07T02:41:06.608Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z2CmOG7Vxz4eNHNSPoQsZw.jpeg" /><figcaption>KloudExams GCP Google Cloud Associate Cloud Engineer Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Associate Cloud Engineer certification</strong>? Knowing what a cloud service does is a starting point — understanding when and why to use it is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing solutions that fit specific requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>Braventra Logistics needs to enable traffic between multiple groups of Compute Engine instances that are currently running in two different Google Cloud projects. Each group of Compute Engine instances is running in its own VPC network. What should you do?</strong></p><p><strong>A. Verify that both projects are in a Google Cloud organization. Create a new VPC and add all instances.</strong></p><p><strong>B. Verify that both projects are in a Google Cloud organization. Share the VPC from one project and request that the Compute Engine instances in the other project use this Shared VPC.</strong></p><p><strong>C. Verify that you are the Project Administrator of both projects. Create two new VPCs and add all instances.</strong></p><p><strong>D. Verify that you are the Project Administrator of both projects. Create a new VPC and add all instances.</strong></p><p>A is incorrect because a normal VPC network belongs to a single project. Simply creating a new standalone VPC does not allow instances owned by another project to attach to it. Cross-project use of a common VPC network requires the Shared VPC model, in which a host project owns the network and attached service projects are permitted to use shared subnets. Therefore, organization membership alone plus a new ordinary VPC does not solve the cross-project networking requirement.</p><blockquote><strong><em>B is correct because Shared VPC is designed to let resources in multiple projects within the same Google Cloud organization use a common VPC network and communicate by using internal IP addresses, subject to firewall rules. One project is enabled as the Shared VPC host project and the other is attached as a service project, after which eligible resources in the service project can use subnets shared from the host project. A current operational caveat is that attaching an existing project does not automatically move already-created VMs onto the Shared VPC; existing VM network interfaces cannot simply be switched to another VPC, so affected instances generally need to be recreated or migrated to use shared subnets. Even with that migration requirement, option B is the only listed design that provides a valid cross-project common VPC architecture.</em></strong></blockquote><p>C is incorrect because creating two additional VPC networks would still leave the workloads on separate networks. Separate VPCs do not gain connectivity merely because the same administrator controls both projects. You would need a supported connectivity mechanism such as VPC Network Peering, Cloud VPN, or a common Shared VPC, depending on requirements. None of those connectivity mechanisms is established by this option.</p><p>D is incorrect because an ordinary VPC network cannot directly contain VM network interfaces from multiple independent projects. Having project-level administrative permissions does not change the ownership scope of a standalone VPC. Shared VPC is the Google Cloud feature that explicitly allows service-project resources to consume subnets from a host project’s VPC. Therefore, simply creating one new VPC while remaining outside Shared VPC does not satisfy the requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/vpc/docs/shared-vpc">cloud.google.com/vpc/docs/shared-vpc</a></p><p><a href="http://cloud.google.com/vpc/docs/provisioning-shared-vpc">cloud.google.com/vpc/docs/provisioning-shared-vpc</a></p><p><strong>Q2.</strong></p><p><strong>Elvaris Software has a development environment in a Google Cloud project that uses Compute Engine and Cloud SQL. You need to create a production environment. Your security requirements prohibit network routes between the development and production environments, and you want to follow Google-recommended practices. What should you do?</strong></p><p><strong>A. Create a new project for production, enable the Compute Engine and Cloud SQL APIs, and replicate the required production configuration in that project.</strong></p><p><strong>B. Create a production subnet in the existing VPC network and create the production Compute Engine and Cloud SQL resources in the existing project.</strong></p><p><strong>C. Create a new project for production and connect it to the existing VPC network by using Shared VPC.</strong></p><p><strong>D. Obtain the Editor basic role (roles/editor) in an existing production project owned by another division and replicate the environment there.</strong></p><blockquote><strong><em>A is correct because separate Google Cloud projects provide strong administrative, policy, quota, billing, and resource-management boundaries. Creating the production resources in a new project with its own VPC network also avoids the automatic VPC routes that would exist between subnets of the same VPC. This cleanly separates development from production and follows the common Google Cloud practice of using separate projects for environments that require isolation. Appropriate IAM, firewall, organization-policy, and connectivity controls can then be configured independently for production.</em></strong></blockquote><p>B is incorrect because subnets in the same VPC network are connected by automatically created subnet routes. Unless additional controls are applied, resources in the development and production subnets can have network reachability to each other. The requirement explicitly prohibits network routes between the environments, so putting both environments in the same VPC does not satisfy the requested isolation model.</p><p>C is incorrect because Shared VPC is designed to let resources in multiple projects use a common VPC network. Service-project resources using shared subnets participate in that network’s routing domain, so this design intentionally creates shared network connectivity rather than prohibiting routes between the development and production environments. It therefore conflicts with the stated security requirement.</p><p>D is incorrect because reusing another division’s production project reduces administrative isolation and grants a broad legacy basic role that is not justified by the requirement. The Editor role provides extensive modification permissions across many services, violating least privilege, and placing unrelated workloads in the same project can complicate ownership, policy, quotas, and billing. Creating a dedicated production project is cleaner and safer.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/creating-managing-projects">cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><a href="http://cloud.google.com/docs/get-started/organize-resources">cloud.google.com/docs/get-started/organize-resources</a></p><p><a href="http://cloud.google.com/vpc/docs/routes">cloud.google.com/vpc/docs/routes</a></p><p><strong>Q3.</strong></p><p><strong>Fyntera Holdings has enabled a domain-restricted sharing organization policy so that IAM access can be granted only to identities in its Cloud Identity domain. An external auditor needs temporary read-only access to review all resources in one Google Cloud project. What should you do?</strong></p><p><strong>A. Ask the auditor to use an external Google Account and grant that account the Viewer basic role (roles/viewer) on the project.</strong></p><p><strong>B. Ask the auditor to use an external Google Account and grant that account the Security Reviewer role (roles/iam.securityReviewer) on the project.</strong></p><p><strong>C. Create a temporary account for the auditor in the organization’s Cloud Identity domain and grant that account the Viewer basic role (roles/viewer) on the project.</strong></p><p><strong>D. Create a temporary account for the auditor in the organization’s Cloud Identity domain and grant that account the Security Reviewer role (roles/iam.securityReviewer) on the project.</strong></p><p>A is incorrect because the domain-restricted sharing policy prevents IAM policy bindings from being granted to principals that are outside the allowed organization or domain. An external consumer Google Account would therefore be blocked by the policy even though the Viewer role itself provides broad read-only access. The auditor needs an identity that is permitted by the organization’s domain restriction.</p><p>B is incorrect because the external Google Account would still violate the domain-restricted sharing constraint. In addition, the Security Reviewer role is focused on viewing IAM policies and security-related configuration; it does not provide broad read access to all project resources that an auditor needs for the stated review.</p><blockquote><strong><em>C is correct because an account created in the organization’s managed Cloud Identity domain can satisfy the domain-restricted sharing policy. Granting that temporary managed identity the Viewer basic role at the project level provides broad read-only access to project resources without granting modification permissions. The account can then be disabled or deleted after the audit. Google Cloud generally recommends more granular predefined roles when practical, but among the supplied choices Viewer is the role that best matches a requirement to inspect all project resources without modifying them.</em></strong></blockquote><p>D is incorrect because creating a managed Cloud Identity account correctly addresses the domain restriction, but the Security Reviewer role does not provide general read access to all project resources. It is intended primarily for reviewing IAM policies and security configuration. Therefore, the auditor could still lack visibility into resources that are within the scope of the requested project-wide audit.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/organization-policy/domain-restricted-sharing">cloud.google.com/organization-policy/domain-restricted-sharing</a></p><p><a href="http://cloud.google.com/iam/docs/roles-overview">cloud.google.com/iam/docs/roles-overview</a></p><p><a href="http://cloud.google.com/iam/docs/choose-predefined-roles">cloud.google.com/iam/docs/choose-predefined-roles</a></p><p><strong>Q4.</strong></p><p><strong>Gryphonex Commerce runs a critical workload on a Compute Engine VM. You need to back up the boot disk regularly, restore it quickly when necessary, and automatically remove older backups to reduce cost. You want to follow Google-recommended practices. What should you do?</strong></p><p><strong>A. Create a Cloud Run function that periodically creates an instance template from the VM.</strong></p><p><strong>B. Create a snapshot schedule for the boot disk with the desired interval and retention policy.</strong></p><p><strong>C. Configure a cron job that periodically uses the gcloud CLI to create a new persistent disk from the boot disk.</strong></p><p><strong>D. Create a Cloud Tasks task that periodically creates a machine image and exports it to Cloud Storage.</strong></p><p>A is incorrect because an instance template stores VM configuration such as machine type, disks to attach, metadata, and networking settings; it is not a point-in-time backup of the changing contents of a boot disk. Repeatedly creating instance templates would therefore not provide the disk-data backup and restore behavior required by the scenario.</p><blockquote><strong><em>B is correct because Compute Engine snapshot schedules are the managed mechanism for taking recurring snapshots of persistent disks. A snapshot schedule can define the backup frequency and a retention policy that automatically deletes older scheduled snapshots. Snapshots are incremental after the first snapshot and can be used to create replacement disks for recovery, providing a straightforward and cost-efficient solution without custom scheduling code.</em></strong></blockquote><p>C is incorrect because repeatedly creating complete persistent disks as backups would consume more storage and require custom scheduling and cleanup logic. It also does not automatically implement a lifecycle or retention policy unless additional scripting is added. Managed snapshot schedules provide purpose-built incremental backup behavior with automatic retention management and require less operational effort.</p><p>D is incorrect because Cloud Tasks is an asynchronous task-delivery service rather than a backup scheduler designed to manage VM disk retention. Machine images can be useful when you need to capture broader VM configuration together with disk data, but creating and exporting them through a custom Cloud Tasks workflow is unnecessarily complex for regular boot-disk backups. Snapshot schedules directly meet the requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/compute/docs/disks/about-snapshot-schedules">cloud.google.com/compute/docs/disks/about-snapshot-schedules</a></p><p><a href="http://cloud.google.com/compute/docs/disks/scheduled-snapshots">cloud.google.com/compute/docs/disks/scheduled-snapshots</a></p><p><a href="http://cloud.google.com/compute/docs/disks/create-snapshots">cloud.google.com/compute/docs/disks/create-snapshots</a></p><p><strong>Q5.</strong></p><p><strong>Jendara Media wants to reduce costs by turning off all configured Google Cloud services and resources in an existing project using the fewest steps. What should you do?</strong></p><p><strong>A. Verify that you have the Owner basic role (roles/owner) or equivalent project-deletion permission, open the project in the Google Cloud console, choose Shut down, and enter the project ID to confirm.</strong></p><p><strong>B. Verify that you have the Owner basic role (roles/owner), and manually delete every resource in the project one service at a time.</strong></p><p><strong>C. Verify that you have the Organization Administrator role (roles/resourcemanager.organizationAdmin), open the project in the Google Cloud console, enter the project ID, and choose Shut down.</strong></p><p><strong>D. Verify that you have the Organization Administrator role (roles/resourcemanager.organizationAdmin), and manually delete every resource in the project one service at a time.</strong></p><blockquote><strong><em>A is correct because shutting down a project is the fewest-step way to stop use of the project’s Google Cloud resources when the entire project is no longer required. Project shutdown marks the project for deletion and initiates the resource cleanup process. The caller needs the resourcemanager.projects.delete permission; the Owner basic role includes this permission, although a more narrowly scoped project-deletion role can be preferable under least privilege. The console requires confirmation of the project before shutdown.</em></strong></blockquote><p>B is incorrect because manually deleting resources from every service is time-consuming and error-prone, especially in a project containing many resource types. Some dependencies can also make deletion order complex. If the objective is to turn off the entire project and all of its configured services with the fewest steps, project shutdown is the appropriate administrative operation.</p><p>C is incorrect because Organization Administrator is not the role specifically required to delete a project. Organization-level administration and project deletion are distinct permissions, and holding an organization-administration role does not automatically mean the principal has resourcemanager.projects.delete on the target project. The project shutdown should be performed by a principal that has the required project-deletion permission.</p><p>D is incorrect because Organization Administrator is both unnecessarily broad for the task and does not make manual deletion the best approach. Deleting service resources individually requires many operations and increases the chance that billable or dependent resources will be missed. Shutting down the project directly is simpler.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/delete-restore-projects">cloud.google.com/resource-manager/docs/delete-restore-projects</a></p><p><a href="http://cloud.google.com/iam/docs/roles-permissions/resourcemanager">cloud.google.com/iam/docs/roles-permissions/resourcemanager</a></p><p><strong>Q6.</strong></p><p><strong>Cendrixa Media has a Compute Engine instance hosting a production application. You want to receive an email if the instance consumes more than 90% of its CPU resources for more than 15 minutes. You want to use Google services. What should you do?</strong></p><p><strong>A. Create a consumer Gmail account. Write a script that monitors CPU usage. When CPU usage exceeds the threshold, have the script send an email through smtp.gmail.com on port 25.</strong></p><p><strong>B. Use a Cloud Monitoring scoping project whose metrics scope includes the Google Cloud project. Create a Cloud Monitoring alerting policy that triggers when VM CPU utilization exceeds 90% for 15 minutes. Configure your email address as a notification channel.</strong></p><p><strong>C. Use a Cloud Monitoring scoping project whose metrics scope includes the Google Cloud project. Write a script that monitors CPU usage and sends the values as a custom metric to Cloud Monitoring. Create an uptime check for the instance in Cloud Monitoring.</strong></p><p><strong>D. In Cloud Logging, create a logs-based metric that extracts CPU usage by using the regular expression CPU Usage: ([0–9]{1,3})%. In Cloud Monitoring, create an alerting policy based on this metric and configure your email address as a notification channel.</strong></p><p>A is incorrect because Compute Engine already exposes CPU utilization as a built-in metric that Cloud Monitoring can evaluate directly, so maintaining a custom polling and email script is unnecessary. The script would become another component to secure, monitor, and operate. In addition, direct outbound SMTP on TCP port 25 from Compute Engine is commonly restricted, making this design even less appropriate. Cloud Monitoring provides a managed alerting and notification mechanism that directly matches the requirement.</p><blockquote><strong><em>B is correct because Cloud Monitoring collects the Compute Engine CPU utilization metric automatically and can evaluate it with a metric-threshold alerting policy. You can configure the condition so that CPU utilization must remain above 90% for a 15-minute retest window before an incident is opened, preventing a short CPU spike from triggering the notification. An email notification channel can then send the alert to the required address. Current Cloud Monitoring uses metrics scopes; a project’s metrics scope includes its own metrics by default and can also include additional projects when centralized monitoring is needed.</em></strong></blockquote><p>C is incorrect because there is no need to create and publish a custom CPU metric when Compute Engine CPU utilization is already available as a system metric. An uptime check tests whether an endpoint is reachable and responsive; it does not determine whether VM CPU utilization has exceeded 90% for 15 minutes. This approach therefore adds unnecessary custom code and uses the wrong monitoring feature for the threshold requirement.</p><p>D is incorrect because logs-based metrics are appropriate when the signal of interest exists in log entries, but VM CPU utilization is already available directly as a Cloud Monitoring metric. Reconstructing CPU usage from application or system log text would be less reliable, require suitable log messages to exist, and add needless parsing. A direct metric-threshold condition is simpler and more accurate.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/monitoring/alerts">cloud.google.com/monitoring/alerts</a></p><p><a href="http://cloud.google.com/monitoring/settings">cloud.google.com/monitoring/settings</a></p><p><a href="http://cloud.google.com/monitoring/support/notification-options">cloud.google.com/monitoring/support/notification-options</a></p><p><strong>Q7.</strong></p><p><strong>Navorex Software built an application on a development laptop that uses Google Cloud services. The application uses Application Default Credentials for authentication and works correctly on the laptop. You want to migrate the application to a Compute Engine VM and configure authentication using Google-recommended practices with minimal application changes. What should you do?</strong></p><p><strong>A. Grant the appropriate Google Cloud IAM roles to the service account attached to the Compute Engine VM.</strong></p><p><strong>B. Create a service account with appropriate access to Google Cloud services and explicitly configure the application to use that account.</strong></p><p><strong>C. Store service-account credentials with appropriate Google Cloud access in a configuration file and deploy that file with the application.</strong></p><p><strong>D. Store your user-account credentials with appropriate Google Cloud access in a configuration file and deploy that file with the application.</strong></p><blockquote><strong><em>A is correct because Application Default Credentials automatically adapts to the environment in which the application runs. On a Compute Engine VM, ADC can obtain short-lived credentials from the metadata server for the service account attached to that VM, so application code that already uses ADC normally requires no credential-specific change. The recommended approach is to attach a suitable user-managed service account where appropriate and grant it only the IAM roles the workload needs. This avoids distributing long-lived keys and preserves the application’s existing ADC-based authentication flow.</em></strong></blockquote><p>B is incorrect because creating a dedicated service account is often a good practice, but the option says to configure the application explicitly to use the account. If that means changing the application’s credential-loading behavior or distributing a key, it is unnecessary because ADC already discovers the VM’s attached service account automatically. The minimal-change solution is to use the service account attached to the VM and assign that identity the required roles.</p><p>C is incorrect because storing service-account credentials in a configuration file usually means distributing a long-lived service-account key. Google recommends avoiding service-account keys whenever a workload can use an attached service identity and short-lived credentials instead. A key file creates secret-storage, rotation, leakage, and revocation risks and is unnecessary on Compute Engine.</p><p>D is incorrect because personal user credentials should not be embedded in a production workload. They couple application availability to an individual person’s account, create serious secret-management risks, and make auditing and offboarding harder. Workloads should use workload identities such as service accounts rather than a developer’s user credentials.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/docs/authentication/application-default-credentials">cloud.google.com/docs/authentication/application-default-credentials</a></p><p><a href="http://cloud.google.com/compute/docs/access/service-accounts">cloud.google.com/compute/docs/access/service-accounts</a></p><p><a href="http://cloud.google.com/iam/docs/best-practices-service-accounts">cloud.google.com/iam/docs/best-practices-service-accounts</a></p><p><strong>Q8.</strong></p><p><strong>Orlentra Systems needs to create a Compute Engine instance in a new Google Cloud project that does not yet exist. What should you do?</strong></p><p><strong>A. Use the Google Cloud CLI to create the new project, enable the Compute Engine API in that project, and then create the VM while specifying the new project.</strong></p><p><strong>B. Enable the Compute Engine API in the Google Cloud console, use the Google Cloud CLI to create the VM, and use the — project flag to specify a new project that does not yet exist.</strong></p><p><strong>C. Use the Google Cloud CLI to create the VM with the — project flag specifying the nonexistent project, and answer yes when prompted to enable the Compute Engine API.</strong></p><p><strong>D. Enable the Compute Engine API in the Google Cloud console. Open the Compute Engine page, start creating a VM, and select a Create In A New Project option in the VM creation form.</strong></p><blockquote><strong><em>A is correct because Google Cloud resources must belong to an existing project, and a service API must be enabled in that project before resources for that service can be created. The correct sequence is therefore to create the project, enable compute.googleapis.com for it, and then run the Compute Engine instance-creation command with that project selected or supplied through the — project flag. This explicit order establishes the resource container and activates the required service before VM creation.</em></strong></blockquote><p>B is incorrect because an API is enabled within the context of an existing project. You cannot enable the Compute Engine API for a project identifier that has not yet been created and then expect the VM-creation command to create the project implicitly. The project must be created first.</p><p>C is incorrect because the — project flag tells the Google Cloud CLI which existing project to use; it does not create a missing project. The CLI can sometimes offer to enable a disabled API for an existing project, but that convenience does not replace project creation. A nonexistent project must first be created through Resource Manager.</p><p>D is incorrect because the Compute Engine VM creation form creates a VM inside the currently selected existing project. It does not provide a Create In A New Project control that simultaneously creates a Resource Manager project as part of VM creation. Project creation and API enablement are separate administrative steps that must occur before the VM can be provisioned.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/creating-managing-projects">cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><a href="http://cloud.google.com/service-usage/docs/enable-disable">cloud.google.com/service-usage/docs/enable-disable</a></p><p><a href="http://cloud.google.com/sdk/gcloud/reference/compute/instances/create">cloud.google.com/sdk/gcloud/reference/compute/instances/create</a></p><p><strong>Q9.</strong></p><p><strong>Avenquor Industries runs one batch process on an on-premises server that takes around 30 hours to complete. The task runs monthly, can be performed offline, and must be restarted if interrupted. You want to migrate this workload to Google Cloud while minimizing cost. What should you do?</strong></p><p><strong>A. Migrate the workload to a Compute Engine Spot VM.</strong></p><p><strong>B. Migrate the workload to a Google Kubernetes Engine cluster with Spot VM nodes.</strong></p><p><strong>C. Migrate the workload to a standard Compute Engine VM. Start and stop the instance as needed.</strong></p><p><strong>D. Create an instance template with the Spot provisioning model enabled. Create a managed instance group from the template and adjust target CPU utilization. Migrate the workload.</strong></p><blockquote><strong><em>A is correct because current Compute Engine Spot VMs are intended for fault-tolerant batch workloads and can provide discounts of up to 91% compared with standard on-demand VM pricing. Unlike the earlier interruptible VM model on which older versions of this question were based, a Spot VM has no fixed 24-hour maximum runtime unless you explicitly configure one. The task can be run offline, occurs only once a month, and the question states that it can be restarted if an interruption occurs, so it can tolerate the principal tradeoff of Spot capacity. A single Spot VM therefore avoids both continuous idle cost and unnecessary orchestration overhead while providing the lowest-cost compute option among the choices. The workload should still be designed so that a preemption causes a clean restart, because Spot capacity has no availability guarantee and can be reclaimed at any time.</em></strong></blockquote><p>B is incorrect because Spot VM nodes in GKE can also reduce compute cost, but Kubernetes adds a cluster control plane, node-pool configuration, workload manifests, scheduling, upgrades, and other operational components that are unnecessary for one monthly batch process. The task does not require container orchestration, horizontal scheduling across multiple Pods, or Kubernetes-specific capabilities. Using GKE solely to execute one process increases complexity and can introduce additional cost without improving the workload’s ability to tolerate Spot VM preemption.</p><p>C is incorrect because a standard Compute Engine VM that is started only for the monthly run is a reasonable choice when uninterrupted execution is more important than price, and instance scheduling can reduce idle cost. However, the question’s primary objective is to minimize cost and explicitly states that the batch process can be restarted after an interruption. Standard VMs use the standard provisioning model and therefore do not receive the large Spot discount. With current Spot VM behavior, there is also no mandatory 24-hour termination that would make a 30-hour run impossible, so the standard VM is no longer the least-cost choice under the stated requirements.</p><p>D is incorrect because a managed instance group and CPU-based autoscaling are unnecessary for a single batch task that requires one execution rather than a horizontally scaled serving fleet. A managed instance group is useful when identical VMs must be maintained, automatically recreated, or scaled according to demand. CPU autoscaling could also create or remove instances without understanding whether the single batch job has completed. This design adds orchestration complexity and does not provide a useful advantage over running the batch process directly on a Spot VM.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/compute/docs/instances/spot">cloud.google.com/compute/docs/instances/spot</a></p><p><a href="http://cloud.google.com/compute/docs/instances/create-use-spot">cloud.google.com/compute/docs/instances/create-use-spot</a></p><p><a href="http://cloud.google.com/compute/docs/instances/schedule-instance-start-stop">cloud.google.com/compute/docs/instances/schedule-instance-start-stop</a></p><p><strong>Q10.</strong></p><p><strong>Elarvix Digital needs to set up application performance monitoring for Google Cloud projects A, B, and C in a single pane of glass. You want to monitor CPU, memory, and disk metrics across the projects. What should you do?</strong></p><p><strong>A. Enable the Cloud Monitoring API and then share charts independently from projects A, B, and C.</strong></p><p><strong>B. Enable the Cloud Monitoring API and then grant the Monitoring Metric Reader role to projects A, B, and C.</strong></p><p><strong>C. Enable the Cloud Monitoring API and then use default dashboards to view all three projects one at a time.</strong></p><p><strong>D. Enable the Cloud Monitoring API, use project A as the scoping project, and add projects B and C to project A’s metrics scope.</strong></p><p>A is incorrect because sharing charts from separate projects does not create one centralized monitoring context. Each chart remains associated with its own project’s metrics scope and dashboards, which requires operators to manage separate views or manually combine information. The requirement is specifically for a single pane of glass across all three projects.</p><p>B is incorrect because the Monitoring Metric Reader role grants a principal permission to read monitoring data, but IAM access alone does not combine the metrics of separate projects into one monitoring scope. Roles determine who can access data; metrics scopes determine which projects’ metrics are visible from a particular scoping project. Therefore, granting a reader role does not itself satisfy the centralized-monitoring requirement.</p><p>C is incorrect because viewing default dashboards sequentially still requires moving from one project to another and does not create a unified operational view. Default dashboards are useful for service-specific telemetry, but the question asks to monitor the projects together from one place rather than inspect each project independently.</p><blockquote><strong><em>D is correct because Cloud Monitoring uses a metrics scope to define the set of resource containers whose metrics can be viewed and monitored from a scoping project. By using project A as the scoping project and adding projects B and C to its metrics scope, operators can create dashboards and alerting policies that use metrics from all three projects in one monitoring context. Compute Engine CPU and disk-related platform metrics are available through Monitoring, while guest-level metrics such as memory usage generally require the Ops Agent or another supported telemetry source on the VMs. The metrics-scope configuration is the part that provides the requested single pane of glass.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/monitoring/settings/multiple-projects">cloud.google.com/monitoring/settings/multiple-projects</a></p><p><a href="http://cloud.google.com/monitoring/settings">cloud.google.com/monitoring/settings</a></p><p><a href="http://cloud.google.com/monitoring/agent/ops-agent">cloud.google.com/monitoring/agent/ops-agent</a></p><p><strong>Q11.</strong></p><p><strong>Luminara Systems has embraced a hybrid cloud strategy in which some applications are deployed on Google Cloud. A Cloud VPN tunnel connects the company’s VPC network in Google Cloud with the company’s on-premises network. Multiple applications in Google Cloud need to connect to an on-premises database server, and you want to avoid changing the IP configuration in all applications when the database IP address changes. What should you do?</strong></p><p><strong>A. Configure Cloud NAT for all subnets of the VPC network to be used when egressing from the VM instances.</strong></p><p><strong>B. Create a private zone in Cloud DNS, and configure the applications with the DNS name.</strong></p><p><strong>C. Configure the IP address of the database as custom metadata for each instance, and query the metadata server.</strong></p><p><strong>D. Query the Compute Engine internal DNS from the applications to retrieve the IP address of the database.</strong></p><p>A is incorrect because Cloud NAT provides outbound address translation for resources that do not have external IP addresses. It does not provide service discovery or a stable logical name for an on-premises database whose private IP address might change. Applications would still need to know the database destination address, so Cloud NAT does not solve the configuration-maintenance problem.</p><blockquote><strong><em>B is correct because Cloud DNS can host a private DNS zone that is visible to selected VPC networks. You can create a DNS record such as database.internal.example that maps to the on-premises database’s private IP address and configure all applications to connect using that DNS name. If the database address changes, you update the DNS record once instead of modifying every application. The existing Cloud VPN provides the network path to the on-premises address, while DNS provides stable name resolution. DNS TTL values should be selected appropriately so clients adopt address changes within the required interval.</em></strong></blockquote><p>C is incorrect because storing the database IP as VM custom metadata would still require updating metadata on every relevant instance, or creating additional automation to keep the metadata synchronized. Applications would also need custom logic to query the metadata server and consume that value. A DNS name is the standard service-discovery abstraction for an endpoint whose IP can change and requires less application-specific handling.</p><p>D is incorrect because Compute Engine internal DNS automatically provides names for Google Cloud VM instances and related internal resources; it does not automatically create a DNS record for an arbitrary on-premises database server. Without explicitly configuring DNS information for the on-premises resource, the applications cannot rely on Compute Engine internal DNS to discover that address.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/dns/docs/zones">cloud.google.com/dns/docs/zones</a></p><p><a href="http://cloud.google.com/dns/docs/overview">cloud.google.com/dns/docs/overview</a></p><p><a href="http://cloud.google.com/network-connectivity/docs/vpn/concepts/overview">cloud.google.com/network-connectivity/docs/vpn/concepts/overview</a></p><p><strong>Q12.</strong></p><p><strong>Braventra Games developed a mobile game that is deployed on Google Cloud. Gamers connect to the game with their personal phones over the internet. The game sends UDP packets to update the servers about gamers’ actions while they are playing in multiplayer mode. Your game backend can scale across multiple virtual machines, and you want to expose the VMs through a single IP address. What should you do?</strong></p><p><strong>A. Configure a global external proxy Network Load Balancer for SSL proxy traffic in front of the application servers.</strong></p><p><strong>B. Configure an internal passthrough Network Load Balancer for UDP traffic in front of the application servers.</strong></p><p><strong>C. Configure an external Application Load Balancer in front of the application servers.</strong></p><p><strong>D. Configure a regional external passthrough Network Load Balancer in front of the application servers.</strong></p><p>A is incorrect because an external proxy Network Load Balancer using an SSL proxy is designed for proxying TCP connections with TLS termination. UDP is connectionless and is not supported by an SSL proxy frontend. The game traffic described in the question consists of UDP packets from internet clients, so an SSL proxy load-balancing architecture does not match the transport protocol.</p><p>B is incorrect because an internal passthrough Network Load Balancer supports UDP but provides an internal frontend that is intended for clients within the VPC network or connected private networks. The gamers are connecting from personal phones over the public internet, so they require an internet-facing external frontend rather than an internal load-balancer IP address.</p><p>C is incorrect because an external Application Load Balancer operates at Layer 7 for HTTP and HTTPS workloads. Raw UDP game traffic is not HTTP traffic and cannot be distributed by an Application Load Balancer. The workload requires a Layer 4 load balancer that supports UDP.</p><blockquote><strong><em>D is correct because a regional external passthrough Network Load Balancer is a Layer 4 internet-facing load balancer that supports UDP and distributes packets across backend Compute Engine VMs in a region. The forwarding rule provides an external frontend IP address that clients can use as a single game-server endpoint, while the load balancer distributes UDP traffic among healthy backend instances. Google Cloud also offers a global external passthrough Network Load Balancer capability, but the regional external passthrough Network Load Balancer described in this choice remains the correct match among the supplied options.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/load-balancing/docs/passthrough-network-load-balancer">cloud.google.com/load-balancing/docs/passthrough-network-load-balancer</a></p><p><a href="http://cloud.google.com/load-balancing/docs/network/networklb-backend-service">cloud.google.com/load-balancing/docs/network/networklb-backend-service</a></p><p><a href="http://cloud.google.com/load-balancing/docs/choosing-load-balancer">cloud.google.com/load-balancing/docs/choosing-load-balancer</a></p><p><strong>Q13.</strong></p><p><strong>Heliovex Trading is developing a financial trading application that will be used globally. Data is stored and queried using a relational structure, and clients around the world must see the exact same state of the data. The application will be deployed in multiple regions to provide low latency to end users. You need to select a storage option for the application data while minimizing latency. What should you do?</strong></p><p><strong>A. Use Bigtable for data storage.</strong></p><p><strong>B. Use Cloud SQL for data storage.</strong></p><p><strong>C. Use Spanner for data storage.</strong></p><p><strong>D. Use Firestore for data storage.</strong></p><p>A is incorrect because Bigtable is a wide-column NoSQL database optimized for very high-throughput, low-latency key-value and time-series workloads. It is not a relational database and does not provide the globally distributed relational SQL and transaction model required by the financial trading application. The stated requirement for relational data and a single strongly consistent global state points to Spanner.</p><p>B is incorrect because Cloud SQL provides managed relational MySQL, PostgreSQL, and SQL Server databases, but an individual Cloud SQL primary is located in a region and is not a globally distributed, synchronously replicated relational database that lets active clients around the world operate against one externally consistent database state. Cross-region replicas can support disaster recovery and read workloads, but they do not provide the same global write and consistency architecture as Spanner.</p><blockquote><strong><em>C is correct because Spanner is a distributed relational database with SQL support, ACID transactions, synchronous replication, and global strong external consistency. A multi-region Spanner configuration can place replicas across geographic regions so data is closer to globally distributed applications while preserving one consistent database state. This combination of relational semantics, worldwide distribution, strong consistency, and high availability is specifically suited to globally deployed financial applications that cannot tolerate divergent views of transactional data.</em></strong></blockquote><p>D is incorrect because Firestore provides a document-oriented NoSQL data model rather than a relational schema. It supports strongly consistent reads and transactions, but it does not provide the relational SQL model described in the requirement. A globally distributed relational workload that requires one externally consistent state is a stronger fit for Spanner.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/spanner/docs/instance-configurations">cloud.google.com/spanner/docs/instance-configurations</a></p><p><a href="http://cloud.google.com/spanner/docs/replication">cloud.google.com/spanner/docs/replication</a></p><p><a href="http://cloud.google.com/spanner">cloud.google.com/spanner</a></p><p><strong>Q14.</strong></p><p><strong>Mavorex Intelligence manages a Google Cloud project for its Business Intelligence department. A data pipeline ingests data into BigQuery through streaming. You want users in the BI department to be able to run custom SQL queries against the latest data in BigQuery. What should you do?</strong></p><p><strong>A. Create a Looker Studio dashboard that uses the related BigQuery tables as a source and give the BI team view access to the Looker Studio dashboard.</strong></p><p><strong>B. Create a service account for the BI team and distribute a new private key to each member of the BI team.</strong></p><p><strong>C. Use Cloud Scheduler to schedule a batch Dataflow job that copies the data from BigQuery to the BI team’s internal data warehouse.</strong></p><p><strong>D. Assign the BigQuery User role (roles/bigquery.user) to a Google group that contains the members of the BI team.</strong></p><p>A is incorrect because view access to a Looker Studio dashboard lets users consume the predefined report but does not give them the ability to submit arbitrary custom SQL queries directly in BigQuery. Dashboards are suitable for curated visualization and self-service report interaction, but the requirement specifically asks for custom SQL access to the latest streamed BigQuery data.</p><p>B is incorrect because human users should authenticate with their own managed identities rather than share a service account. Distributing service-account private keys to multiple users creates long-lived credentials that are difficult to attribute, rotate, and revoke and violates identity and key-management best practices. IAM access should be granted to users through a group.</p><p>C is incorrect because copying the data to another warehouse introduces delay, duplicate storage, another processing pipeline, and additional cost. It also works against the requirement to query the latest data already being streamed into BigQuery. The BI users should query BigQuery directly.</p><blockquote><strong><em>D is correct because granting a BigQuery query-execution role to a Google group is the appropriate scalable IAM model for a BI team. The BigQuery User role, when granted at the project level, includes permission to create query jobs, and using a group makes membership easy to manage as employees join or leave the BI team. A current least-privilege implementation must also ensure that the group has permission to read the tables being queried, normally through BigQuery Data Viewer on the relevant datasets or tables. BigQuery User by itself does not grant arbitrary table-data access, but among the supplied choices D is the only option that correctly provides direct BigQuery query execution to the BI users; the required dataset read role should be granted in addition if it is not already present.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/bigquery/docs/running-queries">cloud.google.com/bigquery/docs/running-queries</a></p><p><a href="http://cloud.google.com/bigquery/docs/access-control">cloud.google.com/bigquery/docs/access-control</a></p><p><a href="http://cloud.google.com/iam/docs/groups-best-practices">cloud.google.com/iam/docs/groups-best-practices</a></p><p><strong>Q15.</strong></p><p><strong>Helvoryn Commerce has an internal application for managing transactional orders. The application is used exclusively by employees in a single physical location. It requires strong consistency, fast queries, and ACID guarantees for multi-table transactional updates. The first version is implemented in PostgreSQL, and you want to deploy it to Google Cloud with minimal code changes. Which database is most appropriate?</strong></p><p><strong>A. BigQuery.</strong></p><p><strong>B. Cloud SQL for PostgreSQL.</strong></p><p><strong>C. Spanner.</strong></p><p><strong>D. Firestore in Datastore mode.</strong></p><p>A is incorrect because BigQuery is a serverless analytical data warehouse optimized for large-scale analytical SQL queries rather than low-latency transactional application workloads. It is not intended to be the primary online transaction-processing database for frequent multi-row application updates and would require significant architectural changes from an existing PostgreSQL OLTP application.</p><blockquote><strong><em>B is correct because Cloud SQL provides a fully managed PostgreSQL database service. The application is already implemented against PostgreSQL, so moving it to Cloud SQL for PostgreSQL preserves the database engine and minimizes changes to SQL syntax, drivers, transaction behavior, and application code. PostgreSQL supports ACID transactions and strongly consistent transactional operations, making Cloud SQL an appropriate managed relational backend for an internal application in one location.</em></strong></blockquote><p>C is incorrect because Spanner provides relational SQL, ACID transactions, horizontal scalability, and strong external consistency, but it is a different database system with different schema, SQL, indexing, key-design, and application considerations. It is particularly valuable for very large, highly available, horizontally scalable, distributed transactional workloads. The scenario does not state that global scale is required and explicitly prioritizes minimal changes from an existing PostgreSQL application, making Cloud SQL the closer fit.</p><p>D is incorrect because Firestore in Datastore mode is a NoSQL document database with an entity-based data model rather than a PostgreSQL-compatible relational engine. Moving a multi-table PostgreSQL transactional application to Datastore mode would require substantial changes to the data model and application access patterns. It therefore conflicts with the requirement to minimize code changes.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/sql/docs/postgres/introduction">cloud.google.com/sql/docs/postgres/introduction</a></p><p><a href="http://cloud.google.com/sql/docs/postgres">cloud.google.com/sql/docs/postgres</a></p><p><strong>Q16.</strong></p><p><strong>Ilyvex Platforms is assigned to maintain a Google Kubernetes Engine cluster named dev that is deployed on Google Cloud. You want to manage the GKE configuration using the command-line interface. You have just downloaded and installed the Google Cloud CLI. You want future gcloud CLI commands to address this specific cluster by default. What should you do?</strong></p><p><strong>A. Use the command gcloud config set container/cluster dev.</strong></p><p><strong>B. Use the command gcloud container clusters update dev.</strong></p><p><strong>C. Create a file called gke.default in the ~/.gcloud folder that contains the cluster name.</strong></p><p><strong>D. Create a file called defaults.json in the ~/.gcloud folder that contains the cluster name.</strong></p><blockquote><strong><em>A is correct because the Google Cloud CLI exposes the container/cluster configuration property specifically for selecting a default GKE cluster. Running gcloud config set container/cluster dev stores dev as the cluster value in the active gcloud configuration. Commands that support the configured default can then use this cluster without requiring the cluster name to be repeatedly supplied through command-specific configuration.</em></strong></blockquote><p>B is incorrect because gcloud container clusters update changes configuration on an existing GKE cluster. It is used for operations such as modifying cluster features and settings; it does not set the local Google Cloud CLI’s default cluster property for future commands. The requirement concerns local CLI configuration rather than a server-side cluster modification.</p><p>C is incorrect because the Google Cloud CLI does not use a manually created gke.default file in the ~/.gcloud directory as the supported mechanism for defining the default GKE cluster. gcloud properties should be changed with the gcloud config commands so that the active named configuration is updated consistently.</p><p>D is incorrect because defaults.json is not the supported Google Cloud CLI configuration file for this purpose. Creating an arbitrary JSON file in the CLI configuration directory does not set the container/cluster property. The supported operation is gcloud config set container/cluster dev.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/how-to/managing-clusters">cloud.google.com/kubernetes-engine/docs/how-to/managing-clusters</a></p><p><a href="http://cloud.google.com/sdk/docs/configurations">cloud.google.com/sdk/docs/configurations</a></p><p><a href="http://cloud.google.com/sdk/gcloud/reference/config/set">cloud.google.com/sdk/gcloud/reference/config/set</a></p><p><strong>Q17.</strong></p><p><strong>Jorvanta Media has a sales team project named Sales Data Digest with the project ID acme-data-digest. You need to set up similar Google Cloud resources for the marketing team, but the marketing resources must be organized independently of the sales resources. What should you do?</strong></p><p><strong>A. Grant the Project Editor basic role to the marketing team for acme-data-digest.</strong></p><p><strong>B. Create a project lien on acme-data-digest and then grant the Project Editor basic role to the marketing team.</strong></p><p><strong>C. Create another project with the ID acme-marketing-data-digest for the marketing team and deploy the resources there.</strong></p><p><strong>D. Create a new project named Marketing Data Digest and use the ID acme-data-digest. Grant the Project Editor basic role to the marketing team.</strong></p><p>A is incorrect because granting the marketing team Editor access to the existing sales project places both teams’ resources in the same project and administrative boundary. The resources would share project-level quotas, billing attribution, IAM inheritance, enabled services, and many policy controls. This does not meet the requirement to organize marketing independently from sales.</p><p>B is incorrect because a project lien protects a project from accidental deletion by preventing project deletion while the lien exists. It does not create a separate organizational boundary for another team’s resources. Marketing would still share the same project with sales, and granting the broad Editor basic role would provide access across that shared project.</p><blockquote><strong><em>C is correct because projects are fundamental Google Cloud organizing and isolation boundaries. Creating a separate project gives the marketing team its own resource namespace, IAM policy scope, quotas, service configuration, billing attribution, and lifecycle while keeping the existing sales project independent. The project ID must be globally unique, so using a distinct ID such as acme-marketing-data-digest is also required.</em></strong></blockquote><p>D is incorrect because a Google Cloud project ID is globally unique and cannot be reused by a second project. The existing sales project already uses acme-data-digest, so a new marketing project cannot use that same ID. Even though a display name does not have to be globally unique, the project ID does.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/docs/get-started/organize-resources">cloud.google.com/docs/get-started/organize-resources</a></p><p><a href="http://cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy">cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy</a></p><p><strong>Q18.</strong></p><p><strong>Faronvex Holdings has multiple projects linked to a single Cloud Billing account in Google Cloud. You need to visualize the costs with specific metrics that should be dynamically calculated based on company-specific criteria. You want to automate the process. What should you do?</strong></p><p><strong>A. In the Google Cloud console, visualize the costs related to the projects in the Reports section.</strong></p><p><strong>B. In the Google Cloud console, visualize the costs related to the projects in the Cost breakdown section.</strong></p><p><strong>C. In the Google Cloud console, use the export functionality of the Cost table. Create a Looker Studio dashboard on top of the CSV export.</strong></p><p><strong>D. Configure Cloud Billing data export to BigQuery for the billing account. Create a Looker Studio dashboard on top of the BigQuery export.</strong></p><p>A is incorrect because Cloud Billing Reports provides interactive built-in cost visualization and filtering, but it is designed around the billing dimensions and calculations provided by Google Cloud. It does not provide the same flexibility as SQL for continuously calculating arbitrary company-specific metrics and transformations. The requirement asks for dynamically calculated custom metrics and an automated process.</p><p>B is incorrect because the Cost breakdown view is intended to help explain how list cost, negotiated savings, committed-use discounts, credits, and other cost components contribute to the bill. It is a predefined billing analysis view rather than a general custom-metric computation platform. It therefore does not provide the required automated company-specific calculation layer.</p><p>C is incorrect because a CSV export from the Cost table is a point-in-time manual export. A Looker Studio report built on that file would not automatically receive new billing records throughout the day unless another process repeatedly recreated and replaced the data source. This fails the automation requirement even though Looker Studio can visualize the exported information.</p><blockquote><strong><em>D is correct because Cloud Billing can automatically export standard or detailed usage cost data to BigQuery throughout the day. BigQuery gives the organization a programmable SQL layer where company-specific metrics, allocations, ratios, labels, and business logic can be calculated dynamically. Looker Studio can then use BigQuery queries, views, or tables as its data source and refresh visualizations as new exported billing records arrive. This architecture provides both automation and the flexibility needed for custom cost metrics.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/billing/docs/how-to/export-data-bigquery">cloud.google.com/billing/docs/how-to/export-data-bigquery</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/export-data-bigquery-setup">cloud.google.com/billing/docs/how-to/export-data-bigquery-setup</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/bq-examples">cloud.google.com/billing/docs/how-to/bq-examples</a></p><p><strong>Q19.</strong></p><p><strong>Helyvora Group completed the acquisition of a startup and is now merging the IT systems of both companies. The startup had a production Google Cloud project in its organization. You need to move this project into your organization and ensure that the project is billed to your organization. You want to accomplish this task with minimal effort. What should you do?</strong></p><p><strong>A. Use the projects.move method to move the project to your organization. Update the Cloud Billing account of the project to your organization’s Cloud Billing account.</strong></p><p><strong>B. Ensure that you have the Organization Administrator IAM role assigned to you in both organizations. Navigate to Resource Manager in the startup’s Google Cloud organization, and drag the project to your company’s organization.</strong></p><p><strong>C. Create a Google Cloud Private Marketplace collection, and upload the resources of the startup’s production project to the collection. Share the collection with your organization, and deploy the resources in your company’s project.</strong></p><p><strong>D. Create an infrastructure-as-code configuration for all resources in the project by using Terraform, and deploy that configuration to a new project in your organization. Delete the project from the startup’s Google Cloud organization.</strong></p><blockquote><strong><em>A is correct because Google Cloud supports migrating an existing project between organization resources by changing the project’s parent through Resource Manager. The Resource Manager API supports the projects.move operation for this purpose, subject to the required migration IAM roles, organization policies, and migration preparation. The migration is a metadata operation: the project ID, project number, most directly attached IAM policies, and deployed resources remain with the project instead of being rebuilt. Moving the project does not automatically change its billing relationship, so you must separately link it to the destination organization’s intended Cloud Billing account. This satisfies both requirements with much less effort than rebuilding the production environment.</em></strong></blockquote><p>B is incorrect because cross-organization project migration has explicit IAM, organization-policy, and migration requirements; it is not simply a drag-and-drop operation that becomes valid merely because a person has Organization Administrator in both organizations. The current Google Cloud console can support selecting a new parent as part of an eligible migration workflow, but the option’s description omits the required project-migration preparation and billing change and incorrectly reduces the operation to a simple UI drag action.</p><p>C is incorrect because Google Cloud Private Marketplace is designed to curate and govern approved Google Cloud Marketplace products for an organization. It is not a mechanism for packaging all of the resources in an arbitrary production project and transferring that project’s live infrastructure, identities, databases, and configuration to another organization. The existing project can be migrated directly without recreating its resources.</p><p>D is incorrect because Terraform is an excellent infrastructure-as-code tool, but recreating an entire production project from Terraform can be a substantial migration project and might not capture all existing data, runtime state, manually configured resources, or service-specific settings. It also introduces a cutover and deletion process that is unnecessary when Resource Manager supports moving the existing project while keeping its resources in place.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/project-migration">cloud.google.com/resource-manager/docs/project-migration</a></p><p><a href="http://cloud.google.com/resource-manager/docs/perform-migration">cloud.google.com/resource-manager/docs/perform-migration</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/modify-project">cloud.google.com/billing/docs/how-to/modify-project</a></p><p><strong>Q20.</strong></p><p><strong>Jorvixa Networks is configuring Cloud DNS. You want to create DNS records to point home.mydomain.com, mydomain.com, and </strong><a href="http://www.mydomain.com/"><strong>www.mydomain.com</strong></a><strong> to the IP address of your Google Cloud load balancer. What should you do?</strong></p><p><strong>A. Create one CNAME record to point mydomain.com to the load balancer, and create two A records to point WWW and HOME to mydomain.com respectively.</strong></p><p><strong>B. Create one CNAME record to point mydomain.com to the load balancer, and create two AAAA records to point WWW and HOME to mydomain.com respectively.</strong></p><p><strong>C. Create one A record to point mydomain.com to the load balancer, and create two CNAME records to point WWW and HOME to mydomain.com respectively.</strong></p><p><strong>D. Create one A record to point mydomain.com to the load balancer, and create two NS records to point WWW and HOME to mydomain.com respectively.</strong></p><p>A is incorrect because mydomain.com is the zone apex, and a standard CNAME record cannot normally coexist at the zone apex with required DNS records such as SOA and NS. In addition, an A record maps a hostname to an IPv4 address rather than mapping a hostname alias to another hostname, so using A records for WWW and HOME that point to the text name mydomain.com is conceptually wrong. The apex should use an address record for the load balancer’s IPv4 address.</p><p>B is incorrect because it still attempts to use a CNAME at the zone apex, and AAAA records are specifically for IPv6 addresses. An AAAA record cannot be used to point WWW or HOME to the hostname mydomain.com unless an IPv6 address is supplied as its value. This does not match the requested aliasing pattern.</p><blockquote><strong><em>C is correct because an A record can map the zone apex mydomain.com directly to the load balancer’s IPv4 address. The hostnames </em></strong><a href="http://www.mydomain.com/"><strong><em>www.mydomain.com</em></strong></a><strong><em> and home.mydomain.com can then be CNAME aliases whose canonical name is mydomain.com, causing clients to resolve through the apex A record. Cloud DNS also supports a custom ALIAS record type for certain apex aliasing scenarios today, but that record type is not one of the supplied choices and is unnecessary when the load balancer IP can be entered directly in an A record.</em></strong></blockquote><p>D is incorrect because NS records identify the authoritative name servers for a DNS zone or delegated child zone. They are not used to make ordinary hostnames such as WWW and HOME aliases of another hostname. Using NS records would change DNS delegation behavior rather than point those names at the load balancer.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/dns/docs/records-overview">cloud.google.com/dns/docs/records-overview</a></p><p><a href="http://cloud.google.com/dns/docs/records">cloud.google.com/dns/docs/records</a></p><p><strong>Q21.</strong></p><p><strong>Morynex Technologies is a startup that was officially registered as a business six months ago. As the customer base grows, the company’s use of Google Cloud increases. You want to allow all engineers to create new projects without asking them for their credit card information. What should you do?</strong></p><p><strong>A. Create a Cloud Billing account, associate a payment method with it, and provide all project creators with permission to associate that Cloud Billing account with their projects.</strong></p><p><strong>B. Grant all engineers permission to create their own Cloud Billing accounts for each new project.</strong></p><p><strong>C. Apply for monthly invoiced billing, and have a single invoice for the project paid by the finance team.</strong></p><p><strong>D. Create a Cloud Billing account, associate it with a monthly purchase order (PO), and send the PO to Google Cloud.</strong></p><blockquote><strong><em>A is correct because a single active Cloud Billing account can pay for the usage of many Google Cloud projects. The organization can centrally configure the payment method once and then grant engineers the permissions necessary to create projects and link those projects to the existing billing account. Current IAM guidance commonly uses Project Creator together with Billing Account User for this delegated project-creation model. Engineers therefore do not need to supply personal credit-card details each time they create a project, while financial control remains centralized.</em></strong></blockquote><p>B is incorrect because allowing every engineer to create a new billing account would cause billing-account sprawl, make financial governance and reporting harder, and potentially require repeated payments-profile or payment-method configuration. Google Cloud documentation specifically recommends controlling Billing Account Creator access because uncontrolled creation can lead to untracked spend. One shared organizational billing account can pay for many projects.</p><p>C is incorrect because monthly invoiced billing is not simply a setting that any recently formed startup can enable on demand. Invoiced billing is subject to Google eligibility, credit, contractual, and spending requirements. Even if the company later becomes eligible, it is unnecessary for the immediate requirement because a self-serve Cloud Billing account with a centrally managed payment method can already be shared across projects.</p><p>D is incorrect because ordinary self-serve Cloud Billing accounts are charged through the payment method associated with the Google payments profile; a purchase order is not a substitute for setting up an eligible invoiced billing relationship. Purchase-order handling applies to particular invoiced-account arrangements and is not the general mechanism for enabling engineers to create projects without personal payment details.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/billing/docs/how-to/create-billing-account">cloud.google.com/billing/docs/how-to/create-billing-account</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/billing-access">cloud.google.com/billing/docs/how-to/billing-access</a></p><p><a href="http://cloud.google.com/resource-manager/docs/creating-managing-projects">cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><strong>Q22.</strong></p><p><strong>Elarvanta Software requires all developers to have the same permissions regardless of the Google Cloud project they are working on. The company’s security policy also restricts developer permissions to Compute Engine, Cloud Run functions, and Cloud SQL. You want to implement the security policy with minimal effort. What should you do?</strong></p><p><strong>A. • Create a custom role with Compute Engine, Cloud Run functions, and Cloud SQL permissions in one project within the Google Cloud organization.</strong></p><p><strong>• Copy the role across all projects created within the organization with the gcloud iam roles copy command.</strong></p><p><strong>• Assign the role to developers in those projects.</strong></p><p><strong>B. • Add all developers to a Google group.</strong></p><p><strong>• Assign the predefined Compute Admin role to the Google group at the Google Cloud organization level.</strong></p><p><strong>C. • Add all developers to a Google group in Cloud Identity.</strong></p><p><strong>• Assign predefined roles for Compute Engine, Cloud Run functions, and Cloud SQL permissions to the Google group for each project in the Google Cloud organization.</strong></p><p><strong>D. • Add all developers to a Google group in Cloud Identity.</strong></p><p><strong>• Create a custom role with Compute Engine, Cloud Run functions, and Cloud SQL permissions at the Google Cloud organization level.</strong></p><p><strong>• Assign the custom role to the Google group.</strong></p><p>A is incorrect because project-level custom roles belong to individual projects. Copying the role into every existing and future project and then assigning it repeatedly creates ongoing administrative work and makes it easier for role definitions to drift between projects. The requirement says developers should have the same permissions throughout the organization with minimal effort, so an organization-level role and inherited role binding are better suited to the task.</p><p>B is incorrect because the Compute Admin role grants broad Compute Engine permissions but does not represent the required permission set for Cloud Run functions and Cloud SQL. It also grants substantially broader Compute Engine authority than might be required by the organization’s security policy. A custom role can express the approved cross-service permission set more precisely.</p><p>C is incorrect because using a group is good practice, and predefined roles are generally preferred when they exactly fit a job function. However, assigning several roles independently in every project creates repeated administration across the organization and future projects. The question states that the same organization-wide security policy should apply to all developers regardless of project, so a centrally defined organization-level custom role and inherited grant requires less ongoing effort.</p><blockquote><strong><em>D is correct because an organization-level custom role can combine the supported permissions that the security policy approves across Compute Engine, Cloud Run functions, and Cloud SQL. Adding developers to a Cloud Identity group provides centralized membership management, and granting the custom role to that group at the organization level allows the binding to be inherited by descendant projects. This avoids copying custom roles or maintaining repeated per-project bindings while keeping the permission set aligned with the security policy. Organization-level grants should be used carefully because of their broad scope, but that scope directly matches this question’s explicit requirement that all developers have the same access across projects.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/iam/docs/creating-custom-roles">cloud.google.com/iam/docs/creating-custom-roles</a></p><p><a href="http://cloud.google.com/iam/docs/groups-best-practices">cloud.google.com/iam/docs/groups-best-practices</a></p><p><a href="http://cloud.google.com/iam/docs/allow-policies">cloud.google.com/iam/docs/allow-policies</a></p><p><strong>Q23.</strong></p><p><strong>Glyndora Systems recently received a new Google Cloud project with an attached Cloud Billing account. You need to create Compute Engine instances, configure VPC firewall rules, and store data in Cloud Storage. You want to follow Google-recommended practices. What should you do?</strong></p><p><strong>A. Use the gcloud services enable cloudresourcemanager.googleapis.com command to enable all required resources.</strong></p><p><strong>B. Use the gcloud services enable compute.googleapis.com command to enable Compute Engine and the gcloud services enable storage-api.googleapis.com command to enable the Cloud Storage JSON API.</strong></p><p><strong>C. Open the Google Cloud console and enable every Google Cloud API from the API dashboard.</strong></p><p><strong>D. Open the Google Cloud console and run gcloud init — project in Cloud Shell.</strong></p><p>A is incorrect because enabling the Cloud Resource Manager API does not automatically enable Compute Engine, Cloud Storage, or every service that a project might use. Google Cloud APIs are enabled independently through Service Usage. Cloud Resource Manager handles resource-hierarchy operations and is not a master switch for all Google Cloud services.</p><blockquote><strong><em>B is correct because Compute Engine operations such as creating VM instances and VPC firewall rules require the Compute Engine API, compute.googleapis.com, to be enabled. Cloud Storage API services can likewise be enabled through Service Usage. In current newly created projects, Cloud Storage-related APIs, including storage.googleapis.com and the Google Cloud Storage JSON API service, are generally enabled by default, so explicitly enabling the JSON API can be redundant; however, it is harmless, and this is the only option that enables the specific service APIs rather than indiscriminately enabling all services. The key required preparation in this scenario is enabling Compute Engine.</em></strong></blockquote><p>C is incorrect because enabling every API in a project is unnecessary and contrary to good service-governance practice. Each enabled API expands the set of services that can be consumed and can increase governance complexity. You should enable only the APIs required by the workloads and administrative tasks.</p><p>D is incorrect because gcloud init configures the local Google Cloud CLI account, project, and related properties. It does not enable the Compute Engine or Cloud Storage services in the project. CLI initialization and API enablement are separate operations.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/service-usage/docs/enable-disable">cloud.google.com/service-usage/docs/enable-disable</a></p><p><a href="http://cloud.google.com/service-usage/docs/enabled-service">cloud.google.com/service-usage/docs/enabled-service</a></p><p><a href="http://cloud.google.com/compute/docs/instances/create-start-instance">cloud.google.com/compute/docs/instances/create-start-instance</a></p><p><strong>Q24.</strong></p><p><strong>Dovarex Analytics runs a Dataproc cluster in a single VPC network and a single subnet with the range 172.16.20.128/25. There are no private IP addresses available in the subnet. You want to add new VMs that can communicate with the cluster using the minimum number of steps. What should you do?</strong></p><p><strong>A. Expand the existing subnet range to 172.16.20.0/24.</strong></p><p><strong>B. Create a new secondary IPv4 range in the VPC network and configure the VMs to use that range.</strong></p><p><strong>C. Create a new VPC network for the VMs and configure VPC Network Peering between the VM VPC network and the Dataproc cluster VPC network.</strong></p><p><strong>D. Create a new VPC network for the VMs with a subnet of 172.32.0.0/16. Configure VPC Network Peering between the Dataproc VPC network and the VM VPC network, and configure custom route exchange.</strong></p><blockquote><strong><em>A is correct because Google Cloud supports in-place expansion of a subnet’s primary IPv4 range to a larger supernet as long as the expanded range does not conflict with other network ranges. The current 172.16.20.128/25 range is contained in 172.16.20.0/24, so expanding it to /24 makes more primary internal IP addresses available while preserving the existing subnet and its connected workloads. Traffic is not interrupted by the subnet expansion, and the new VMs remain in the same VPC routing domain as the Dataproc cluster, which minimizes configuration steps.</em></strong></blockquote><p>B is incorrect because ordinary VM primary network-interface addresses are allocated from the subnet’s primary IPv4 range. Secondary ranges are used for alias IP ranges and services such as VPC-native GKE Pod and Service address allocation. Adding a secondary range does not replenish the exhausted pool used for the VMs’ primary internal IP addresses.</p><p>C is incorrect because a second VPC network and VPC Network Peering could provide private communication, but this adds another subnet, a second routing domain, two-sided peering configuration, firewall considerations, and additional network administration. The existing subnet can simply be expanded, so a new network is unnecessary.</p><p>D is incorrect because this option introduces even more components than C by requiring another VPC, another subnet, peering, and custom route exchange. Custom route exchange is useful for specific peering scenarios involving custom routes, but it is unnecessary when the existing subnet’s primary range can be expanded directly.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/vpc/docs/create-modify-vpc-networks">cloud.google.com/vpc/docs/create-modify-vpc-networks</a></p><p><a href="http://cloud.google.com/vpc/docs/subnets">cloud.google.com/vpc/docs/subnets</a></p><p><strong>Q25.</strong></p><p><strong>Havoriq Web Services wants to create a secure website with autoscaling based on Compute Engine CPU load. You also want to improve performance by storing static content in Cloud Storage. Which resources are needed to distribute user traffic?</strong></p><p><strong>A. An external Application Load Balancer with a Google-managed SSL certificate and a URL map that routes requests for static content to a Cloud Storage backend bucket.</strong></p><p><strong>B. A regional external passthrough Network Load Balancer pointing to the backend instances. The web servers forward requests to Cloud Storage when needed.</strong></p><p><strong>C. An internal Application Load Balancer together with Identity-Aware Proxy to allow only HTTPS traffic.</strong></p><p><strong>D. An external Application Load Balancer with a URL map that routes requests for static content to a Cloud Storage backend bucket. Install the HTTPS certificates on the backend instances.</strong></p><blockquote><strong><em>A is correct because an external Application Load Balancer can terminate HTTPS at Google’s load-balancing layer, use a Google-managed certificate, and route requests through a URL map. Dynamic paths can be sent to a backend service containing the Compute Engine managed instance group, while paths for static assets can be routed directly to a Cloud Storage backend bucket. This avoids forcing dynamic web servers to proxy static content and lets the application tier scale independently. The managed SSL certificate also reduces certificate-management work because Google provisions and renews it.</em></strong></blockquote><p>B is incorrect because a passthrough Network Load Balancer operates at Layer 4 and does not provide HTTP path-based routing through a URL map. The web servers would need to receive requests for static assets and redirect or proxy those requests to Cloud Storage themselves, adding application work and failing to use the load-balancer feature designed for separate dynamic and static backends.</p><p>C is incorrect because an internal Application Load Balancer is intended for internal clients in a VPC or connected private networks. It does not provide the public internet-facing HTTPS frontend required by a public website. Identity-Aware Proxy can protect suitable applications but does not replace the public external load balancer needed in this scenario.</p><p>D is incorrect because an external Application Load Balancer terminates HTTPS at the frontend proxy, so the client-facing certificate should be attached to the load balancer rather than installed on the backend VMs for normal HTTPS termination. Backend encryption can be configured separately when required, but putting the customer-facing certificate only on the instances does not provide the managed frontend certificate architecture described by the requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/load-balancing/docs/https/setup-global-ext-https-buckets">cloud.google.com/load-balancing/docs/https/setup-global-ext-https-buckets</a></p><p><a href="http://cloud.google.com/load-balancing/docs/url-map-concepts">cloud.google.com/load-balancing/docs/url-map-concepts</a></p><p><a href="http://cloud.google.com/load-balancing/docs/ssl-certificates/google-managed-certs">cloud.google.com/load-balancing/docs/ssl-certificates/google-managed-certs</a></p><p><strong>Q26.</strong></p><p><strong>Jorvexa Technologies has just installed the Google Cloud CLI on a new corporate laptop. You need to list the company’s existing Compute Engine instances. What must you do before you run gcloud compute instances list? (Choose two.)</strong></p><p><strong>A. Run gcloud auth login and complete the supported sign-in flow to authenticate the Google Cloud CLI.</strong></p><p><strong>B. Create a Google Cloud service account, download a service account key, and place the key file in a folder where the Google Cloud CLI can find it.</strong></p><p><strong>C. Download a Cloud Identity user account key and place the key file in a folder where the Google Cloud CLI can find it.</strong></p><p><strong>D. Run gcloud config set compute/zone $my_zone to set the default Compute Engine zone.</strong></p><p><strong>E. Run gcloud config set project $my_project to set the default project for the Google Cloud CLI.</strong></p><blockquote><strong><em>A is correct because the Google Cloud CLI must have an authenticated identity before it can call the Compute Engine API on the user’s behalf. For an employee using the CLI interactively on a corporate laptop, gcloud auth login starts the supported authentication flow and stores the resulting CLI credentials. The user must also have sufficient IAM permission to list Compute Engine instances. Modern gcloud authentication normally uses a browser-based sign-in flow rather than requiring the user to manually copy an access token into the CLI.</em></strong></blockquote><p>B is incorrect because an interactive human user does not need to create a service account or download a service account key merely to administer Google Cloud from the gcloud CLI. Service account keys are long-lived credentials with significant security risk, and Google recommends avoiding them when normal user authentication or another keyless authentication method is available.</p><p>C is incorrect because Cloud Identity user accounts do not provide downloadable private key files analogous to service account keys. Human users authenticate through supported Google or federated sign-in flows. Therefore, this is not a valid preparation step for the gcloud CLI.</p><p>D is incorrect because gcloud compute instances list is not a zonal-only listing operation that requires a default zone. The command can list VM instances across zones in the selected project. Setting compute/zone can make commands that operate on a particular zonal resource more convenient, but it is not required before listing instances.</p><blockquote><strong><em>E is correct because the gcloud CLI needs to know which Google Cloud project should be used when a command is issued without an explicit — project flag. Setting the project property with gcloud config set project $my_project selects the target project in the active configuration. The same project could alternatively be specified directly on the command line, but given the choices, configuring the default project is the expected setup step.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/sdk/docs/authenticate">cloud.google.com/sdk/docs/authenticate</a></p><p><a href="http://cloud.google.com/sdk/docs/configurations">cloud.google.com/sdk/docs/configurations</a></p><p><a href="http://cloud.google.com/sdk/gcloud/reference/config/set">cloud.google.com/sdk/gcloud/reference/config/set</a></p><p><strong>Q27.</strong></p><p><strong>Lumivora Apps wants to deploy a new containerized application to Google Cloud by using a Kubernetes manifest. You want full control over the Kubernetes workload deployment while minimizing infrastructure configuration. What should you do?</strong></p><p><strong>A. Deploy the application on GKE Autopilot.</strong></p><p><strong>B. Deploy the application on Cloud Run.</strong></p><p><strong>C. Deploy the application on GKE Standard.</strong></p><p><strong>D. Deploy the application as a Cloud Run function.</strong></p><blockquote><strong><em>A is correct because GKE Autopilot uses the Kubernetes API and standard compatible Kubernetes workload manifests while Google manages most node infrastructure, scaling, security defaults, and underlying compute configuration. The application team can continue controlling Kubernetes objects such as Deployments, Services, ConfigMaps, resource requests, scheduling requirements, and other supported workload settings. Autopilot then provisions and manages the infrastructure required by those manifests. This gives the Kubernetes deployment model while minimizing infrastructure administration.</em></strong></blockquote><p>B is incorrect because Cloud Run minimizes infrastructure management even further, but it does not deploy an arbitrary Kubernetes manifest as the primary workload configuration model. Cloud Run uses its own service and job resource model. If the team specifically requires control through Kubernetes manifests, GKE is the appropriate platform.</p><p>C is incorrect because GKE Standard provides the greatest control over cluster and node infrastructure, but that control also requires more configuration and ongoing management of node pools, machine types, scaling settings, and other infrastructure details. The requirement is to retain Kubernetes workload control while minimizing infrastructure configuration, which is the purpose of Autopilot.</p><p>D is incorrect because Cloud Run functions use a functions-oriented source and event-trigger model rather than general Kubernetes manifests. A containerized application that needs Kubernetes deployment semantics is better suited to GKE Autopilot.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/concepts/autopilot-overview">cloud.google.com/kubernetes-engine/docs/concepts/autopilot-overview</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/get-started/deploy-workloads">cloud.google.com/kubernetes-engine/docs/get-started/deploy-workloads</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/how-to/prepare-migrate-cluster-mode">cloud.google.com/kubernetes-engine/docs/how-to/prepare-migrate-cluster-mode</a></p><p><strong>Q28.</strong></p><p><strong>Noryvex Finance is deploying an application on Google Cloud that requires a relational database. Company security policy requires an encrypted and authenticated database connection with minimal certificate management and integration with Identity and Access Management (IAM). What should you do?</strong></p><p><strong>A. Deploy a Cloud SQL database with SSL/TLS configured for encrypted connections, configure client certificates, and use a database username and password.</strong></p><p><strong>B. Deploy a Cloud SQL database with SSL/TLS configured for encrypted connections, configure client certificates, and use IAM database authentication.</strong></p><p><strong>C. Deploy a Cloud SQL database, configure IAM database authentication, and connect through the Cloud SQL Auth Proxy.</strong></p><p><strong>D. Deploy a Cloud SQL database, configure a database username and password, and connect through the Cloud SQL Auth Proxy.</strong></p><p>A is incorrect because TLS client certificates can provide encrypted and authenticated connections, but they require certificate provisioning, distribution, renewal, and management. Static database usernames and passwords also do not integrate database login with Google Cloud IAM. This therefore fails the requirement to minimize management and use IAM.</p><p>B is incorrect because IAM database authentication integrates database login with IAM, but explicitly managing client SSL/TLS certificates still adds certificate lifecycle work that can be avoided. Cloud SQL Connectors such as the Cloud SQL Auth Proxy automatically create secure, encrypted connections and manage short-lived connection certificates on behalf of the application.</p><blockquote><strong><em>C is correct because the Cloud SQL Auth Proxy uses IAM to authorize the connection to the Cloud SQL instance and automatically establishes an encrypted TLS connection using short-lived certificates, eliminating the need to manage client SSL certificates manually. Cloud SQL IAM database authentication can also map the application’s Google Cloud identity to a database login and, with automatic IAM authentication enabled, the proxy can refresh the required access tokens. This provides both encrypted transport and IAM-integrated authentication with minimal credential-management overhead.</em></strong></blockquote><p>D is incorrect because the Cloud SQL Auth Proxy provides a secure encrypted connection and IAM-based connection authorization, but using a static database username and password does not satisfy the requirement to integrate database authentication itself with IAM. Where the selected Cloud SQL database engine supports IAM database authentication, using IAM for the database login reduces reliance on long-lived passwords.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/sql/docs/postgres/sql-proxy">cloud.google.com/sql/docs/postgres/sql-proxy</a></p><p><a href="http://cloud.google.com/sql/docs/postgres/iam-logins">cloud.google.com/sql/docs/postgres/iam-logins</a></p><p><a href="http://cloud.google.com/sql/docs/postgres/connect-auth-proxy">cloud.google.com/sql/docs/postgres/connect-auth-proxy</a></p><p><strong>Q29.</strong></p><p><strong>Dervonix Governance recently implemented an automated process for provisioning Google Cloud projects. You have discovered several projects that were manually created before the new process was introduced. You need to identify the creators of these projects so that you can inform them of the new process. What should you do?</strong></p><p><strong>A. Review the value of a creator label attached to each project.</strong></p><p><strong>B. Use Resource Manager to identify the responsible department from the folder and project hierarchy.</strong></p><p><strong>C. Review the Cloud Billing account linked to each project, and contact the owner of the billing account.</strong></p><p><strong>D. Use the IAM page in the Google Cloud console to identify the user with the Owner role.</strong></p><p>A is incorrect because Google Cloud does not automatically add a standard creator label containing the identity of the person who created a project. Labels are user-defined metadata and are useful only if the organization had already implemented a process that populated such a label. The question describes projects that predate the automated process, so this metadata cannot be assumed to exist.</p><p>B is incorrect because the resource hierarchy can identify the organization or folder that contains a project, but a project’s parent does not identify the individual who originally created it. A folder can contain projects created by many different users.</p><p>C is incorrect because a Cloud Billing account identifies the billing relationship for the project, not the person who created the project. A billing account owner can link, administer, or pay for many projects created by other users.</p><blockquote><strong><em>D is correct because Google Cloud initially grants the Owner basic role on a newly created project to the user who creates the project. Reviewing the project IAM policy can therefore identify the original creator when that initial ownership binding has not subsequently been changed. This is not a guaranteed historical audit method: projects can acquire additional Owners, the original creator can later lose the Owner role, and current IAM shows present authorization rather than immutable project history. For an exact historical answer when retained logs are available, Resource Manager Admin Activity audit logs for the CreateProject operation and its principal identity are more authoritative, but that option is not provided in this question.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/access-control-proj">cloud.google.com/resource-manager/docs/access-control-proj</a></p><p><a href="http://cloud.google.com/resource-manager/docs/audit-logging">cloud.google.com/resource-manager/docs/audit-logging</a></p><p><a href="http://cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy">cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy</a></p><p><strong>Q30.</strong></p><p><strong>Orvexia Monitoring is developing an application that runs on a Google Kubernetes Engine (GKE) cluster and requires service account permissions to collect performance data from Google Cloud services. The organization prohibits creating new service account keys. You need to grant the application access while following Google-recommended security practices. What should you do?</strong></p><p><strong>A. Use your personal user credentials to authenticate the application and ensure that your account has the required permissions.</strong></p><p><strong>B. Create a new project, generate a service account key in that project, and grant the service account the required IAM roles in the application project.</strong></p><p><strong>C. Override the organization policy in the project and create a long-lived service account key for the GKE workload.</strong></p><p><strong>D. Configure Workload Identity Federation for GKE so that the workload receives the required Google Cloud identity and, when necessary for the target API, impersonates an IAM service account.</strong></p><p>A is incorrect because a production workload should not run with a developer’s personal user credentials. User credentials are tied to the employee’s lifecycle, can be revoked when the person changes roles or leaves, and make workload audit trails less clear. Applications should use workload identities that are independent of individual users.</p><p>B is incorrect because placing a service account in another project does not avoid the security risk of a user-managed service account key. The private key remains a long-lived credential that must be distributed to the cluster, stored securely, rotated, and revoked if compromised. The organization policy intentionally prohibits that authentication pattern.</p><p>C is incorrect because overriding a security organization policy solely to create a long-lived key weakens the organization’s control and is unnecessary. GKE provides a native workload identity mechanism that supplies short-lived credentials without private key files.</p><blockquote><strong><em>D is correct because Workload Identity Federation for GKE lets Kubernetes workloads authenticate to Google Cloud APIs without service account keys. For most supported Google Cloud APIs, you can grant IAM roles directly to the Kubernetes ServiceAccount principal, which is the simpler current approach. When an API requires or benefits from an IAM service account identity, Workload Identity Federation for GKE also supports linking the Kubernetes ServiceAccount to an IAM service account and using service account impersonation. Both models provide short-lived credentials and maintain workload-specific least-privilege access without violating the organization’s prohibition on keys.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/concepts/workload-identity">cloud.google.com/kubernetes-engine/docs/concepts/workload-identity</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/how-to/workload-identity">cloud.google.com/kubernetes-engine/docs/how-to/workload-identity</a></p><p><a href="http://cloud.google.com/iam/docs/workload-identities">cloud.google.com/iam/docs/workload-identities</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3cd266f7116" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[HashiCorp Certified Terraform Associate (004) Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/hashicorp-certified-terraform-associate-004-practice-questions-97f19d6f6d1b?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/97f19d6f6d1b</guid>
            <category><![CDATA[terraform-certification]]></category>
            <category><![CDATA[hashicorp-terraform]]></category>
            <category><![CDATA[terraform-associate-exam]]></category>
            <category><![CDATA[terraform]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 02:40:38 GMT</pubDate>
            <atom:updated>2026-10-07T02:40:38.199Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n1Baoj53cW81-PAM-RxtKQ.jpeg" /><figcaption>KloudExams HashiCorp Certified Terraform Associate (004) Practice Questions</figcaption></figure><p>Preparing for the <strong>HashiCorp Certified: Terraform Associate (004) certification</strong>? Knowing what Terraform commands, providers, and state do is a starting point — understanding how they work together to provision and manage infrastructure is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing configurations and workflows that fit specific infrastructure requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through official HashiCorp documentation and hands-on Terraform practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>The fictional company Aetherquill Systems is adopting Infrastructure as Code (IaC). IaC configuration can be stored in a version control system along with application code.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><blockquote><strong><em>A is correct because Infrastructure as Code treats infrastructure definitions as code files, which means they can be stored in a version control system such as Git just like application source code. Version control provides history, change review, branching, collaboration, rollback of configuration changes, and an auditable record of how the infrastructure definition evolved. Terraform’s core workflow explicitly assumes that configuration is authored as code and commonly stored in version control. The word “can” is important: infrastructure configuration may be kept in the same repository as application code or in a separate infrastructure repository, depending on the organization’s repository strategy. Either model is compatible with IaC.</em></strong></blockquote><p>B is incorrect because there is no Terraform restriction that prevents infrastructure configuration from being version controlled alongside application code. In fact, keeping Terraform configuration in version control is a standard practice because it enables repeatable changes and peer review. A team may intentionally separate application and infrastructure repositories for ownership or lifecycle reasons, but that is an architectural choice rather than a limitation of IaC. Therefore, the statement is not false merely because separate repositories are also valid.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/intro/core-workflow">developer.hashicorp.com/terraform/intro/core-workflow</a></p><p><a href="http://developer.hashicorp.com/terraform/language/style">developer.hashicorp.com/terraform/language/style</a></p><p><strong>Q2.</strong></p><p><strong>The fictional company Brindleforge Cloudworks is comparing Infrastructure as Code (IaC) with provisioning performed only through a graphical user interface (GUI). Which is an advantage of using IaC that is not available when provisioning only through a GUI?</strong></p><p><strong>A. Lets you version, reuse, and share infrastructure configuration.</strong></p><p><strong>B. Secures your credentials.</strong></p><p><strong>C. Provisions the same resources at a lower cost.</strong></p><p><strong>D. Prevents manual modifications to your resources.</strong></p><blockquote><strong><em>A is correct because IaC expresses infrastructure in files that can be versioned, copied, reviewed, packaged into reusable modules, and shared with other teams. Terraform configuration therefore becomes a durable artifact rather than a sequence of manual clicks. Version control can show exactly what changed, reusable configuration can standardize common infrastructure patterns, and the same configuration can be used repeatedly. If a GUI generates or exports a reusable configuration artifact, the reusable artifact itself is effectively participating in an IaC-style workflow; the advantage comes from having infrastructure represented as code rather than from the GUI alone.</em></strong></blockquote><p>B is incorrect because using IaC does not automatically secure credentials. Terraform often needs credentials to call provider APIs, and teams must still protect those credentials using appropriate authentication and secret-management practices. Hard-coding credentials into Terraform configuration can make the security situation worse, especially if the configuration is committed to version control. Terraform has features for marking values sensitive and, in modern versions, for using ephemeral values in supported contexts, but credential security is a separate responsibility from the basic adoption of IaC.</p><p>C is incorrect because IaC does not change the cloud provider’s price for a given resource simply because the resource was provisioned from Terraform instead of a GUI. IaC can help teams standardize deployments, detect unnecessary changes, and automate cost-control practices, but the underlying service charges are based on the resources and services consumed. A virtual machine, database, or network service does not inherently become cheaper just because Terraform created it.</p><p>D is incorrect because IaC does not inherently block people or external systems from making out-of-band changes directly in the provider console or API. Terraform can detect differences between configuration, state, and remote objects during planning or refresh-related operations, and governance controls can restrict manual access, but the Terraform configuration by itself does not make manual modification impossible. Preventing such changes requires access controls, policies, and operational governance in addition to IaC.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/aws-get-started/infrastructure-as-code">developer.hashicorp.com/terraform/tutorials/aws-get-started/infrastructure-as-code</a></p><p><a href="http://developer.hashicorp.com/terraform/intro/core-workflow">developer.hashicorp.com/terraform/intro/core-workflow</a></p><p><a href="http://developer.hashicorp.com/terraform/language/manage-sensitive-data">developer.hashicorp.com/terraform/language/manage-sensitive-data</a></p><p><strong>Q3.</strong></p><p><strong>The fictional company Cinderlace Infrastructure often uses direct API calls to create and manage cloud infrastructure. In what way does Terraform differ from conventional direct infrastructure-management approaches?</strong></p><p><strong>A. Terraform replaces cloud provider APIs with its own protocols, enabling automated deployments.</strong></p><p><strong>B. Terraform describes infrastructure with version-controlled, repeatable configurations that specify the desired state.</strong></p><p><strong>C. Terraform is merely a wrapper for cloud provider APIs, so there is little to no difference in calling the API directly.</strong></p><p><strong>D. Terraform enforces infrastructure through imperative scripts to ensure tasks are completed in the proper order.</strong></p><p>A is incorrect because Terraform does not replace an infrastructure platform’s APIs with a universal Terraform protocol. Terraform providers integrate with upstream cloud, SaaS, and infrastructure APIs and translate Terraform resource operations into interactions with those systems. The provider abstraction gives Terraform a consistent workflow, but the underlying service APIs continue to exist and remain the mechanisms through which most providers manage remote objects.</p><blockquote><strong><em>B is correct because Terraform uses declarative configuration to describe the desired end state of infrastructure. Those configuration files can be stored in version control, reviewed, reused, and applied repeatedly. Terraform also maintains state, builds a dependency graph, and creates an execution plan that identifies the actions required to reconcile managed infrastructure with the configuration. This is substantially different from manually issuing individual API calls and personally tracking the sequence and current state of each object.</em></strong></blockquote><p>C is incorrect because describing Terraform as merely an API wrapper omits major parts of Terraform’s behavior. Providers do communicate with APIs, but Terraform Core evaluates configuration and expressions, tracks resource instances in state, infers dependencies, builds a graph, plans changes, handles resource lifecycle behavior, and coordinates provider operations. These capabilities create a declarative infrastructure-management workflow that direct API calls do not provide on their own.</p><p>D is incorrect because Terraform configuration is primarily declarative rather than an imperative list of API operations. Users describe the intended resources and relationships, while Terraform determines ordering from dependencies and provider requirements. Terraform can support provisioners for special cases, but provisioners do not turn the Terraform language into a general imperative scripting model and HashiCorp recommends purpose-built approaches over provisioners when possible.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/aws-get-started/infrastructure-as-code">developer.hashicorp.com/terraform/tutorials/aws-get-started/infrastructure-as-code</a></p><p><a href="http://developer.hashicorp.com/terraform/language/providers">developer.hashicorp.com/terraform/language/providers</a></p><p><a href="http://developer.hashicorp.com/terraform/internals/graph">developer.hashicorp.com/terraform/internals/graph</a></p><p><strong>Q4.</strong></p><p><strong>The fictional company Orivanta Labs is reviewing workflows enabled by Infrastructure as Code. Which of these workflows is specifically enabled by treating infrastructure changes as code?</strong></p><p><strong>A. Automatic scaling of resources based on application load.</strong></p><p><strong>B. Cost optimization of infrastructure deployment.</strong></p><p><strong>C. Role-based access control of cloud resources.</strong></p><p><strong>D. Reviewing the proposed changes for potential security issues.</strong></p><p>A is incorrect because automatic scaling is a runtime capability provided by services such as autoscaling groups, container platforms, or other infrastructure systems. Terraform can create and configure those scaling resources, but scaling in response to load does not require the infrastructure to be managed as code. The provider platform can perform autoscaling regardless of whether the original configuration was entered through Terraform, another IaC tool, or a GUI.</p><p>B is incorrect because cost optimization is not unique to IaC. Teams can select less expensive resource types, shut down unused services, negotiate pricing, or use provider cost-management tools without adopting Terraform. IaC can help enforce or repeat cost-conscious patterns, but it does not create a uniquely IaC-only cost-optimization workflow.</p><p>C is incorrect because role-based access control is a capability of identity and access-management systems. Terraform can provision roles, policies, and permissions where providers expose those resources, but cloud platforms can also configure RBAC directly through their APIs or consoles. RBAC therefore does not depend on using IaC.</p><blockquote><strong><em>D is correct because representing infrastructure changes as version-controlled configuration and generating an execution plan creates a reviewable artifact before the change is applied. Reviewers can inspect the configuration diff and planned infrastructure actions for unexpected exposure, destructive changes, overly broad access, or other security concerns. Terraform’s team workflow explicitly uses plans and pull-request review to evaluate changes before applying them, making this the option most specifically associated with an IaC workflow.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/intro/core-workflow">developer.hashicorp.com/terraform/intro/core-workflow</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><strong>Q5.</strong></p><p><strong>The fictional company Vantrelle Technologies wants to experiment interactively with Terraform expressions before placing them into configuration. Which command should it use?</strong></p><p><strong>A. terraform console</strong></p><p><strong>B. terraform workspace</strong></p><p><strong>C. terraform validate</strong></p><p><strong>D. terraform test</strong></p><blockquote><strong><em>A is correct because terraform console opens an interactive command-line console that evaluates Terraform expressions. It is specifically designed for experimenting with expression syntax, built-in functions, references, and values available in the current configuration or state. This makes it the appropriate command when a practitioner wants immediate feedback while exploring how a Terraform expression evaluates.</em></strong></blockquote><p>B is incorrect because terraform workspace is a command group for managing Terraform CLI workspaces, which are separate instances of state data associated with the same working directory. It can list, create, select, show, and delete workspaces, but it is not an expression evaluator.</p><p>C is incorrect because terraform validate checks whether a Terraform configuration is syntactically valid and internally consistent. It is useful for detecting invalid configuration before planning, but it does not provide an interactive prompt where arbitrary expressions can be entered and evaluated. A valid configuration can be checked with validate, while expression experimentation belongs in terraform console.</p><p>D is incorrect because terraform test runs Terraform test files and validates assertions and test behavior defined for modules or configurations. It is useful for automated testing, not as an interactive expression scratchpad. A practitioner who wants to type an expression and immediately inspect the result should use terraform console.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/console">developer.hashicorp.com/terraform/cli/commands/console</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/workspace">developer.hashicorp.com/terraform/cli/commands/workspace</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands">developer.hashicorp.com/terraform/cli/commands</a></p><p><strong>Q6.</strong></p><p><strong>The fictional company Zephyrcove Systems wants to use providers published by third parties. You can install Community and Partner providers using terraform init.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><blockquote><strong><em>A is correct because terraform init examines the configuration’s provider requirements and installs the required provider plugins. Providers published in the public Terraform Registry can include Community, Partner, Partner Premier, and Official providers, and Terraform uses the provider source address to locate the appropriate provider package. The key requirement is that the configuration declares the provider correctly and that Terraform can reach an allowed installation source or configured mirror.</em></strong></blockquote><p>B is incorrect because terraform init is precisely the command that initializes a working directory and installs required providers. Third-party providers are not excluded merely because they are maintained by a community maintainer or HashiCorp technology partner. Terraform can also install providers from third-party registries or configured mirrors when their source addresses and CLI installation settings are valid, so installation is not restricted only to providers maintained by HashiCorp.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/init">developer.hashicorp.com/terraform/cli/commands/init</a></p><p><a href="http://developer.hashicorp.com/terraform/registry/providers">developer.hashicorp.com/terraform/registry/providers</a></p><p><a href="http://developer.hashicorp.com/terraform/docs/partnerships">developer.hashicorp.com/terraform/docs/partnerships</a></p><p><strong>Q7.</strong></p><p><strong>The fictional company Quenlora Systems needs to determine from which paths Terraform is loading the providers referenced by its .tf files. How can an engineer enable detailed Terraform logging to obtain this diagnostic information?</strong></p><p><strong>A. Set the environment variable TF_VAR_log=TRACE.</strong></p><p><strong>B. Set the environment variable TF_LOG=TRACE.</strong></p><p><strong>C. Set the environment variable TF_LOG=PATH.</strong></p><p><strong>D. Set verbose logging for each provider in the Terraform configuration.</strong></p><p>A is incorrect because environment variables beginning with TF_VAR_ are used to assign values to declared Terraform input variables. TF_VAR_log=TRACE would therefore attempt to provide the value TRACE to an input variable named log, assuming such a variable exists. It does not enable Terraform diagnostic logging and will not cause Terraform Core to emit detailed provider-loading traces.</p><blockquote><strong><em>B is correct because TF_LOG controls Terraform’s diagnostic log verbosity, and TRACE is the most verbose standard log level. Setting TF_LOG=TRACE causes detailed logs to be emitted to standard error and is the appropriate general-purpose setting when investigating internal operations such as plugin discovery and provider loading. If persistent output is required, TF_LOG_PATH can additionally specify a file, but TF_LOG or an applicable subsystem logging variable must still enable logging.</em></strong></blockquote><p>C is incorrect because PATH is not a valid logging level for TF_LOG. Valid logging levels include TRACE, DEBUG, INFO, WARN, and ERROR, while TF_LOG_PATH is a separate environment variable that specifies the file path where enabled log output should be written. Setting TF_LOG=PATH therefore confuses the logging-level variable with the output-path setting.</p><p>D is incorrect because generic Terraform provider-loading diagnostics are not enabled by adding a verbose setting to every provider block in the Terraform configuration. Terraform logging is controlled through logging environment variables such as TF_LOG, TF_LOG_CORE, and TF_LOG_PROVIDER rather than through a universal provider-block argument. Individual providers can have their own provider-specific configuration arguments, but there is no standard Terraform configuration argument named verbose that must be set on each provider for this purpose.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/config/environment-variables">developer.hashicorp.com/terraform/cli/config/environment-variables</a></p><p><a href="http://developer.hashicorp.com/terraform/internals/debugging">developer.hashicorp.com/terraform/internals/debugging</a></p><p><a href="http://developer.hashicorp.com/terraform/plugin/log/managing">developer.hashicorp.com/terraform/plugin/log/managing</a></p><p><strong>Q8.</strong></p><p><strong>The fictional company Brinovex Digital has written a new Terraform configuration. Which command must it run before running terraform plan or terraform apply for the first time in that working directory?</strong></p><p><strong>A. terraform validate</strong></p><p><strong>B. terraform workspace</strong></p><p><strong>C. terraform import</strong></p><p><strong>D. terraform init</strong></p><p>A is incorrect because terraform validate checks whether a Terraform configuration is syntactically valid and internally consistent, but it does not initialize the working directory. Validation itself normally requires the directory to have already been initialized so that referenced providers and modules are available. Therefore, validate does not replace the required initialization step before normal planning or applying.</p><p>B is incorrect because terraform workspace is a command group for managing Terraform CLI workspaces, such as listing, creating, selecting, and deleting them. Selecting or creating a workspace does not perform provider installation, module installation, or backend initialization. A new working directory still needs terraform init before normal plan or apply operations.</p><p>C is incorrect because terraform import associates an existing remote object with a Terraform resource address in state. Import is used when bringing existing infrastructure under Terraform management; it is not the command that prepares a new working directory for normal Terraform operations. The working directory must be initialized before commands that depend on providers and backend configuration can operate correctly.</p><blockquote><strong><em>D is correct because terraform init initializes a Terraform working directory and is the required setup step before normal terraform plan or terraform apply operations. During initialization, Terraform configures the selected backend, installs required provider plugins, retrieves referenced modules, and creates or updates initialization metadata. Terraform also indicates when reinitialization is required after relevant backend, module, or provider-requirement changes.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/init">developer.hashicorp.com/terraform/cli/init</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/init">developer.hashicorp.com/terraform/cli/commands/init</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><strong>Q9.</strong></p><p><strong>The fictional company Navoryx Cloud has reviewed a Terraform execution plan and approves it. What does terraform apply change when it successfully executes the approved plan? (Choose two.)</strong></p><p><strong>A. The Terraform state</strong></p><p><strong>B. The cloud infrastructure</strong></p><p><strong>C. The execution plan</strong></p><p><strong>D. The Terraform code</strong></p><p><strong>E. The .terraform directory</strong></p><blockquote><strong><em>A is correct because a successful terraform apply updates Terraform state to record the resulting bindings and current known attributes of the managed resource instances. State is how Terraform associates resource instances in the configuration with real infrastructure objects. Depending on the configured backend, that state may be stored locally or remotely, but applying changes updates the workspace’s state to reflect the results of the operation.</em></strong></blockquote><blockquote><strong><em>B is correct because terraform apply performs the resource actions in the execution plan against the target provider APIs. Those actions can create, update, replace, or destroy infrastructure so that the real objects move toward the desired configuration. This is the primary distinction between terraform plan, which previews proposed changes, and terraform apply, which performs approved changes.</em></strong></blockquote><p>C is incorrect because an execution plan is an input or intermediate description of the actions Terraform intends to perform; applying it does not modify the already-created plan into a new desired configuration. If a saved plan file is supplied to terraform apply, Terraform executes the actions recorded in that plan. A future planning operation creates a new plan based on the then-current configuration, state, and remote objects.</p><p>D is incorrect because terraform apply does not rewrite the Terraform configuration files that express the desired infrastructure. The configuration remains source code controlled by the user, while apply changes managed infrastructure and updates state. Commands such as terraform fmt may rewrite formatting, but apply does not automatically modify the Terraform code to match provider-side results.</p><p>E is incorrect because the .terraform directory is primarily initialization metadata and cache data, including installed provider plugins, downloaded modules, the current CLI workspace selection, and backend-related metadata. Although other Terraform operations can update information associated with the working directory, changing .terraform is not one of the two conceptual results the question asks about after approving and successfully applying an execution plan. The principal changes are to real infrastructure and Terraform state.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><a href="http://developer.hashicorp.com/terraform/language/resources">developer.hashicorp.com/terraform/language/resources</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state">developer.hashicorp.com/terraform/language/state</a></p><p><strong>Q10.</strong></p><p><strong>The fictional company Veltrona Labs is deploying new infrastructure with Terraform. Which is the correct workflow?</strong></p><p><strong>A. 1. Write Terraform configuration.</strong></p><p><strong>2. Run terraform apply to create infrastructure.</strong></p><p><strong>3. Use terraform validate to confirm Terraform deployed resources correctly.</strong></p><p><strong>B. 1. Write Terraform configuration.</strong></p><p><strong>2. Run terraform init to initialize the working directory.</strong></p><p><strong>3. Run terraform apply.</strong></p><p><strong>C. 1. Write Terraform configuration.</strong></p><p><strong>2. Run terraform plan to initialize the working directory.</strong></p><p><strong>3. Run terraform apply to create the infrastructure.</strong></p><p><strong>D. 1. Write Terraform configuration.</strong></p><p><strong>2. Run terraform show to view proposed changes.</strong></p><p><strong>3. Run terraform apply to create new infrastructure.</strong></p><p>A is incorrect because terraform validate is a configuration validation command, not a post-deployment check that verifies whether resources were successfully created in the provider platform. In addition, a newly created working directory normally needs terraform init before apply so that Terraform can initialize its backend and install required providers and modules. Therefore, this sequence omits an essential initialization step and misstates the purpose of validate.</p><blockquote><strong><em>B is correct because after authoring a new Terraform configuration, terraform init prepares the working directory by initializing backend settings and installing the required providers and modules. terraform apply can then create an execution plan, display it for approval in interactive use, and perform the resulting infrastructure changes. Running a separate terraform plan between init and apply is a common and recommended review step, but it is not strictly required because terraform apply can create its own plan when no saved plan file is supplied.</em></strong></blockquote><p>C is incorrect because terraform plan does not initialize an uninitialized Terraform working directory. Planning depends on initialization having already installed required providers and modules and prepared the backend. The correct initialization command is terraform init; plan is used afterward to preview proposed changes.</p><p>D is incorrect because terraform show displays human-readable information from Terraform state or a saved plan file; it does not create the normal execution plan required for an initial deployment and it does not initialize the working directory. You would ordinarily run terraform init before planning or applying. terraform show is useful for inspecting existing state or a previously saved plan, not as a substitute for initialization.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/intro/core-workflow">developer.hashicorp.com/terraform/intro/core-workflow</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/init">developer.hashicorp.com/terraform/cli/init</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><strong>Q11.</strong></p><p><strong>The fictional company Nexorune Technologies has not run terraform plan separately. terraform apply will fail because a plan must always be generated with a separate terraform plan command first.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><p>A is incorrect because a separate terraform plan command is not mandatory before a normal interactive terraform apply. When terraform apply is run without a saved plan file, Terraform automatically creates a new execution plan, displays the proposed actions, and requests approval before performing them. A separate plan step is useful for review and is common in team or automation workflows, but it is not a prerequisite for the apply command.</p><blockquote><strong><em>B is correct because terraform apply supports an automatic planning mode when no saved plan is supplied. Terraform creates the plan immediately before the apply, prompts for approval unless an applicable noninteractive option changes that behavior, and then executes the approved actions. A previously saved plan is optional and is primarily useful when you need to review and preserve an exact plan for later application.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/cli/apply">developer.hashicorp.com/terraform/tutorials/cli/apply</a></p><p><strong>Q12.</strong></p><p><strong>The fictional company Daxenmere Labs runs terraform validate against a Terraform configuration. Which condition will cause validation to return an error?</strong></p><p><strong>A. None of these will return an error.</strong></p><p><strong>B. The configuration references an input variable, but no corresponding variable block is declared.</strong></p><p><strong>C. The Terraform state does not match the current remote infrastructure.</strong></p><p><strong>D. The configuration uses tabs for indentation instead of spaces.</strong></p><p>A is incorrect because Terraform validation can detect configuration errors such as references to undeclared input variables. terraform validate checks syntax and internal consistency, including whether references resolve to declared objects in the configuration. Therefore, at least one of the listed conditions can cause validate to fail.</p><blockquote><strong><em>B is correct because a reference such as var.region requires a corresponding variable “region” block to declare that root-module input variable. If the configuration references an undeclared input variable, Terraform cannot resolve that reference as a valid configuration object and terraform validate reports an error. This is an internal configuration-consistency issue, which is exactly the kind of problem validate is intended to detect.</em></strong></blockquote><p>C is incorrect because terraform validate does not evaluate existing state and does not compare that state with remote infrastructure. State synchronization and drift detection occur in run operations such as terraform plan, which communicate with providers and inspect current managed objects. A stale or divergent state therefore is not, by itself, a syntax or internal-consistency error for terraform validate.</p><p>D is incorrect because tabs versus spaces is a formatting and style concern rather than an invalid Terraform-language construct that validate rejects merely on indentation choice. terraform fmt is the command that rewrites Terraform configuration into the canonical style. A configuration can therefore validate successfully even if its current indentation does not match the canonical formatting that terraform fmt would produce.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/validate">developer.hashicorp.com/terraform/cli/commands/validate</a></p><p><a href="http://developer.hashicorp.com/terraform/language/values/variables">developer.hashicorp.com/terraform/language/values/variables</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/fmt">developer.hashicorp.com/terraform/cli/commands/fmt</a></p><p><strong>Q13.</strong></p><p><strong>The fictional company Meridionix Cloud is using terraform plan during development. How can terraform plan aid the development process?</strong></p><p><strong>A. It lets you compare your expectations with the proposed execution plan without permanently modifying managed infrastructure or committing normal plan-time refresh information to state.</strong></p><p><strong>B. It reconciles Terraform state with deployed resources and permanently writes that refreshed state as its primary behavior.</strong></p><p><strong>C. It initializes the working directory that contains the Terraform configuration files.</strong></p><p><strong>D. It formats Terraform configuration files.</strong></p><blockquote><strong><em>A is correct because terraform plan creates an execution plan that previews the actions Terraform proposes to make based on the configuration, prior state, and current remote objects. A normal plan lets developers check whether the proposed changes match their intent without actually carrying out those infrastructure actions. Although normal planning reads remote objects to build an up-to-date in-memory view, it does not persist those refresh results to state merely by completing a speculative plan.</em></strong></blockquote><p>B is incorrect because permanently synchronizing state with remote-object changes is not the primary behavior of a normal terraform plan. Refresh-only planning can preview state-only changes, but persisting those changes requires a refresh-only apply operation such as terraform apply -refresh-only. A normal plan should therefore not be described as permanently rewriting state merely to reconcile it with remote infrastructure.</p><p>C is incorrect because terraform plan does not initialize a Terraform working directory. terraform init performs initialization, including backend setup and installation of required providers and modules. Planning is performed after initialization and is used to preview proposed changes.</p><p>D is incorrect because formatting configuration is the job of terraform fmt, which rewrites Terraform files into Terraform’s canonical style. terraform plan may expose configuration problems while evaluating a proposed run, but it does not rewrite source formatting. Its purpose is to calculate and display proposed changes rather than format configuration files.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><a href="http://developer.hashicorp.com/terraform/intro/core-workflow">developer.hashicorp.com/terraform/intro/core-workflow</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state/purpose">developer.hashicorp.com/terraform/language/state/purpose</a></p><p><strong>Q14.</strong></p><p><strong>The fictional company Nexalith Works is initializing a Terraform working directory. Which task does terraform init not perform?</strong></p><p><strong>A. Sources any remote modules and downloads them.</strong></p><p><strong>B. Connects to the configured backend.</strong></p><p><strong>C. Validates that values are set for all required input variables.</strong></p><p><strong>D. Sources all providers used in the configuration and downloads them.</strong></p><p>A is incorrect because installing referenced child modules is one of the initialization tasks performed by terraform init. Terraform examines module blocks and retrieves module source code from the locations specified by their source arguments. Remote module packages are installed into Terraform’s working-directory data so that subsequent operations can evaluate the full module tree.</p><p>B is incorrect because backend initialization is one of the main responsibilities of terraform init. Terraform initializes the backend configured for the root module and prepares access to the state associated with that backend. Depending on backend changes and migration circumstances, initialization can also require additional backend-related input or confirmation.</p><blockquote><strong><em>C is correct because terraform init does not generally verify that runtime values have been supplied for every required root-module input variable. Terraform evaluates most input variables when it creates a plan, and a variable without a default normally requires a value before Terraform can generate that plan. Current Terraform also supports const variables for certain early operations, but that does not change the general rule that init is not the command whose job is to confirm values for all required runtime input variables.</em></strong></blockquote><p>D is incorrect because provider installation is a core terraform init responsibility. Terraform analyzes provider requirements, selects versions according to configured constraints and the dependency lock file, and downloads or otherwise installs the required provider plugins using the configured provider installation methods. A working directory normally must be initialized before planning or applying so the necessary providers are available.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/init">developer.hashicorp.com/terraform/cli/commands/init</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/init">developer.hashicorp.com/terraform/cli/init</a></p><p><a href="http://developer.hashicorp.com/terraform/language/values/variables">developer.hashicorp.com/terraform/language/values/variables</a></p><p><strong>Q15.</strong></p><p><strong>The fictional company Orbinex Systems needs Terraform to deliberately replace one managed resource during its next apply. Why would Orbinex Systems use the -replace flag with terraform apply?</strong></p><p><strong>A. You want to force Terraform to destroy a resource on the next apply.</strong></p><p><strong>B. You want Terraform to ignore a resource on the next apply.</strong></p><p><strong>C. You want Terraform to destroy all the infrastructure in your workspace.</strong></p><p><strong>D. You want to force Terraform to destroy and recreate a resource on the next apply.</strong></p><p>A is incorrect because -replace does not request a destroy-only action for the selected resource. It tells Terraform to plan a replacement, which means Terraform will create a replacement instance as part of the operation as well as remove the existing instance, with the exact ordering influenced by resource and lifecycle behavior. If the objective were only to remove infrastructure, replacement would be the wrong operation.</p><p>B is incorrect because -replace does the opposite of ignoring a resource: it explicitly selects a resource instance that Terraform should replace even if Terraform would otherwise plan an in-place update or no change. Ignoring particular attribute changes is handled through lifecycle configuration such as ignore_changes, not through -replace.</p><p>C is incorrect because -replace targets one explicitly addressed resource instance per -replace option; it does not switch Terraform into a mode that destroys the entire workspace’s managed infrastructure. Destroying all managed objects is handled by terraform destroy or the destroy planning mode, not by the resource-replacement option.</p><blockquote><strong><em>D is correct because -replace=ADDRESS instructs Terraform to plan replacement of the specified resource instance instead of an update or no-op. This is useful when an object is degraded, misconfigured in a way not otherwise detected, or needs to be reprovisioned even though its configuration may not require replacement. The -replace workflow is Terraform’s current explicit mechanism for requesting replacement through the plan or apply workflow.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><a href="http://developer.hashicorp.com/terraform/language/resources/configure">developer.hashicorp.com/terraform/language/resources/configure</a></p><p><strong>Q16.</strong></p><p><strong>The fictional company Kalystron Digital wants to understand how Terraform orders resource operations while creating an execution plan. How does Terraform determine dependencies between resources?</strong></p><p><strong>A. Terraform builds a resource graph based on your configuration and your state data, if present.</strong></p><p><strong>B. Terraform requires resources in your configuration to be listed in the order they will be created to determine dependencies.</strong></p><p><strong>C. Terraform requires all dependencies between resources to be specified using the depends_on meta-argument.</strong></p><p><strong>D. Terraform requires resource dependencies to be defined as modules and sourced in order.</strong></p><blockquote><strong><em>A is correct because Terraform constructs a dependency graph from the configuration and incorporates relevant planning and state information when building the graph for an operation. References from one resource’s expressions to attributes of another create implicit dependency edges, while explicit depends_on declarations can add dependencies that Terraform cannot infer from data references. Terraform then uses the graph to determine a safe operation order and can process independent objects in parallel.</em></strong></blockquote><p>B is incorrect because the physical order of resource blocks in Terraform configuration files does not determine creation order. Terraform configuration is declarative, so resources can be arranged for readability rather than execution sequence. Dependencies are derived from references and explicit dependency declarations, and Terraform’s graph determines the operational order.</p><p>C is incorrect because depends_on is needed only for dependencies Terraform cannot infer automatically from expression references or other configuration relationships. When one resource argument references an attribute of another resource, Terraform recognizes that as an implicit dependency without requiring depends_on. HashiCorp recommends using expression references where possible and reserving depends_on for hidden dependencies.</p><p>D is incorrect because resources do not need to be wrapped in modules or sourced in a particular order to establish dependencies. Modules are a configuration-organization and reuse mechanism, and dependencies can exist directly between resources in the same module or across module boundaries through references. Terraform derives the execution relationship from its dependency graph rather than source ordering.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/internals/graph">developer.hashicorp.com/terraform/internals/graph</a></p><p><a href="http://developer.hashicorp.com/terraform/language/meta-arguments/depends_on">developer.hashicorp.com/terraform/language/meta-arguments/depends_on</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/configuration-language/dependencies">developer.hashicorp.com/terraform/tutorials/configuration-language/dependencies</a></p><p><strong>Q17.</strong></p><p><strong>The fictional company Merovanta Tech has a main.tf configuration containing an application server, a database, and a load balancer. Merovanta Tech runs terraform apply and Terraform successfully creates all three resources. The team later decides that it does not need the load balancer and runs terraform destroy without any flags. What will happen?</strong></p><p><strong>A. Terraform will prompt you to confirm that you want to destroy all the infrastructure.</strong></p><p><strong>B. Terraform will prompt you to pick which resource you want to destroy.</strong></p><p><strong>C. Terraform will immediately destroy all the infrastructure.</strong></p><p><strong>D. Terraform will destroy the application server because it is listed first in the code.</strong></p><p><strong>E. Terraform will destroy the main.tf file.</strong></p><blockquote><strong><em>A is correct because terraform destroy is the convenience command for creating and applying a destroy plan for all objects managed by the current Terraform configuration. Without -auto-approve, Terraform presents the proposed destruction and asks for confirmation before proceeding. Therefore, running terraform destroy merely because one resource is no longer needed would propose destruction of all managed infrastructure in that configuration, not just the load balancer.</em></strong></blockquote><p>B is incorrect because terraform destroy without a targeting option does not interactively ask the user to select an individual managed resource. Its default behavior is to propose destroying all objects managed by the configuration. If only the load balancer should be removed from normal management, the usual declarative workflow is to remove its resource block from the configuration, review terraform plan, and apply the resulting change.</p><p>C is incorrect because terraform destroy without flags does not normally proceed immediately. Terraform shows the destroy plan and requests explicit approval before taking destructive actions. The -auto-approve option can suppress the interactive approval prompt, but the question states that no flags were used.</p><p>D is incorrect because Terraform does not destroy resources based on their textual order in main.tf. Terraform builds a dependency graph and uses dependency relationships to order operations safely. Furthermore, an unqualified terraform destroy proposes destroying all managed objects, not simply whichever resource block appears first in the file.</p><p>E is incorrect because Terraform manages infrastructure objects and state; terraform destroy does not delete the Terraform configuration source file. The main.tf file remains on disk unless a user or some other filesystem operation removes it. Destroying infrastructure does not erase the configuration that described that infrastructure.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/destroy">developer.hashicorp.com/terraform/cli/commands/destroy</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/configuration-language/dependencies">developer.hashicorp.com/terraform/tutorials/configuration-language/dependencies</a></p><p><strong>Q18.</strong></p><p><strong>The fictional company Klyrion Cloud needs Terraform to construct a collection of repeatable nested configuration blocks from a collection value. What type of block is designed for this purpose?</strong></p><p><strong>A. dynamic</strong></p><p><strong>B. nesting</strong></p><p><strong>C. for_each</strong></p><p><strong>D. repeated</strong></p><blockquote><strong><em>A is correct because a dynamic block is Terraform’s language feature for constructing repeatable nested configuration blocks programmatically. It iterates over a collection using its for_each argument and emits one nested block for each selected element using the dynamic block’s content definition.</em></strong></blockquote><p>B is incorrect because nesting is not a Terraform block type that dynamically generates nested blocks. Resources and other constructs can have provider-defined nested blocks, but Terraform uses the dynamic block language feature when those repeatable nested blocks need to be generated from collection data.</p><p>C is incorrect because for_each is a meta-argument or iteration argument rather than a standalone block type for constructing nested blocks. A dynamic block commonly contains for_each internally, but the block type that actually generates the nested configuration blocks is dynamic.</p><p>D is incorrect because repeated is not a Terraform configuration block type. Repeatable nested blocks are either written explicitly according to the provider schema or generated with a dynamic block when iteration is required.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/expressions/dynamic-blocks">developer.hashicorp.com/terraform/language/expressions/dynamic-blocks</a></p><p><a href="http://developer.hashicorp.com/terraform/language/expressions">developer.hashicorp.com/terraform/language/expressions</a></p><p><a href="http://developer.hashicorp.com/terraform/language/meta-arguments/for_each">developer.hashicorp.com/terraform/language/meta-arguments/for_each</a></p><p><strong>Q19.</strong></p><p><strong>The fictional company Lumetora Systems has a root Terraform module that calls several child modules. When Lumetora Systems runs terraform apply, Terraform CLI automatically prints output values declared inside both the root module and every child module.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><p>A is incorrect because Terraform CLI displays root-module output values after an apply, but child-module outputs are not automatically printed just because the child module declares them. Child outputs form an interface that the parent module can reference. If an operator wants a child-module value displayed as a root result, the root module must declare its own output that references the child output.</p><blockquote><strong><em>B is correct because only outputs exposed by the root module are automatically presented as the configuration’s top-level outputs after a normal apply. A child module can expose a value to its parent through an output block, and the parent accesses it with module.&lt;CHILD_MODULE_NAME&gt;.&lt;OUTPUT_NAME&gt;. The parent must re-export that value through a root output if it should appear as a root-module output in Terraform CLI.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/values/outputs">developer.hashicorp.com/terraform/language/values/outputs</a></p><p><a href="http://developer.hashicorp.com/terraform/language/block/output">developer.hashicorp.com/terraform/language/block/output</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/modules/module-use">developer.hashicorp.com/terraform/tutorials/modules/module-use</a></p><p><strong>Q20.</strong></p><p><strong>The fictional company Astranova Cloud is evaluating modules published in the public Terraform Registry. All modules published in the public Terraform Registry are maintained and tested by HashiCorp.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><p>A is incorrect because the public Terraform Registry is an open ecosystem and does not contain only modules maintained by HashiCorp. HashiCorp documentation states that modules in the public registry are created and maintained by HashiCorp, HashiCorp partners, and members of the Terraform community. Partner modules can receive additional review for stability and compatibility, but that does not make every module in the registry a HashiCorp-maintained or HashiCorp-tested module.</p><blockquote><strong><em>B is correct because anyone who satisfies the public module publishing requirements can publish a module to the public Terraform Registry. The registry includes modules from HashiCorp, partners, and the broader community, so users should review the module owner, source repository, documentation, version history, and other relevant trust indicators before adopting a module.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/registry">developer.hashicorp.com/terraform/registry</a></p><p><a href="http://developer.hashicorp.com/terraform/registry/modules/publish">developer.hashicorp.com/terraform/registry/modules/publish</a></p><p><a href="http://developer.hashicorp.com/terraform/registry/modules/use">developer.hashicorp.com/terraform/registry/modules/use</a></p><p><strong>Q21.</strong></p><p><strong>The fictional company Merovanta Systems provisioned virtual machines on Google Cloud using the Google Cloud CLI (gcloud). It now wants to bring those existing VMs under Terraform management using the configuration-driven import workflow represented by the choices below. Which two listed elements are part of that workflow? (Choose two.)</strong></p><p><strong>A. Add an import block to the configuration.</strong></p><p><strong>B. Run terraform apply -refresh-only.</strong></p><p><strong>C. Add a resource block for the existing VM.</strong></p><p><strong>D. Run terraform state pull.</strong></p><blockquote><strong><em>A is correct because an import block tells Terraform which existing remote object to associate with a Terraform resource address. The block specifies a destination through the to argument and identifies the existing object using an id or, for supported resources and providers, an identity value. Terraform can then perform the import as part of the normal plan and apply workflow.</em></strong></blockquote><p>B is incorrect because terraform apply -refresh-only reconciles state with changes to remote objects that Terraform already tracks. It does not create a new state binding for an independently created VM that is not yet managed by Terraform. Import is required to establish that management relationship.</p><blockquote><strong><em>C is correct because Terraform needs resource configuration for the object it will manage after import. In the standard configuration-driven workflow, the import target corresponds to a resource block that describes the desired configuration. Current Terraform can also generate configuration for some imports by using an import block with the configuration-generation workflow, but the end result still requires resource configuration for ongoing management.</em></strong></blockquote><p>D is incorrect because terraform state pull downloads and prints the current Terraform state from its storage location. It is an inspection and state-retrieval command; it does not associate a previously unmanaged Google Cloud VM with a Terraform resource address.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/block/import">developer.hashicorp.com/terraform/language/block/import</a></p><p><a href="http://developer.hashicorp.com/terraform/language/import/single-resource">developer.hashicorp.com/terraform/language/import/single-resource</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/import">developer.hashicorp.com/terraform/cli/import</a></p><p><strong>Q22.</strong></p><p><strong>The fictional company Orynthia Cloud changes its Terraform configuration to use a new backend and needs to migrate the existing state to that backend. Which Terraform command is used for backend reinitialization and state migration?</strong></p><p><strong>A. terraform init</strong></p><p><strong>B. terraform state rm</strong></p><p><strong>C. terraform push</strong></p><p><strong>D. terraform apply</strong></p><blockquote><strong><em>A is correct because backend configuration is initialized and changed through terraform init. With current Terraform, when migrating existing state to a changed backend, the explicit supported form is terraform init -migrate-state; Terraform can then copy existing state to the newly configured backend and may prompt for migration decisions. Therefore, terraform init is the correct command family among the listed choices, with -migrate-state providing the current explicit migration behavior.</em></strong></blockquote><p>B is incorrect because terraform state rm removes selected resource bindings from Terraform state without destroying the corresponding remote objects. It does not change the configured backend or migrate the state snapshot from one backend to another.</p><p>C is incorrect because terraform push is not the Terraform CLI command for backend migration. Terraform has a terraform state push subcommand for manually uploading a local state file to the configured backend, but that low-level state operation is not the supported backend reconfiguration and migration workflow described in the question.</p><p>D is incorrect because terraform apply performs infrastructure and state changes represented by an execution plan after the working directory and backend are properly initialized. A backend change requires reinitialization first, and Terraform directs users to terraform init rather than using apply to migrate state storage.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/init">developer.hashicorp.com/terraform/cli/commands/init</a></p><p><a href="http://developer.hashicorp.com/terraform/language/backend">developer.hashicorp.com/terraform/language/backend</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/cloud/settings">developer.hashicorp.com/terraform/cli/cloud/settings</a></p><p><strong>Q23.</strong></p><p><strong>The fictional company Velmoryn Systems wants to bring an existing database under Terraform management by using an id-based import block. What information is required to define the import action? (Choose two.)</strong></p><p><strong>A. The connection string that Terraform will use to connect to and manage the database.</strong></p><p><strong>B. The import ID associated with the existing database on the provider platform.</strong></p><p><strong>C. The database platform and version that the existing resource is running.</strong></p><p><strong>D. The path to the .tf file that contains the database resource block.</strong></p><p><strong>E. The destination resource address of the block that will manage the database.</strong></p><p>A is incorrect because an import block does not require a database connection string as a Terraform language argument. Terraform communicates with the remote platform through the configured provider, and provider authentication or connection settings are handled separately from the import block. The import block identifies which existing object should be imported and which Terraform resource address should receive that object.</p><blockquote><strong><em>B is correct because an id-based import block needs an identifier that the relevant provider resource type recognizes for the existing remote object. The exact format of the import ID is provider-specific, so for one resource it might be a simple object ID while another resource might require a compound string. Current Terraform also supports provider-defined resource identity for supported resources, but when using the id-based workflow described here, the import ID is required.</em></strong></blockquote><p>C is incorrect because an import block does not require the database platform and software version merely to establish the import mapping. Those characteristics may later appear as resource attributes or arguments and may matter when reconciling the imported object’s configuration, but they are not the built-in information that identifies the import destination and the existing object.</p><p>D is incorrect because Terraform does not identify an import destination by the filesystem path of a .tf file. Terraform loads the configuration files in the module together, and the import block’s to argument specifies a Terraform resource address such as aws_db_instance.example. The physical filename containing the resource block is not part of the import identity.</p><blockquote><strong><em>E is correct because every import block requires a to argument that identifies the destination Terraform resource address. Terraform uses that address to bind the existing remote object to the corresponding resource instance in state. For a normal single-resource import, the destination address must correspond to the resource configuration that will manage the imported object.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/block/import">developer.hashicorp.com/terraform/language/block/import</a></p><p><a href="http://developer.hashicorp.com/terraform/language/import">developer.hashicorp.com/terraform/language/import</a></p><p><a href="http://developer.hashicorp.com/terraform/language/import/single-resource">developer.hashicorp.com/terraform/language/import/single-resource</a></p><p><strong>Q24.</strong></p><p><strong>The fictional company Solvirex Cloud corrects a typo in a Terraform resource label, changing aws_s3_bucket.photoes to aws_s3_bucket.photos. Solvirex wants Terraform to recognize that the same existing bucket now has a new Terraform resource address, without destroying and recreating the bucket. Which configuration should Solvirex use?</strong></p><p><strong>A. Remove the old resource from the configuration and re-import it.</strong></p><p><strong>B. Run terraform apply -refresh-only.</strong></p><p><strong>C. Do nothing because Terraform automatically detects resource-address renames.</strong></p><p><strong>D. Add a moved block to the configuration.</strong></p><p>A is incorrect because removing and re-importing can sometimes be used to re-establish state mappings, but it is unnecessarily disruptive and error-prone for a simple Terraform address rename. Terraform provides moved blocks specifically to describe resource-address changes declaratively and preserve the existing object through the refactor.</p><p>B is incorrect because refresh-only mode updates Terraform’s recorded state attributes to match changes detected in remote objects. It does not tell Terraform that a resource address in the configuration has been renamed from aws_s3_bucket.photoes to aws_s3_bucket.photos. Without an explicit address migration, Terraform can interpret the old address as removed and the new address as a new resource.</p><p>C is incorrect because Terraform does not infer that two differently named resource addresses represent the same real-world object merely because their configurations look similar. Changing a resource’s local label changes its Terraform address, so Terraform needs an explicit moved declaration or another supported state-migration operation to preserve the association.</p><blockquote><strong><em>D is correct because a moved block maps the previous resource address in its from argument to the new resource address in its to argument. During planning, Terraform recognizes that the existing object tracked at aws_s3_bucket.photoes should now be tracked as aws_s3_bucket.photos, allowing the rename to occur in state without destroying and recreating the bucket.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/block/moved">developer.hashicorp.com/terraform/language/block/moved</a></p><p><a href="http://developer.hashicorp.com/terraform/language/modules/develop/refactoring">developer.hashicorp.com/terraform/language/modules/develop/refactoring</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/state/move">developer.hashicorp.com/terraform/cli/state/move</a></p><p><strong>Q25.</strong></p><p><strong>The fictional company Avenlith Technologies has several users and automation jobs that can run Terraform against the same state. What Terraform feature prevents multiple state-writing operations from operating on that state at the same time when the selected backend supports the feature?</strong></p><p><strong>A. Remote state storage</strong></p><p><strong>B. Version control</strong></p><p><strong>C. State locking</strong></p><p><strong>D. Provider constraints</strong></p><p>A is incorrect because remote state storage centralizes Terraform state, but remote storage by itself does not necessarily provide locking. HashiCorp explicitly notes that not all backends support state locking. A backend must implement locking for Terraform to prevent concurrent state-writing operations in this way.</p><p>B is incorrect because version control records revisions of configuration files and other tracked artifacts but does not provide Terraform’s runtime state-locking semantics. Storing state directly in a typical version-control repository is also discouraged because state can contain secrets and ordinary version control does not coordinate Terraform state writes safely.</p><blockquote><strong><em>C is correct because state locking prevents another Terraform operation from acquiring the state lock while a state-writing operation already holds it. Terraform performs locking automatically for operations that could write state when the configured backend supports locking. This is the Terraform feature specifically designed to prevent concurrent state modifications.</em></strong></blockquote><p>D is incorrect because provider constraints control which provider versions Terraform may select. They help produce predictable dependency selections but have no role in coordinating concurrent access to Terraform state.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/state/locking">developer.hashicorp.com/terraform/language/state/locking</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state/backends">developer.hashicorp.com/terraform/language/state/backends</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state">developer.hashicorp.com/terraform/language/state</a></p><p><strong>Q26.</strong></p><p><strong>The fictional company Quenvaris Labs initially used Terraform’s default local backend and has already completed its first terraform apply. Changing the backend to another supported backend afterward is:</strong></p><p><strong>A. discouraged</strong></p><p><strong>B. mandatory</strong></p><p><strong>C. impossible</strong></p><p><strong>D. optional</strong></p><p>A is incorrect because HashiCorp does not prohibit or generally discourage changing a backend after infrastructure has already been applied. Terraform explicitly supports backend reconfiguration and state migration. Operators should perform the migration carefully and protect state, but changing backends is a supported lifecycle operation.</p><p>B is incorrect because Terraform does not require every configuration to leave the default local backend after its first apply. A user can continue using the local backend if it suits the workflow, although remote state is commonly recommended for teams and collaborative environments.</p><p>C is incorrect because Terraform can change both a backend’s configuration and the backend type. After changing the backend configuration, the user re-runs terraform init and can migrate existing state to the new backend, typically using terraform init -migrate-state for the explicit migration workflow.</p><blockquote><strong><em>D is correct because changing from the default local backend to a different supported backend after the first apply is optional. Terraform allows the backend configuration to be changed later and provides initialization and migration behavior so existing state can be copied to the new backend rather than requiring infrastructure to be recreated.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/backend">developer.hashicorp.com/terraform/language/backend</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/init">developer.hashicorp.com/terraform/cli/commands/init</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state/remote">developer.hashicorp.com/terraform/language/state/remote</a></p><p><strong>Q27.</strong></p><p><strong>The fictional company Teralynx Systems is evaluating Terraform backends. __________ backends support state locking.</strong></p><p><strong>A. Only local</strong></p><p><strong>B. Some</strong></p><p><strong>C. All</strong></p><p><strong>D. No</strong></p><p>A is incorrect because state locking is not exclusive to the local backend. Several backends can provide locking through their underlying storage or coordination mechanisms. The local backend itself can lock state using local system APIs, but it is not the only backend with locking capability.</p><blockquote><strong><em>B is correct because HashiCorp explicitly states that not all Terraform backends support state locking. Backends are responsible for implementing locking where possible, and the documentation for each backend describes whether and how it supports locking. Therefore, only some backends support this capability.</em></strong></blockquote><p>C is incorrect because state locking is an optional backend capability rather than a universal property of every Terraform backend. Users must check the selected backend’s current documentation before relying on locking behavior.</p><p>D is incorrect because Terraform does support state locking, and multiple backends implement it. When a backend supports locking, Terraform automatically attempts to acquire a lock for operations that could write state.</p><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/state/backends">developer.hashicorp.com/terraform/language/state/backends</a></p><p><a href="http://developer.hashicorp.com/terraform/language/state/locking">developer.hashicorp.com/terraform/language/state/locking</a></p><p><a href="http://developer.hashicorp.com/terraform/language/backend">developer.hashicorp.com/terraform/language/backend</a></p><p><strong>Q28.</strong></p><p><strong>The fictional company Quorivex Technologies is reviewing Terraform’s state refresh behavior. When must it explicitly run Terraform in refresh-only mode as a mandatory prerequisite?</strong></p><p><strong>A. Before every terraform plan.</strong></p><p><strong>B. Before every terraform apply.</strong></p><p><strong>C. Before every terraform import operation.</strong></p><p><strong>D. None of the above.</strong></p><p>A is incorrect because a normal terraform plan already performs an implicit in-memory refresh of managed remote objects by default before Terraform calculates proposed configuration changes. An explicit refresh-only operation is therefore not required before every plan. The -refresh=false option can disable the normal refresh behavior, but that is an explicit exception rather than the default.</p><p>B is incorrect because a normal terraform apply without a saved plan creates a fresh execution plan, and normal planning includes an implicit refresh by default. Terraform therefore does not require users to run a separate refresh-only operation before every apply.</p><p>C is incorrect because Terraform does not require a separate refresh-only operation as a mandatory prerequisite to importing an existing resource. Import workflows establish a binding between an existing remote object and a Terraform resource address, while refresh-only mode serves a different purpose: reconciling state and root output values with remote changes that occurred outside Terraform.</p><blockquote><strong><em>D is correct because none of the listed operations universally requires a separate explicit refresh-only operation beforehand. Terraform plan and normal apply perform an implicit refresh by default, and refresh-only mode is an optional specialized workflow for reviewing and accepting out-of-band changes into Terraform state without changing the remote objects.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/plan">developer.hashicorp.com/terraform/cli/commands/plan</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/commands/apply">developer.hashicorp.com/terraform/cli/commands/apply</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/state/refresh">developer.hashicorp.com/terraform/tutorials/state/refresh</a></p><p><strong>Q29.</strong></p><p><strong>The fictional company Orelivex Systems has one Terraform configuration and wants to add one cloud block for HCP Terraform and a second cloud block for a separate Terraform Enterprise instance in that same configuration. Terraform supports multiple cloud blocks in one configuration for this purpose.</strong></p><p><strong>A. True</strong></p><p><strong>B. False</strong></p><p>A is incorrect because Terraform permits only one cloud block per configuration. A cloud block can configure the Terraform CLI integration with either HCP Terraform or a Terraform Enterprise installation, but the same configuration cannot contain multiple cloud blocks to connect simultaneously to both services.</p><blockquote><strong><em>B is correct because the terraform block reference explicitly limits a configuration to one cloud block. The cloud block represents the configuration’s integration with one HCP Terraform or Terraform Enterprise destination. Terraform also prevents combining a cloud block with a backend block in the same configuration for state storage.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/language/block/terraform">developer.hashicorp.com/terraform/language/block/terraform</a></p><p><a href="http://developer.hashicorp.com/terraform/cli/cloud">developer.hashicorp.com/terraform/cli/cloud</a></p><p><a href="http://developer.hashicorp.com/terraform/cloud-docs">developer.hashicorp.com/terraform/cloud-docs</a></p><p><strong>Q30.</strong></p><p><strong>The fictional company Kalyvora Cloud wants its team to handle API tokens and other secrets securely when using Terraform and HCP Terraform. Which of the following are appropriate mechanisms for keeping those secrets out of version-controlled plaintext configuration? (Choose three.)</strong></p><p><strong>A. In a terraform.tfvars file, checked into your version control system.</strong></p><p><strong>B. In an environment variable and referenced with TF_VAR_variablename.</strong></p><p><strong>C. In HashiCorp Vault.</strong></p><p><strong>D. In a plaintext document on a shared drive.</strong></p><p><strong>E. In an HCP Terraform variable, with the sensitive option checked.</strong></p><p>A is incorrect because a terraform.tfvars file can contain sensitive values, and committing such a file to version control can expose those secrets to everyone who can read the repository and can preserve them indefinitely in repository history. Secret values should be supplied through secured runtime mechanisms rather than committed as plaintext.</p><blockquote><strong><em>B is correct because Terraform can read root-module input values from environment variables whose names use the TF_VAR_&lt;NAME&gt; convention, allowing the value to remain outside the Terraform configuration files. In HCP Terraform, environment variables can be stored as workspace or variable-set values and marked sensitive. For provider authentication specifically, HashiCorp recommends provider-supported environment variables or dynamic credentials when possible because they can reduce the chance of credentials flowing into Terraform state or logs.</em></strong></blockquote><blockquote><strong><em>C is correct because HashiCorp Vault is specifically designed to store, control access to, and dynamically generate secrets. HCP Terraform can also integrate with Vault-backed dynamic credential workflows so Terraform runs receive short-lived credentials instead of relying on long-lived static secrets stored in configuration.</em></strong></blockquote><p>D is incorrect because a plaintext document on a shared drive provides poor secret isolation and can expose API tokens to unintended users, backups, search indexes, synchronization clients, or other systems. A dedicated secrets system or sensitive workspace variable provides stronger access controls and secret-handling semantics.</p><blockquote><strong><em>E is correct because HCP Terraform lets users mark Terraform or environment variables as sensitive. Once a workspace variable is marked sensitive, HCP Terraform treats its value as write-only in the UI and API so users cannot later retrieve the plaintext value through those interfaces. Sensitive variables still require careful downstream handling because Terraform operations can use them and certain configurations can cause secret values to appear in logs or state.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://developer.hashicorp.com/terraform/cloud-docs/variables/managing-variables">developer.hashicorp.com/terraform/cloud-docs/variables/managing-variables</a></p><p><a href="http://developer.hashicorp.com/terraform/cloud-docs/architectural-details/security-model">developer.hashicorp.com/terraform/cloud-docs/architectural-details/security-model</a></p><p><a href="http://developer.hashicorp.com/terraform/tutorials/cloud/dynamic-credentials">developer.hashicorp.com/terraform/tutorials/cloud/dynamic-credentials</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97f19d6f6d1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kubernetes and Cloud Native Associate (KCNA) Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/kubernetes-and-cloud-native-associate-kcna-practice-questions-e79560e06157?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e79560e06157</guid>
            <category><![CDATA[kubernetes-certification]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 11:29:30 GMT</pubDate>
            <atom:updated>2026-10-06T11:29:30.164Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I97-I8iOodn4d91r_77GOg.jpeg" /><figcaption>KloudExams Kubernetes and Cloud Native Associate (KCNA) Practice Questions</figcaption></figure><p>Preparing for the <strong>Kubernetes and Cloud Native Associate (KCNA) certification</strong>? Knowing the terminology of containers, Kubernetes, and cloud native technologies is a starting point — understanding how these concepts fit together is what builds a strong foundation for practical learning.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise applying Kubernetes and cloud native concepts to different situations. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Kubernetes documentation, official cloud native learning resources, and hands-on exploration.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>At the fictional company Velunexa-41 Systems, which Kubernetes API object is the recommended way to run a scalable, stateless application on a cluster?</strong></p><p><strong>A. ReplicaSet</strong></p><p><strong>B. Deployment</strong></p><p><strong>C. DaemonSet</strong></p><p><strong>D. Pod</strong></p><p>A is incorrect because a ReplicaSet can maintain a specified number of identical Pod replicas, but Kubernetes documentation recommends using a Deployment rather than managing ReplicaSets directly for ordinary stateless application workloads. A Deployment provides a higher-level declarative interface and manages ReplicaSets on the user’s behalf. This gives the application lifecycle capabilities that are normally required in production, including controlled rollouts, rolling updates, rollback history, scaling, and replacement of old ReplicaSets as the Pod template changes. A ReplicaSet is therefore an important internal building block of a Deployment, but it is usually not the preferred object that an application operator creates and manages directly.</p><blockquote><strong><em>B is correct because a Deployment is the standard Kubernetes workload resource for managing a scalable application whose Pods are interchangeable and do not need stable per-replica identity. A Deployment declares the desired Pod template and replica count and then manages ReplicaSets to reconcile the actual cluster state with that desired state. It supports horizontal scaling, rolling updates, rollback, rollout status, and controlled replacement of replicas. Kubernetes documentation specifically describes Deployment as a good fit for stateless application workloads and identifies it as the most common way to run an application on a cluster. For these reasons, Deployment is the recommended answer rather than directly creating Pods or ReplicaSets.</em></strong></blockquote><p>C is incorrect because a DaemonSet is designed for a different placement model. Its purpose is to ensure that a copy of a Pod runs on all, or a selected subset of, eligible Nodes. Typical DaemonSet workloads include node-level log collectors, monitoring agents, storage daemons, or networking components. The replica count is therefore primarily tied to matching Nodes rather than to an application’s independent scaling requirement. A general stateless web or service workload that needs an arbitrary number of replicas is normally modeled with a Deployment, not a DaemonSet.</p><p>D is incorrect because an individual Pod is the basic execution unit in Kubernetes, but directly managing standalone Pods does not provide the controller behavior expected for a scalable application. A standalone Pod does not by itself maintain a desired replica count, perform rolling updates, create replacement replicas after deletion through a higher-level workload controller, or provide rollout and rollback semantics. Kubernetes normally expects users to manage application Pods through workload resources such as Deployments. The Deployment controller and its ReplicaSets continuously reconcile the desired state and create replacement Pods when necessary.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/deployment/">kubernetes.io/docs/concepts/workloads/controllers/deployment/</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/">kubernetes.io/docs/concepts/workloads/</a></p><p><a href="http://kubernetes.io/docs/tasks/run-application/run-stateless-application-deployment/">kubernetes.io/docs/tasks/run-application/run-stateless-application-deployment/</a></p><p><strong>Q2.</strong></p><p><strong>At the fictional company Torvane-Pulse Data, an organization must process a large burst of data every Monday on a cloud-based Kubernetes cluster. About 1,000 one-hour compute jobs must finish by Monday night, but that extra capacity is not needed for the rest of the week. Which option is the most cost-effective approach among the choices?</strong></p><p><strong>A. Run a group of Nodes with the exact required size to complete the batch on time, and use a combination of taints, tolerations, and nodeSelector settings to reserve these Nodes for the batch jobs.</strong></p><p><strong>B. Leverage Kubernetes Cluster Autoscaler to automatically add and remove Nodes as they are needed.</strong></p><p><strong>C. Commit to a specific level of spending to get discounted prices, for example through reserved-capacity mechanisms.</strong></p><p><strong>D. Use PriorityClass so that the weekly batch job gets priority over other workloads running on the cluster and can be completed on time.</strong></p><p>A is incorrect because taints, tolerations, and nodeSelector settings are scheduling controls; they do not by themselves create or remove cloud compute capacity. Keeping a permanently sized group of Nodes large enough for the Monday burst would make the capacity available when needed, but those Nodes could remain idle or underutilized during the rest of the week unless some separate mechanism scales the underlying infrastructure down. Reserving Nodes exclusively for the batch also reduces opportunities for other workloads to use that capacity. This can be operationally predictable, but for a highly periodic burst followed by long idle periods it generally sacrifices the principal cloud cost advantage of paying for elastic capacity only while it is required.</p><blockquote><strong><em>B is correct because Cluster Autoscaler is a Kubernetes SIG Autoscaling node-autoscaler implementation that can add Nodes to configured node groups when Pods cannot be scheduled because the existing cluster lacks suitable capacity, and can remove Nodes that are no longer needed when workloads subside and removal is safe. That behavior matches a large weekly burst: the cluster can expand to provide capacity for pending batch Pods and later consolidate back toward its normal size. The result reduces the need to pay continuously for a permanently oversized worker fleet. Actual savings depend on cloud-provider pricing, workload resource requests, autoscaler configuration, node-group design, quotas, startup times, and scheduling constraints, but among the listed choices this is the option directly designed to align node capacity with changing demand.</em></strong></blockquote><p>C is incorrect because reserved-capacity or committed-spend discounts primarily lower the unit price in exchange for a usage or spending commitment; they do not inherently eliminate the cost of capacity that is unnecessary most of the week. Commitments can be economical for a predictable baseline that runs continuously, and an organization might combine commitments for baseline capacity with elastic capacity for bursts. However, the scenario specifically emphasizes a short weekly spike followed by much lower demand. An autoscaling approach better matches resource supply to that workload pattern instead of solving the problem by committing to a fixed spending level.</p><p>D is incorrect because PriorityClass influences Pod scheduling priority and preemption; it does not provision new Nodes or increase the amount of compute capacity available to the cluster. Giving batch Pods high priority could cause lower-priority workloads to be preempted so the batch can use existing resources, which may harm other services and still might not provide enough capacity to complete all jobs by the deadline. Priority is useful when deciding which workloads should receive scarce resources first, but node autoscaling is the mechanism that addresses insufficient cluster capacity by adding and later removing Nodes.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/cluster-administration/node-autoscaling/">kubernetes.io/docs/concepts/cluster-administration/node-autoscaling/</a></p><p><a href="http://kubernetes.io/docs/concepts/scheduling-eviction/taint-and-toleration/">kubernetes.io/docs/concepts/scheduling-eviction/taint-and-toleration/</a></p><p><a href="http://kubernetes.io/docs/concepts/scheduling-eviction/pod-priority-preemption/">kubernetes.io/docs/concepts/scheduling-eviction/pod-priority-preemption/</a></p><p><strong>Q3.</strong></p><p><strong>At the fictional company Luminexor-Mesh Systems, which current Kubernetes SIG Network approach standardizes service-mesh traffic configuration through Gateway API and its GAMMA work?</strong></p><p><strong>A. Service Mesh Specification (SMS)</strong></p><p><strong>B. Service Mesh Technology (SMT)</strong></p><p><strong>C. Gateway API for Mesh Management and Administration (GAMMA)</strong></p><p><strong>D. Service Mesh Function (SMF)</strong></p><p>A is incorrect because Service Mesh Specification, abbreviated SMS, is not the current Kubernetes SIG Network standard named in the question. Kubernetes service-mesh standardization work is being developed through Gateway API, whose APIs can be used not only for north-south gateway traffic but also for mesh or east-west traffic. Inventing a similarly named specification does not correspond to the actual Kubernetes networking project.</p><p>B is incorrect because Service Mesh Technology, abbreviated SMT, is not a Kubernetes API standard or SIG Network initiative. Service mesh is a technology category implemented by multiple projects, but interoperability and Kubernetes-native configuration require defined APIs rather than a generic label. The current standards work relevant to Kubernetes service meshes is part of Gateway API’s mesh support.</p><blockquote><strong><em>C is correct because Gateway API for Mesh Management and Administration (GAMMA) is the Kubernetes Gateway API initiative that standardizes how Gateway API resources apply to service-mesh use cases, including east-west traffic between Services. Mesh support entered the Gateway API Standard Channel and is considered generally available for the supported functionality. A key model is that route resources can attach directly to Kubernetes Services for mesh traffic instead of requiring a Gateway resource for every in-mesh route.</em></strong></blockquote><p>D is incorrect because Service Mesh Function, abbreviated SMF, is not the name of the Kubernetes SIG Network service-mesh API initiative. Individual service meshes certainly implement networking functions such as routing, traffic splitting, retries, policy, telemetry, and mutual authentication, but that does not create a standard called Service Mesh Function. Gateway API for Service Mesh through GAMMA is the current named Kubernetes community effort represented by the choices.</p><p><strong>Reference links:</strong></p><p><a href="http://gateway-api.sigs.k8s.io/docs/mesh/mesh-overview/">gateway-api.sigs.k8s.io/docs/mesh/mesh-overview/</a></p><p><a href="http://gateway-api.sigs.k8s.io/docs/mesh/gamma/">gateway-api.sigs.k8s.io/docs/mesh/gamma/</a></p><p><a href="http://gateway-api.sigs.k8s.io/">gateway-api.sigs.k8s.io/</a></p><p><strong>Q4.</strong></p><p><strong>At the fictional company Quenvera Release Engineering, which kubectl command can watch the progress of the latest rollout for supported workload resources such as Deployments, DaemonSets, and StatefulSets?</strong></p><p><strong>A. kubectl rollout watch</strong></p><p><strong>B. kubectl rollout progress</strong></p><p><strong>C. kubectl rollout state</strong></p><p><strong>D. kubectl rollout status</strong></p><p>A is incorrect because kubectl rollout watch is not a valid kubectl rollout subcommand. kubectl has a general watch capability through commands such as kubectl get with — watch, but the rollout command family uses a specific subcommand for following rollout completion. Inventing a plausible word such as watch does not make it part of the kubectl rollout interface. For workload rollout progress, the supported command is kubectl rollout status.</p><p>B is incorrect because kubectl rollout progress is not a valid kubectl rollout subcommand. The concept of rollout progress exists in Kubernetes, and Deployments can have progress-related conditions and deadlines, but the command-line subcommand used to observe a rollout is status. kubectl rollout also provides other operations such as history, pause, resume, restart, and undo, depending on the workload and command, but progress is not the command requested here.</p><p>C is incorrect because kubectl rollout state is not a valid rollout subcommand. Kubernetes objects have status fields and conditions representing observed state, but kubectl uses the word status for the rollout-monitoring command. Certification questions often test recognition of exact kubectl command syntax, so an intuitively reasonable but nonexistent subcommand such as state should not be confused with kubectl rollout status.</p><blockquote><strong><em>D is correct because kubectl rollout status shows the status of a rollout and, by default, watches the latest rollout until it completes. It can be used with supported workload resources including Deployments, DaemonSets, and StatefulSets. For example, kubectl rollout status deployment/my-app follows the Deployment rollout. The command also provides options such as — watch=false when the user wants the current status without continuing to watch, and — timeout to limit how long kubectl waits. Therefore status is the correct rollout subcommand for monitoring rollout progress.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_rollout/kubectl_rollout_status/">kubernetes.io/docs/reference/kubectl/generated/kubectl_rollout/kubectl_rollout_status/</a></p><p><a href="http://kubernetes.io/docs/reference/generated/kubectl/kubectl-commands/">kubernetes.io/docs/reference/generated/kubectl/kubectl-commands/</a></p><p><strong>Q5.</strong></p><p><strong>At the fictional company Norvexa Service Fabric, which type of Kubernetes Service normally requires you to create and manage the corresponding EndpointSlice objects manually when you want to define its network backends yourself?</strong></p><p><strong>A. LoadBalancer</strong></p><p><strong>B. Services without selectors</strong></p><p><strong>C. NodePort</strong></p><p><strong>D. ClusterIP with selectors</strong></p><p>A is incorrect because a Service of type LoadBalancer can use a selector to identify Pods in the same way as other selector-based Services. When a Service has a selector, the Kubernetes control plane automatically manages EndpointSlice objects that represent the matching backend Pods. The LoadBalancer type primarily changes how the Service is exposed, typically by integrating with an external load balancer provided by the environment. It does not inherently require administrators to create EndpointSlices manually. A selectorless Service is the relevant case when the administrator explicitly supplies backend network endpoints.</p><blockquote><strong><em>B is correct because Kubernetes does not automatically create EndpointSlice objects for a Service that has no selector. This design allows a selectorless Service to represent backends that Kubernetes is not selecting as Pods, including services running outside the cluster, a backend in another environment, or manually managed addresses during a migration. The administrator or another controller can create EndpointSlice objects directly and associate them with the Service using the kubernetes.io/service-name label. EndpointSlice is the current Kubernetes API for representing Service backends and should be used for this purpose.</em></strong></blockquote><p>C is incorrect because NodePort is a Service type that exposes the Service on a port of each Node, but a NodePort Service can still use a normal selector. When it has a selector, the control plane automatically creates and updates EndpointSlice objects for matching Pods. The NodePort setting affects how clients reach the Service from outside or through Node addresses; it does not determine whether backend discovery is automatic. The selector field is what determines whether Kubernetes can automatically derive the Service’s Pod backends.</p><p>D is incorrect because a ClusterIP Service with a selector is the standard case in which Kubernetes automatically discovers matching Pods and creates or updates EndpointSlice objects on the Service’s behalf. As Pods are added, removed, become ready, or change, the control plane updates the relevant EndpointSlices. Manual EndpointSlice management is therefore unnecessary for an ordinary selector-based ClusterIP Service. The manual-management case described in the question applies to Services without selectors.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/service/">kubernetes.io/docs/concepts/services-networking/service/</a></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/endpoint-slices/">kubernetes.io/docs/concepts/services-networking/endpoint-slices/</a></p><p><a href="http://kubernetes.io/docs/reference/kubernetes-api/discovery/endpoint-slice-v1/">kubernetes.io/docs/reference/kubernetes-api/discovery/endpoint-slice-v1/</a></p><p><strong>Q6.</strong></p><p><strong>At the fictional company Qubrixa Release Systems, what is the default update strategy for a Kubernetes Deployment when spec.strategy.type is not explicitly set?</strong></p><p><strong>A. RollingUpdate</strong></p><p><strong>B. Blue/Green deployment</strong></p><p><strong>C. Canary deployment</strong></p><p><strong>D. Recreate</strong></p><blockquote><strong><em>A is correct because RollingUpdate is the default value of a Deployment’s spec.strategy.type. With a RollingUpdate, the Deployment controller gradually scales up the new ReplicaSet while scaling down the old ReplicaSet instead of terminating every old Pod at once. The behavior can be tuned with maxSurge and maxUnavailable, which both default to 25 percent when applicable. This strategy is intended to allow controlled application updates while maintaining a configurable level of availability during the transition.</em></strong></blockquote><p>B is incorrect because blue/green deployment is a release strategy in which separate old and new environments are maintained and traffic is switched between them, but it is not a built-in Deployment strategy type selected by default. Kubernetes primitives such as Deployments, Services, Gateway API resources, and delivery tooling can be combined to implement blue/green workflows, but the Deployment API’s native strategy.type values are RollingUpdate and Recreate. Therefore blue/green is an application-delivery pattern rather than the default Deployment strategy field.</p><p>C is incorrect because canary deployment is a progressive-delivery technique that exposes a limited portion of users or traffic to a new version before broader rollout. Kubernetes can support canary releases through multiple Deployments, routing controls, service-mesh or Gateway capabilities, and progressive-delivery tools, but canary is not the default value of Deployment spec.strategy.type. A standard Deployment defaults to RollingUpdate.</p><p>D is incorrect because Recreate is a valid Deployment strategy type, but it is not the default. When Recreate is selected for a Deployment update, the old Pods are terminated before the new revision’s Pods are created as part of that Deployment upgrade behavior. This can introduce downtime and is useful only when the workload requires old and new versions not to run concurrently. If no strategy type is specified, Kubernetes defaults the Deployment to RollingUpdate instead.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/deployment/">kubernetes.io/docs/concepts/workloads/controllers/deployment/</a></p><p><a href="http://kubernetes.io/docs/reference/kubernetes-api/apps/deployment-v1/">kubernetes.io/docs/reference/kubernetes-api/apps/deployment-v1/</a></p><p><strong>Q7.</strong></p><p><strong>At the fictional company Orivex Application Systems, which Kubernetes component is the smallest deployable unit of computing?</strong></p><p><strong>A. StatefulSet</strong></p><p><strong>B. Deployment</strong></p><p><strong>C. Pod</strong></p><p><strong>D. Container</strong></p><p>A is incorrect because a StatefulSet is a higher-level Kubernetes workload resource that creates and manages Pods requiring stable identity, ordered lifecycle behavior, or stable storage associations. The scheduler does not place a StatefulSet itself onto a Node; the StatefulSet controller creates Pod objects, and those Pods become the individual scheduling units. StatefulSet is therefore a workload-management abstraction above the smallest deployable unit.</p><p>B is incorrect because a Deployment is a controller-level workload resource used to manage ReplicaSets and Pods, commonly for stateless applications. A Deployment specifies desired application state, replica count, update strategy, and Pod template, but the scheduler does not deploy the Deployment object directly onto Nodes. The Deployment ultimately causes Pods to be created, and those Pods are the units that Kubernetes schedules and runs.</p><blockquote><strong><em>C is correct because Kubernetes documentation defines a Pod as the smallest deployable unit of computing that Kubernetes can create and manage. A Pod contains one or more containers that are co-located and co-scheduled and share a network context along with any configured shared storage. Kubernetes scheduling operates on Pods, not individual application containers. Higher-level workload resources such as Deployments, StatefulSets, DaemonSets, and Jobs manage Pods according to different lifecycle requirements.</em></strong></blockquote><p>D is incorrect because a container is an execution and packaging unit, but Kubernetes does not normally schedule a standalone container as an independent Kubernetes object. Containers are defined inside Pods, and all containers in a Pod are scheduled together onto the same Node. Even when a Pod contains only one application container, the Kubernetes deployable and schedulable unit remains the Pod.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/workloads/pods/">kubernetes.io/docs/concepts/workloads/pods/</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/">kubernetes.io/docs/concepts/workloads/</a></p><p><strong>Q8.</strong></p><p><strong>At the fictional company Lumetra Deployment Services, which methods can be used to manually change the replica count of a Kubernetes Deployment using kubectl?</strong></p><p><strong>A. With kubectl edit deployment exclusively.</strong></p><p><strong>B. With kubectl scale-up deployment exclusively.</strong></p><p><strong>C. With kubectl scale deployment and kubectl edit deployment.</strong></p><p><strong>D. With kubectl scale deployment exclusively.</strong></p><p>A is incorrect because kubectl edit can be used to change a Deployment’s spec.replicas field, but it is not the only available kubectl method. kubectl scale provides a dedicated command for changing the replica count, and kubectl patch is another supported way to modify spec.replicas. Therefore saying that editing is the exclusive scaling method makes the option incorrect even though kubectl edit itself can accomplish the change.</p><p>B is incorrect because kubectl scale-up is not the standard kubectl command for changing a Deployment’s replica count. Kubernetes uses kubectl scale with the — replicas option. For example, kubectl scale deployment/my-app — replicas=5 sets a new desired replica count for the Deployment. The invented scale-up form should not be confused with the valid scale subcommand.</p><blockquote><strong><em>C is correct because both kubectl scale deployment and kubectl edit deployment can change the desired number of Deployment replicas. kubectl scale provides a purpose-built interface such as kubectl scale deployment/my-app — replicas=5, while kubectl edit opens the live resource in an editor and lets the user change spec.replicas directly. Current Kubernetes documentation also describes kubectl patch as another valid way to modify the replica count, but among the choices provided, option C correctly identifies two legitimate methods.</em></strong></blockquote><p>D is incorrect because kubectl scale is a valid and convenient method, but it is not the only manual method available. A user can edit spec.replicas with kubectl edit or update the field with kubectl patch, and declarative configuration workflows can also change the desired replica count. The word exclusively therefore makes this option incorrect.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_scale/">kubernetes.io/docs/reference/kubectl/generated/kubectl_scale/</a></p><p><a href="http://kubernetes.io/docs/tasks/run-application/scale-deployment/">kubernetes.io/docs/tasks/run-application/scale-deployment/</a></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_edit/">kubernetes.io/docs/reference/kubectl/generated/kubectl_edit/</a></p><p><strong>Q9.</strong></p><p><strong>At the fictional company Velatrix Workload Systems, what do Kubernetes Deployments and StatefulSets have in common?</strong></p><p><strong>A. They manage Pods based on a shared Pod template.</strong></p><p><strong>B. They support the OnDelete update strategy.</strong></p><p><strong>C. They both guarantee ordered, graceful deployment and scaling.</strong></p><p><strong>D. They both maintain a sticky identity for each of their Pods.</strong></p><blockquote><strong><em>A is correct because both Deployments and StatefulSets are higher-level workload resources that use a Pod template to describe the Pods they manage. In both APIs, spec.template contains a PodTemplateSpec that defines items such as containers, images, labels, volumes, security settings, and other Pod configuration. Controllers create Pods from that template as needed to reconcile desired state. The resulting Pods are not identical in every respect because StatefulSet assigns each Pod a distinct ordinal identity, but the workload specification from which their application configuration is derived is shared through the Pod template.</em></strong></blockquote><p>B is incorrect because OnDelete is a StatefulSet update strategy, not a Deployment update strategy. A StatefulSet can use OnDelete so that changes to its Pod template are not automatically rolled out to existing Pods; the user must delete those Pods for replacements to be created from the updated template. Deployments instead use RollingUpdate or Recreate as their supported strategy types. Therefore OnDelete is not a capability shared by both resource types.</p><p>C is incorrect because ordered and graceful deployment and scaling are characteristic StatefulSet guarantees rather than general Deployment guarantees. StatefulSet Pods have ordinal identities and, under its ordered Pod-management behavior, can be created, updated, or terminated according to defined ordering and readiness rules. Deployment replicas are intended to be interchangeable, and a Deployment does not give each replica an ordinal identity or the same ordered scaling semantics. Both controllers manage replica lifecycles, but only StatefulSet is specifically designed around ordered stateful identity.</p><p>D is incorrect because maintaining a stable or sticky identity for each managed Pod is a defining StatefulSet property, not a Deployment property. A StatefulSet Pod keeps a predictable ordinal name and can maintain stable storage and network identity across replacement. Deployment Pods are managed through ReplicaSets and are generally treated as interchangeable replicas; replacement Pods receive new identities. Therefore sticky per-Pod identity is not common to both workload types.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/deployment/">kubernetes.io/docs/concepts/workloads/controllers/deployment/</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/statefulset/">kubernetes.io/docs/concepts/workloads/controllers/statefulset/</a></p><p><a href="http://kubernetes.io/docs/reference/kubernetes-api/apps/stateful-set-v1/">kubernetes.io/docs/reference/kubernetes-api/apps/stateful-set-v1/</a></p><p><strong>Q10.</strong></p><p><strong>At the fictional company Quoralith Metrics Systems, which of the following is a valid PromQL expression that selects the http_requests_total time series whose job label equals apiserver?</strong></p><p><strong>A. SELECT * from http_requests_total WHERE job=apiserver</strong></p><p><strong>B. http_requests_total WHERE (job=”apiserver”)</strong></p><p><strong>C. SELECT * from http_requests_total</strong></p><p><strong>D. http_requests_total{job=”apiserver”}</strong></p><p>A is incorrect because PromQL is not SQL and does not use SELECT, FROM, and WHERE syntax to select time series. Prometheus has its own functional query language in which a metric name acts as a vector selector and label matchers are placed inside curly braces. SQL-like syntax may look familiar to database users, but it is not valid PromQL. To filter http_requests_total by the job label, the label matcher must be written using PromQL’s selector syntax.</p><p>B is incorrect because PromQL does not use a WHERE clause after a metric name. Label filtering is expressed directly as part of the vector selector using curly braces. Parentheses can appear in PromQL for function calls and grouping expressions, but the construction WHERE (job=”apiserver”) is not valid PromQL grammar. The correct expression places job=”apiserver” inside braces following the metric name.</p><p>C is incorrect because SELECT * from http_requests_total is SQL-style syntax rather than PromQL. If the goal were to select all series for this metric without filtering by labels, the valid PromQL expression could simply be http_requests_total. The question additionally requires a job label value of apiserver, so a label matcher is needed even after correcting the SQL syntax.</p><blockquote><strong><em>D is correct because PromQL instant vector selectors use a metric name followed by optional label matchers enclosed in curly braces. The expression http_requests_total{job=”apiserver”} selects time series named http_requests_total whose job label is exactly apiserver. PromQL also supports other matcher operators such as !=, =~, and !~, but equality with = is the correct operator for the exact label value requested here. This option corrects the malformed selector syntax present in older versions of the question bank.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://prometheus.io/docs/prometheus/latest/querying/basics/">prometheus.io/docs/prometheus/latest/querying/basics/</a></p><p><a href="http://prometheus.io/docs/prometheus/latest/querying/operators/">prometheus.io/docs/prometheus/latest/querying/operators/</a></p><p><strong>Q11.</strong></p><p><strong>At the fictional company Ceryntra Runtime Platforms, which Node component is responsible for actually executing and managing the containers that make up Kubernetes workloads after kubelet requests runtime operations?</strong></p><p><strong>A. The kubelet.</strong></p><p><strong>B. The kube-proxy.</strong></p><p><strong>C. The kube-apiserver.</strong></p><p><strong>D. The container runtime.</strong></p><p>A is incorrect because kubelet is the Node agent that ensures the containers described by assigned Pod specifications are running, but kubelet does not itself implement the low-level container execution engine. Instead, kubelet acts as a client of a CRI-compatible container runtime. It sends runtime operations over the Container Runtime Interface, such as requests to create Pod sandboxes, pull images, start or stop containers, and retrieve runtime status. The distinction is that kubelet reconciles the desired Pod state, while the container runtime performs the actual container lifecycle operations.</p><p>B is incorrect because kube-proxy is concerned with Kubernetes Service networking where kube-proxy is used. It watches Service and EndpointSlice information and configures operating-system networking behavior so traffic sent to a Service can reach appropriate backends. It does not pull container images, create container processes, or provide the CRI runtime service used by kubelet. Service traffic handling and container execution are separate Node responsibilities.</p><p>C is incorrect because kube-apiserver is a control-plane component rather than the Node-level runtime that executes workload containers. The API server exposes the Kubernetes API and coordinates cluster state, but it does not directly launch application processes on worker Nodes. kubelet observes desired Pod state through Kubernetes APIs and then asks the local container runtime to carry out the required runtime operations.</p><blockquote><strong><em>D is correct because every Kubernetes Node needs a working container runtime so kubelet can launch Pods and their containers. The runtime exposes the CRI runtime and image services used by kubelet and performs container lifecycle operations such as image handling, sandbox creation, container creation, startup, shutdown, and status inspection. Common current Kubernetes runtimes include containerd and CRI-O. The kubelet remains responsible for reconciliation, but the container runtime is the Node component that actually executes and manages the containers.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/containers/cri/">kubernetes.io/docs/concepts/containers/cri/</a></p><p><a href="http://kubernetes.io/docs/setup/production-environment/container-runtimes/">kubernetes.io/docs/setup/production-environment/container-runtimes/</a></p><p><a href="http://kubernetes.io/docs/concepts/architecture/">kubernetes.io/docs/concepts/architecture/</a></p><p><strong>Q12.</strong></p><p><strong>At the fictional company Lumora Network Engineering, Kubernetes IPv4/IPv6 dual-stack networking provides which capability?</strong></p><p><strong>A. Translates an IPv4 request from a Service to an IPv6 Service.</strong></p><p><strong>B. Allows you to access an IPv4 address by using an IPv6 address.</strong></p><p><strong>C. Requires NetworkPolicies to prevent Services from mixing requests.</strong></p><p><strong>D. Allows you to create IPv4 and IPv6 dual-stack Services.</strong></p><p>A is incorrect because Kubernetes dual-stack support does not mean that Kubernetes automatically translates arbitrary IPv4 requests into IPv6 requests between Services. Address-family translation is a separate networking concern and may require specialized gateway or translation technology. Kubernetes dual-stack instead lets Pods, Nodes, and Services operate with both address families when the cluster network and relevant infrastructure are configured to support them.</p><p>B is incorrect because an IPv6 address is not simply an alternate notation through which a client accesses an unrelated IPv4 address. IPv4 and IPv6 are distinct address families. Dual-stack systems support both families simultaneously, and applications or networking components select an address that matches a usable family. Translation mechanisms can bridge address families in some architectures, but that is not the defining Kubernetes dual-stack behavior.</p><p>C is incorrect because NetworkPolicy is not required merely to stop IPv4 and IPv6 Service traffic from being “mixed.” NetworkPolicy controls allowed ingress and egress connections for selected Pods when supported by the network implementation. Dual-stack configuration is expressed through cluster networking and Service fields such as ipFamilyPolicy and ipFamilies. Security policy remains important, but NetworkPolicy is not a prerequisite whose purpose is to separate the two IP families.</p><blockquote><strong><em>D is correct because Kubernetes dual-stack support allows Services to have both IPv4 and IPv6 cluster IP addresses when the cluster is configured for both address families. A Service can use ipFamilyPolicy values such as PreferDualStack or RequireDualStack, and its ipFamilies field can identify the requested address-family order. Kubernetes can then allocate both IPv4 and IPv6 addresses for the Service. Dual-stack capability also extends to Pod and Node networking, subject to support from the installed network implementation and infrastructure.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/dual-stack/">kubernetes.io/docs/concepts/services-networking/dual-stack/</a></p><p><a href="http://kubernetes.io/docs/tasks/network/validate-dual-stack/">kubernetes.io/docs/tasks/network/validate-dual-stack/</a></p><p><strong>Q13.</strong></p><p><strong>At the fictional company Cerynova Event Systems, what is CloudEvents?</strong></p><p><strong>A. It is a specification for describing event data in common formats specifically for Kubernetes network traffic management and cloud providers.</strong></p><p><strong>B. It is a specification for describing event data in common formats only across major public cloud providers.</strong></p><p><strong>C. It is a specification for describing event data in common formats to provide interoperability across services, platforms, and systems.</strong></p><p><strong>D. It is a Kubernetes specification for describing event data specifically for Apple cloud and device platforms.</strong></p><p>A is incorrect because CloudEvents is not a Kubernetes networking specification. Kubernetes-based systems can produce and consume CloudEvents, and event-driven Kubernetes projects may adopt the standard, but CloudEvents is deliberately independent of Kubernetes and of any particular network-traffic implementation. It standardizes event metadata and representations so events can move across diverse producers, consumers, middleware, protocols, and platforms.</p><p>B is incorrect because CloudEvents is not limited to major public cloud providers. One of its central design goals is vendor neutrality and interoperability across many types of services and systems, including cloud services, SaaS products, event brokers, serverless platforms, Kubernetes-based systems, and custom applications. Major cloud providers participate in and support CloudEvents, but restricting the specification to those providers would undermine the portability it was created to enable.</p><blockquote><strong><em>C is correct because CloudEvents is a vendor-neutral specification for describing event data in common formats so that events can interoperate across services, platforms, and systems. Event producers historically used different event envelopes and metadata conventions, forcing consumers and middleware to handle many incompatible formats. CloudEvents standardizes common context attributes and defines event formats and protocol bindings. The project is a graduated CNCF project and is used across a broad event-driven ecosystem.</em></strong></blockquote><p>D is incorrect because CloudEvents is not specific to Kubernetes, iCloud, iOS, macOS, or any single vendor’s products. Its purpose is exactly the opposite: to establish portable event conventions that different platforms and ecosystems can understand. The standard can be implemented by products from many vendors and by open-source projects without being tied to one operating system or cloud provider.</p><p><strong>Reference links:</strong></p><p><a href="http://cloudevents.io/">cloudevents.io/</a></p><p><a href="http://github.com/cloudevents/spec/blob/main/cloudevents/spec.md">github.com/cloudevents/spec/blob/main/cloudevents/spec.md</a></p><p><a href="http://github.com/cloudevents/spec/blob/main/cloudevents/primer.md">github.com/cloudevents/spec/blob/main/cloudevents/primer.md</a></p><p><strong>Q14.</strong></p><p><strong>At the fictional company Solvaryn GitOps Engineering, which option best describes direct benefits of adopting a GitOps deployment approach based on declarative desired state and continuous reconciliation?</strong></p><p><strong>A. Reduce failed deployments, operational costs, and fragile release processes.</strong></p><p><strong>B. Reduce failed deployments, configuration drift, and fragile release processes.</strong></p><p><strong>C. Reduce failed deployments, operational costs, and learn Git.</strong></p><p><strong>D. Reduce failed deployments, configuration drift, and improve your reputation.</strong></p><p>A is incorrect because GitOps can indirectly reduce operational effort and costs by standardizing automation and reducing manual work, but operational cost is not as direct or universal a consequence as configuration-drift control. GitOps continuously compares actual system state with a versioned declarative desired state and attempts to reconcile differences. That mechanism directly addresses drift and fragile manual release processes. Cost improvement can occur, but it depends on organization size, implementation quality, tooling choices, and operating model, so it is less fundamental to the GitOps definition than the benefits listed in option B.</p><blockquote><strong><em>B is correct because GitOps uses declarative desired state that is versioned and immutable, automatically pulled by software agents, and continuously reconciled against actual system state. This design can reduce configuration drift because unauthorized or accidental changes can be detected and corrected, and it makes releases more repeatable because deployment state is represented in version-controlled configuration rather than undocumented manual procedures. Version history and automated reconciliation also improve rollback and recovery, which can reduce failed or prolonged deployments. These benefits follow directly from the core OpenGitOps principles.</em></strong></blockquote><p>C is incorrect because learning Git may be necessary for teams adopting a Git-centered workflow, but it is not an operational benefit comparable to drift control, deployment reliability, or reproducibility. Operational cost can sometimes decrease, yet that result is organization-specific rather than inherent. The defining GitOps value is that desired state is declarative, versioned, automatically pulled, and continuously reconciled, so the option that explicitly includes configuration-drift reduction is more technically aligned with the model.</p><p>D is incorrect because configuration-drift reduction and deployment reliability are genuine GitOps benefits, but improved reputation is not an engineering guarantee or defined GitOps outcome. A successful delivery practice may improve customer or internal confidence, but reputation depends on many factors outside the deployment model. The technically direct benefit in the choices is reducing fragile release processes through repeatable reconciliation, which option B states explicitly.</p><p><strong>Reference links:</strong></p><p><a href="https://docs.google.com/document/u/0/d/1ptHUszDI71hYxHjo-xGrwCxAOatPpx7zYZqt8aJgfcg/edit">opengitops.dev/</a></p><p><a href="https://docs.google.com/document/u/0/d/1ptHUszDI71hYxHjo-xGrwCxAOatPpx7zYZqt8aJgfcg/edit">opengitops.dev/about/</a></p><p><a href="https://docs.google.com/document/u/0/d/1ptHUszDI71hYxHjo-xGrwCxAOatPpx7zYZqt8aJgfcg/edit">opengitops.dev/blog/1.0-announcement/</a></p><p><strong>Q15.</strong></p><p><strong>At the fictional company Ceryntra Open Source Labs, which choice most closely matches the common conflict-resolution and decision-making approach used by CNCF-hosted open-source projects under open governance?</strong></p><p><strong>A. Financial Analysis</strong></p><p><strong>B. Discussion and Voting</strong></p><p><strong>C. Flipism Technique</strong></p><p><strong>D. Project Founder Say</strong></p><p>A is incorrect because financial analysis is not the primary governance mechanism by which technical open-source disagreements are resolved. Financial considerations can matter for budgets, infrastructure, events, or organizational planning, but open-source technical governance centers on transparent discussion, documented decision processes, maintainer responsibilities, consensus, and voting rules where needed. A technical proposal is therefore not normally accepted or rejected simply by performing a financial analysis.</p><blockquote><strong><em>B is correct because public discussion followed by consensus or, when required, formal voting best matches modern CNCF project governance among the available choices. CNCF governance guidance commonly recommends lazy consensus for normal decisions: participants discuss a proposal openly, and it proceeds when there are no blocking objections within the defined process. Formal votes are then used for matters such as unresolved disagreements, governance changes, elections, budgets, or other decisions the project’s governance document identifies as requiring a vote. Therefore “Discussion and Voting” is the closest option, although a mature project’s exact rules are defined by its own published governance model.</em></strong></blockquote><p>C is incorrect because flipism means making a decision by an arbitrary or random mechanism such as a coin flip and is not an accepted governance model for CNCF technical projects. Open-source governance aims for transparent, reviewable decision processes in which contributors can understand how and why a decision was made. Random choice would not provide the accountability, technical reasoning, contributor participation, or documented decision history expected of a healthy open project.</p><p>D is incorrect because CNCF-hosted projects are expected to use open, vendor-neutral governance rather than permanent unilateral control by a project founder. Founders may remain influential maintainers and their technical judgment can carry substantial weight, especially early in a project’s life, but sustainable governance defines roles, contribution paths, decision procedures, and leadership succession so the project does not depend on one individual’s decree. Larger projects commonly distribute authority across maintainers, councils, or elected bodies.</p><p><strong>Reference links:</strong></p><p><a href="http://contribute.cncf.io/projects/best-practices/governance/">contribute.cncf.io/projects/best-practices/governance/</a></p><p><a href="http://contribute.cncf.io/projects/best-practices/governance/templates/governance-elections/">contribute.cncf.io/projects/best-practices/governance/templates/governance-elections/</a></p><p><a href="http://github.com/cncf/toc/blob/main/PRINCIPLES.md">github.com/cncf/toc/blob/main/PRINCIPLES.md</a></p><p><strong>Q16.</strong></p><p><strong>At the fictional company Ceryvora Security Engineering, which option lists the Kubernetes Pod Security Standards from most permissive to most restrictive?</strong></p><p><strong>A. Privileged, Restricted, Baseline</strong></p><p><strong>B. Baseline, Restricted, Privileged</strong></p><p><strong>C. Baseline, Privileged, Restricted</strong></p><p><strong>D. Privileged, Baseline, Restricted</strong></p><p>A is incorrect because Restricted is the most restrictive profile and therefore cannot appear before Baseline in a sequence ordered from most permissive to most restrictive. Baseline is intended to prevent known privilege escalations while remaining compatible with many common containerized workloads, whereas Restricted applies stronger hardening requirements. The correct progression places Baseline between Privileged and Restricted.</p><p>B is incorrect because this sequence begins with Baseline even though Privileged is the most permissive policy. The Privileged profile is intentionally unrestricted and provides the widest level of permission. Restricted belongs at the opposite end of the sequence, so this ordering is incorrect at both ends.</p><p>C is incorrect because Privileged cannot appear between Baseline and Restricted when ordering by increasing restrictiveness. Privileged imposes the fewest restrictions, Baseline provides an intermediate policy intended to block known privilege escalations, and Restricted applies current Pod hardening best practices. Therefore Privileged must appear first.</p><blockquote><strong><em>D is correct because the three Kubernetes Pod Security Standards are ordered from least restrictive to most restrictive as Privileged, Baseline, and Restricted. Privileged is an intentionally unrestricted policy. Baseline aims to prevent known privilege escalations while minimizing restrictions on common workloads. Restricted builds on Baseline and enforces stronger hardening controls such as non-root execution, restricted capabilities, seccomp requirements, and prevention of privilege escalation for applicable Linux workloads. Pod Security Admission can enforce these profiles at namespace scope.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/security/pod-security-standards/">kubernetes.io/docs/concepts/security/pod-security-standards/</a></p><p><a href="http://kubernetes.io/docs/concepts/security/pod-security-admission/">kubernetes.io/docs/concepts/security/pod-security-admission/</a></p><p><strong>Q17.</strong></p><p><strong>At the fictional company Velqora Service Engineering, which Kubernetes Service type exposes a Service on a cluster-internal virtual IP and is normally reachable only from within the cluster network?</strong></p><p><strong>A. ClusterIP</strong></p><p><strong>B. LoadBalancer</strong></p><p><strong>C. ExternalName</strong></p><p><strong>D. NodePort</strong></p><blockquote><strong><em>A is correct because ClusterIP is the default Kubernetes Service type and exposes the Service through a cluster-internal virtual IP address. Clients inside the cluster can access the Service using that cluster IP or, more commonly, the Service’s DNS name. Kubernetes networking then directs Service traffic toward the eligible backend endpoints represented through EndpointSlices. The ClusterIP is normally routable only within the cluster network, which makes this Service type appropriate when an application should be reachable by other cluster workloads but should not be directly exposed through a Node port or external load balancer.</em></strong></blockquote><p>B is incorrect because LoadBalancer is specifically intended to expose a Service externally through a load-balancer implementation provided by the environment. A LoadBalancer Service builds on Service networking and can receive an externally reachable address when the cluster’s infrastructure supports it. Although the Service also has cluster-internal behavior, its defining purpose is not to limit access to the cluster. Therefore it does not match the requirement of a Service exposed only on a cluster-internal virtual IP.</p><p>C is incorrect because ExternalName is a special Service type that maps the Service’s DNS name to an external DNS name by returning a CNAME record. It does not allocate a normal cluster virtual IP and does not proxy traffic to a selected set of Pods. Although clients inside the cluster commonly resolve an ExternalName Service through cluster DNS, the Service represents an external name rather than the cluster-internal virtual-IP exposure model asked about in the question.</p><p>D is incorrect because NodePort exposes the Service on a port of every Node in addition to providing the underlying ClusterIP behavior. Clients that can reach a Node address and the allocated NodePort can potentially access the Service from outside the cluster network, subject to infrastructure routing and firewall rules. NodePort therefore provides a broader exposure mechanism than a ClusterIP-only Service.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/service/">kubernetes.io/docs/concepts/services-networking/service/</a></p><p><a href="http://kubernetes.io/docs/reference/kubernetes-api/core/service-v1/">kubernetes.io/docs/reference/kubernetes-api/core/service-v1/</a></p><p><strong>Q18.</strong></p><p><strong>At the fictional company Vireluxa Troubleshooting Systems, kubectl reports CrashLoopBackOff for a container in a Kubernetes Pod. Which of the following is the most likely explanation?</strong></p><p><strong>A. The container’s image is missing or cannot be pulled.</strong></p><p><strong>B. There are insufficient resources available to schedule the Pod.</strong></p><p><strong>C. The Pod is unable to communicate with the Kubernetes API server.</strong></p><p><strong>D. The application inside the container starts, crashes, and is repeatedly restarted.</strong></p><p>A is incorrect because a missing or inaccessible container image normally produces image-pull-related waiting reasons such as ErrImagePull or ImagePullBackOff rather than CrashLoopBackOff. In an image-pull failure, the container has not successfully started because kubelet and the container runtime cannot obtain the required image. CrashLoopBackOff instead indicates that the container has been started, has subsequently terminated, and Kubernetes is delaying repeated restart attempts.</p><p>B is incorrect because insufficient schedulable resources normally leave the Pod in Pending with FailedScheduling events explaining that no Node satisfies its CPU, memory, affinity, volume, or other requirements. A container cannot enter a repeated runtime crash cycle before the Pod is successfully scheduled and the container process starts. Resource limits can indirectly cause crashes after startup, for example through an OOM kill, but generic inability to schedule is not what CrashLoopBackOff means.</p><p>C is incorrect because loss of communication with kube-apiserver does not by itself define CrashLoopBackOff. Node or control-plane communication failures can result in Node conditions becoming Unknown, stale status information, or broader cluster-management problems. CrashLoopBackOff is specifically associated with repeated failure and restart of a container process on a Node rather than with the Pod being unable to contact the Kubernetes API.</p><blockquote><strong><em>D is correct because CrashLoopBackOff indicates that a container repeatedly starts, terminates, and is restarted, while kubelet applies an increasing backoff delay between restart attempts. Common causes include an application process exiting with an error, invalid runtime configuration, failed dependencies, incorrect commands or arguments, permissions problems, failing liveness probes, or resource-related termination such as OOM conditions. Kubernetes troubleshooting documentation demonstrates CrashLoopBackOff with a container that starts and immediately exits with a failure. The status is therefore a symptom of a restart loop rather than one specific root cause, but an application that crashes after starting is the best explanation among the options.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/tasks/debug/debug-application/debug-running-pod/">kubernetes.io/docs/tasks/debug/debug-application/debug-running-pod/</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/pods/pod-lifecycle/">kubernetes.io/docs/concepts/workloads/pods/pod-lifecycle/</a></p><p><a href="http://kubernetes.io/docs/tasks/debug/debug-application/debug-pods/">kubernetes.io/docs/tasks/debug/debug-application/debug-pods/</a></p><p><strong>Q19.</strong></p><p><strong>At the fictional company Lumoryx Helm Engineering, which file in a current Helm chart defines the chart’s default configuration values that users can override during installation or upgrade?</strong></p><p><strong>A. templates/deployment.yaml</strong></p><p><strong>B. Chart.yaml</strong></p><p><strong>C. values.yaml</strong></p><p><strong>D. Chart.lock</strong></p><p>A is incorrect because files under the templates directory are templates used to generate Kubernetes manifests rather than the central default-values file. A template such as templates/deployment.yaml can reference values through Helm’s .Values object and use those values to populate fields in the rendered Deployment. The defaults consumed by those templates are normally stored in values.yaml, not defined by the template file itself.</p><p>B is incorrect because Chart.yaml contains metadata about the Helm chart, including fields such as the chart API version, name, chart version, description, application version, and optional dependency declarations. It identifies and describes the chart rather than serving as the normal store for user-configurable default template values. The chart’s values are kept separately in values.yaml so users can override them without changing chart metadata.</p><blockquote><strong><em>C is correct because values.yaml is the Helm chart file that contains default configuration values. Chart templates access those values through the .Values object. During helm install or helm upgrade, a user can override the defaults by supplying additional values files with -f or — values, by using — set and related flags, or through other supported Helm value inputs. Helm merges those overrides according to its documented precedence rules. This separation lets one chart template be reused across different environments and deployment configurations.</em></strong></blockquote><p>D is incorrect because Chart.lock records a resolved snapshot of chart dependencies and is used to keep dependency resolution reproducible. It is not the file that contains default application configuration values consumed by templates. Current Helm charts declare dependencies in Chart.yaml, and dependency-management operations can generate Chart.lock. The obsolete Helm 2 dependency-file terminology has therefore been replaced in this option with the current Chart.lock file, but values.yaml remains the correct answer.</p><p><strong>Reference links:</strong></p><p><a href="http://helm.sh/docs/topics/charts/">helm.sh/docs/topics/charts/</a></p><p><a href="http://helm.sh/docs/chart_template_guide/values_files/">helm.sh/docs/chart_template_guide/values_files/</a></p><p><a href="http://helm.sh/docs/chart_template_guide/getting_started/">helm.sh/docs/chart_template_guide/getting_started/</a></p><p><strong>Q20.</strong></p><p><strong>At the fictional company Cerynova Troubleshooting Labs, a Pod is repeatedly entering CrashLoopBackOff. Which is the most appropriate initial troubleshooting approach?</strong></p><p><strong>A. Use kubectl top pod to check CPU usage and then scale the Deployment to more replicas.</strong></p><p><strong>B. Use kubectl describe pod to review state and recent events, then use kubectl logs to inspect the failing container’s output.</strong></p><p><strong>C. Use kubectl get nodes to verify Node capacity and then use kubectl apply -f to restart the Pod.</strong></p><p><strong>D. Use kubectl exec to open a shell in the repeatedly crashing container and then inspect a host kubelet log file from inside that container.</strong></p><p>A is incorrect because resource usage can be relevant to some failures, but kubectl top does not explain the full range of causes behind CrashLoopBackOff, and increasing replica count does not fix a container that repeatedly crashes because of bad configuration, a missing dependency, a failing command, permissions, an application defect, or another startup problem. Scaling could simply create more failing replicas. The more useful first step is to inspect Pod status and events and then examine the failing container’s logs.</p><blockquote><strong><em>B is correct because kubectl describe pod provides detailed information about the Pod’s container states, restart counts, conditions, configuration, scheduling information, and recent events. Those events can reveal issues such as failed mounts, probe failures, image-related problems, resource problems, or other runtime conditions. kubectl logs then retrieves the container’s stdout and stderr output, which frequently contains the application error that caused termination. If the container has already restarted, kubectl logs — previous can retrieve logs from the previous terminated instance. Together, describe and logs form the standard first-line troubleshooting workflow for CrashLoopBackOff.</em></strong></blockquote><p>C is incorrect because checking Nodes can be useful when a Pod is Pending due to scheduling or capacity problems, but CrashLoopBackOff means the container has been starting and terminating repeatedly after the Pod reached a Node. kubectl apply -f is a configuration-management command, not a generic Pod restart command, and blindly reapplying the same manifest may simply reproduce the same failure. Troubleshooting should identify the cause before attempting corrective redeployment.</p><p>D is incorrect because kubectl exec requires a container to remain running long enough for the command session to be established, which may be difficult in a rapid crash loop. It is also incorrect to assume that a host’s kubelet log exists at a particular path inside the application container; containers normally do not have access to arbitrary host log files. kubectl describe and kubectl logs are safer and more direct initial tools. If deeper debugging is required, Kubernetes also supports mechanisms such as ephemeral debug containers.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/tasks/debug/debug-application/debug-running-pod/">kubernetes.io/docs/tasks/debug/debug-application/debug-running-pod/</a></p><p><a href="http://kubernetes.io/docs/tasks/debug/debug-application/debug-pods/">kubernetes.io/docs/tasks/debug/debug-application/debug-pods/</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/pods/pod-lifecycle/">kubernetes.io/docs/concepts/workloads/pods/pod-lifecycle/</a></p><p><strong>Q21.</strong></p><p><strong>At the fictional company Lumivanta Service Platforms, what Kubernetes abstraction allows a group of Pods to be exposed through a stable network access mechanism inside a cluster?</strong></p><p><strong>A. Service</strong></p><p><strong>B. Deployment</strong></p><p><strong>C. Unit</strong></p><p><strong>D. Daemon</strong></p><blockquote><strong><em>A is correct because a Kubernetes Service defines a stable network abstraction over a logical set of backends, usually Pods. Pods are ephemeral and can be recreated with new IP addresses, so clients should not normally track individual Pod addresses. A selector-based Service identifies matching Pods, and Kubernetes manages EndpointSlice objects representing the current backends. The Service then provides a stable DNS name and, for applicable Service types, a virtual IP and port through which clients can reach the application. ClusterIP, the default Service type, is specifically intended for cluster-internal access.</em></strong></blockquote><p>B is incorrect because a Deployment manages application Pods through ReplicaSets and provides declarative replica management, rolling updates, and rollback. It does not itself create the stable client-facing network endpoint used to access those Pods. A common design is to run the application with a Deployment and expose its interchangeable replicas through a separate Service. The two resources therefore complement each other but serve different responsibilities.</p><p>C is incorrect because Unit is not a built-in Kubernetes API resource used to expose Pods. Kubernetes provides objects such as Service, Deployment, Pod, ConfigMap, Secret, Job, and many others, but there is no standard resource called Unit that provides Service networking. The stable networking abstraction requested in the question is Service.</p><p>D is incorrect because Daemon is not a built-in Kubernetes resource kind. DaemonSet is a Kubernetes workload controller that ensures Pods run on all or selected Nodes, but even a DaemonSet does not replace the Service abstraction when stable network access to its Pods is required. The question is specifically about exposing a group of Pods through a network abstraction, which is the role of Service.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/service/">kubernetes.io/docs/concepts/services-networking/service/</a></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/">kubernetes.io/docs/concepts/services-networking/</a></p><p><a href="http://kubernetes.io/docs/tutorials/kubernetes-basics/expose/expose-intro/">kubernetes.io/docs/tutorials/kubernetes-basics/expose/expose-intro/</a></p><p><strong>Q22.</strong></p><p><strong>At the fictional company Nexovanta Node Operations, which kubectl command can a cluster administrator use to temporarily prevent ordinary new Pods from being scheduled onto a Node for maintenance?</strong></p><p><strong>A. kubectl cordon</strong></p><p><strong>B. kubectl annotate</strong></p><p><strong>C. kubectl label</strong></p><p><strong>D. kubectl taint</strong></p><blockquote><strong><em>A is correct because kubectl cordon marks a Node as unschedulable. Once cordoned, kube-scheduler does not place ordinary new Pods onto that Node, while Pods that are already running there remain in place. This makes cordon useful when an administrator wants to prepare a Node for maintenance without immediately evicting its existing workloads. When the maintenance period is over, kubectl uncordon marks the Node schedulable again. DaemonSet Pods have special scheduling behavior and can still run on unschedulable Nodes when appropriate, but cordon is the standard administrative command for the stated requirement.</em></strong></blockquote><p>B is incorrect because kubectl annotate adds or updates non-identifying metadata on Kubernetes objects. An annotation can be consumed by automation that independently changes scheduling behavior, but adding an arbitrary annotation does not itself set the Node’s spec.unschedulable field. To directly prevent normal scheduling, the administrator should cordon the Node.</p><p>C is incorrect because kubectl label can add or modify Node labels, and those labels can influence scheduling if Pods use nodeSelector, node affinity, or other selector-based rules. However, a label alone does not globally mark the Node as unavailable for new scheduling. Changing labels could also have unintended effects on only the workloads whose placement rules reference those labels. cordon is the purpose-built reversible maintenance operation.</p><p>D is incorrect because a taint can repel Pods that lack matching tolerations and can therefore be used to control Node placement, but it is not the simplest or standard command for temporarily marking a maintenance Node unschedulable. Taints are policy-oriented and can express different effects such as NoSchedule, PreferNoSchedule, and NoExecute. kubectl cordon directly sets the Node’s unschedulable state and is the documented administrative mechanism for pausing normal scheduling during maintenance.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_cordon/">kubernetes.io/docs/reference/kubectl/generated/kubectl_cordon/</a></p><p><a href="http://kubernetes.io/docs/concepts/architecture/nodes/">kubernetes.io/docs/concepts/architecture/nodes/</a></p><p><a href="http://kubernetes.io/docs/reference/kubectl/quick-reference/">kubernetes.io/docs/reference/kubectl/quick-reference/</a></p><p><strong>Q23.</strong></p><p><strong>At the fictional company Velorian Observability Engineering, which statement best describes observability in a cloud-native environment?</strong></p><p><strong>A. Observability is the practice of collecting and analyzing logs, metrics, and traces to understand system behavior.</strong></p><p><strong>B. Observability is the use of alerts and dashboards only to notify administrators about system outages.</strong></p><p><strong>C. Observability is the process of restarting failed Pods to maintain application availability.</strong></p><p><strong>D. Observability is the configuration of health probes to check whether containers are running correctly.</strong></p><blockquote><strong><em>A is correct because observability is the practice of obtaining visibility into a system’s internal state and behavior from the telemetry it emits. Kubernetes documentation describes metrics, logs, and traces as the three traditional pillars of observability. Metrics provide numerical time-series measurements, logs provide chronological event records, and traces reveal relationships and latency across request paths. By collecting and correlating those signals, operators and developers can understand health, performance, failures, dependencies, and unexpected behavior rather than relying only on predetermined failure checks.</em></strong></blockquote><p>B is incorrect because alerts and dashboards are consumers and presentations of observability data, not the complete definition of observability. Alerts can notify teams when known conditions cross thresholds, and dashboards can summarize metrics or other telemetry, but mature observability also supports exploratory investigation of failures that were not anticipated in advance. Logs, traces, events, profiles, and contextual metadata can all contribute to understanding why a system behaves the way it does. Restricting observability to outage notifications is therefore too narrow.</p><p>C is incorrect because restarting failed Pods is part of Kubernetes self-healing and workload reconciliation rather than observability. Controllers and kubelet can restart containers or create replacement Pods to maintain desired state, but those actions do not by themselves provide information about why the failure occurred. Observability supplies the telemetry that helps people or automated systems understand the failure; self-healing acts on workload state.</p><p>D is incorrect because health probes are one input into application lifecycle management rather than a full observability system. Liveness, readiness, and startup probes answer targeted health questions and can trigger actions such as restarts or removal from Service traffic. They do not provide the broad historical, contextual, and distributed visibility obtained through metrics, logs, and traces. Probes can complement observability but do not define it.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/cluster-administration/observability/">kubernetes.io/docs/concepts/cluster-administration/observability/</a></p><p><a href="http://opentelemetry.io/docs/concepts/observability-primer/">opentelemetry.io/docs/concepts/observability-primer/</a></p><p><a href="http://opentelemetry.io/docs/concepts/signals/">opentelemetry.io/docs/concepts/signals/</a></p><p><strong>Q24.</strong></p><p><strong>At the fictional company Cinderova Gateway Engineering, Gateway API is the successor to which Kubernetes API?</strong></p><p><strong>A. NetworkPolicy</strong></p><p><strong>B. Service</strong></p><p><strong>C. NodePort</strong></p><p><strong>D. Ingress</strong></p><p>A is incorrect because NetworkPolicy addresses network isolation and allowed traffic flows between Pods and other network peers. Gateway API does not replace NetworkPolicy. A cluster can use both technologies at the same time: Gateway API describes traffic-handling infrastructure and routes, while NetworkPolicy controls whether selected network connections are permitted by the cluster network implementation.</p><p>B is incorrect because Service remains a fundamental Kubernetes API used to provide stable network access to backend endpoints. Gateway API routes frequently target Kubernetes Services as their backends, so Service is part of the networking model used together with Gateway API rather than an API being superseded by it. Gateway API provides higher-level traffic routing and infrastructure configuration on top of backend abstractions such as Services.</p><p>C is incorrect because NodePort is a value of the Kubernetes Service type field rather than a standalone API that Gateway API replaces. A NodePort Service exposes a Service on ports of the cluster Nodes, while Gateway API provides a separate and more expressive model for traffic-handling infrastructure and protocol-aware routing. Both can exist in the same cluster.</p><blockquote><strong><em>D is correct because Kubernetes documentation explicitly identifies Gateway API as the successor to the Ingress API. Ingress remains a stable supported API, but its feature set is frozen and new networking capabilities are being developed through Gateway API. Gateway API offers a richer, role-oriented, portable, and extensible resource model with objects such as GatewayClass, Gateway, HTTPRoute, GRPCRoute, and other route types. Migration from Ingress is not automatic because Gateway API uses different resources rather than introducing a new version of the Ingress kind.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/gateway/">kubernetes.io/docs/concepts/services-networking/gateway/</a></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/ingress/">kubernetes.io/docs/concepts/services-networking/ingress/</a></p><p><a href="http://gateway-api.sigs.k8s.io/">gateway-api.sigs.k8s.io/</a></p><p><strong>Q25.</strong></p><p><strong>At the fictional company Velunexa GitOps Delivery, in a Kubernetes environment, how should an Argo CD ApplicationSet be configured when the goal is continuous deployment from Git-managed application manifests?</strong></p><p><strong>A. Manually deploy the application with kubectl commands every time Git changes are made.</strong></p><p><strong>B. Define the ApplicationSet declaratively in YAML so it generates Applications that reference the Git repository containing the desired manifests, and configure automated synchronization for those generated Applications.</strong></p><p><strong>C. Define only deployment parameters in an ApplicationSet and apply it once with kubectl, without configuring a desired-state source or automated synchronization.</strong></p><p><strong>D. Package the application as a Helm chart and deploy it directly outside the GitOps reconciliation flow.</strong></p><p>A is incorrect because repeated manual kubectl deployment bypasses the GitOps reconciliation model that Argo CD is designed to provide. In Argo CD, desired application configuration is stored in a source such as Git and represented through Application resources. The controller compares live state with desired state and can synchronize differences automatically. A manual deployment process would reintroduce imperative cluster changes, reduce drift control, and make Git no longer the authoritative deployment path.</p><blockquote><strong><em>B is correct because ApplicationSet is a Kubernetes custom resource that generates one or more Argo CD Application resources from generator input and an Application template. Those generated Applications can reference a Git repository, revision, and path containing the desired Kubernetes manifests or configuration. If the generated Application template also enables automated synchronization, Argo CD can detect differences between the Git-managed desired state and the live cluster and apply new manifests automatically. The ApplicationSet itself can also be stored in Git and reconciled through an Argo CD parent Application, preserving the GitOps model for both application generation and application deployment.</em></strong></blockquote><p>C is incorrect because applying an ApplicationSet object once with kubectl can create the resource, but deployment parameters alone do not establish continuous deployment. The generated Applications need a desired-state source, destination, and relevant sync policy. Without a Git or other supported source and without automated synchronization, Argo CD does not have the configuration required to automatically deploy each new desired-state change. A one-time kubectl apply is therefore not sufficient for the continuous-deployment behavior described in the question.</p><p>D is incorrect because Helm is a valid application packaging mechanism and Argo CD can deploy Helm charts through GitOps, but deploying the chart directly outside the Argo CD reconciliation flow does not satisfy the requirement. The key requirement is continuous reconciliation of desired state, not whether the manifests originate from raw YAML, Kustomize, or Helm. A Helm chart can be referenced by the generated Applications, but direct out-of-band deployment would bypass Argo CD’s desired-state controller.</p><p><strong>Reference links:</strong></p><p><a href="http://argo-cd.readthedocs.io/en/latest/operator-manual/applicationset/">argo-cd.readthedocs.io/en/latest/operator-manual/applicationset/</a></p><p><a href="http://argo-cd.readthedocs.io/en/latest/user-guide/application-specification/">argo-cd.readthedocs.io/en/latest/user-guide/application-specification/</a></p><p><a href="http://argo-cd.readthedocs.io/en/latest/user-guide/ci_automation/">argo-cd.readthedocs.io/en/latest/user-guide/ci_automation/</a></p><p><strong>Q26.</strong></p><p><strong>At the fictional company Ceryvanta Traffic Platforms, which Kubernetes API provides a standardized and portable configuration model for gateway load balancing, TLS termination, and rich traffic routing across multiple implementations?</strong></p><p><strong>A. Gateway API</strong></p><p><strong>B. CoreDNS</strong></p><p><strong>C. Ingress</strong></p><p><strong>D. kube-proxy</strong></p><blockquote><strong><em>A is correct because Gateway API is a standardized, role-oriented, portable, and extensible set of Kubernetes networking APIs for configuring traffic-handling infrastructure and routes. Gateway resources describe instances of traffic-handling infrastructure such as load balancers or proxies, while resources such as HTTPRoute and GRPCRoute define protocol-aware routing to backends. Gateway listeners can be configured for TLS termination, and route rules support features such as host and path matching, header matching, weighted traffic splitting, and other portable capabilities depending on the conformance level and implementation. Gateway API itself is configuration rather than a packet-processing engine; a compatible implementation performs the actual load balancing and traffic handling.</em></strong></blockquote><p>B is incorrect because CoreDNS provides DNS resolution and service discovery rather than gateway load balancing and application traffic routing. It can resolve Kubernetes Service names and forward external DNS queries, but it does not interpret HTTPRoute resources, terminate application TLS connections, or implement Gateway API routing behavior.</p><p>C is incorrect because Ingress provides HTTP and HTTPS routing and can support TLS termination through an Ingress controller, but it is the older and less expressive API model. Kubernetes has frozen Ingress feature development and identifies Gateway API as its successor for new networking capabilities. Ingress implementations also historically relied heavily on controller-specific annotations for advanced functions that Gateway API attempts to represent through portable structured fields.</p><p>D is incorrect because kube-proxy implements Kubernetes Service virtual-IP networking where it is used. It operates at the Node networking layer and directs Service traffic to suitable backends, but it does not provide the high-level role-oriented gateway resource model, protocol-specific route APIs, portable TLS listener configuration, or advanced application routing represented by Gateway API.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/gateway/">kubernetes.io/docs/concepts/services-networking/gateway/</a></p><p><a href="http://gateway-api.sigs.k8s.io/">gateway-api.sigs.k8s.io/</a></p><p><a href="http://gateway-api.sigs.k8s.io/api-types/gateway/">gateway-api.sigs.k8s.io/api-types/gateway/</a></p><p><strong>Q27.</strong></p><p><strong>At the fictional company Lumivexa Certificate Systems, how can a kubernetes.io/tls Secret named tls-secret be created from an existing PEM-encoded TLS certificate and its matching private key using kubectl?</strong></p><p><strong>A. kubectl create secret generic tls-secret — cert=path/to/tls.crt — key=path/to/tls.key</strong></p><p><strong>B. kubectl create secret tls tls-secret — cert=path/to/tls.crt — key=path/to/tls.key</strong></p><p><strong>C. kubectl create secret tls tls-secret — from-file=cert=path/to/tls.crt — from-file=key=path/to/tls.key</strong></p><p><strong>D. kubectl create secret generic tls-secret — from-file=cert=path/to/tls.crt — from-file=key=path/to/tls.key</strong></p><p>A is incorrect because kubectl create secret generic creates an Opaque Secret rather than using the dedicated TLS Secret subcommand. The — cert and — key flags belong to kubectl create secret tls rather than the generic subcommand. If an application deliberately needed arbitrary files stored in an Opaque Secret, — from-file could be used with secret generic, but that would not create the standard kubernetes.io/tls Secret type requested by the question.</p><blockquote><strong><em>B is correct because kubectl create secret tls is the purpose-built command for creating a TLS Secret from a public certificate and its associated private key. The documented syntax is kubectl create secret tls NAME — cert=path/to/cert/file — key=path/to/key/file. The certificate must be PEM encoded and match the supplied private key. Kubernetes creates a Secret of type kubernetes.io/tls containing the conventional tls.crt and tls.key data entries. Such Secrets are commonly referenced by TLS-capable components including Ingress and Gateway implementations.</em></strong></blockquote><p>C is incorrect because the dedicated tls subcommand expects the — cert and — key options rather than generic — from-file mappings named cert and key. A TLS Secret also uses the conventional keys tls.crt and tls.key. The command shown therefore does not match the documented kubectl create secret tls interface. Option B uses the correct specialized TLS Secret syntax.</p><p>D is incorrect because kubectl create secret generic with — from-file can store files in an Opaque Secret, but it does not create the standard kubernetes.io/tls Secret type simply because the source files happen to contain a certificate and private key. The question specifically requests a tls-type Secret, so the dedicated kubectl create secret tls command is the appropriate method.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_create/kubectl_create_secret_tls/">kubernetes.io/docs/reference/kubectl/generated/kubectl_create/kubectl_create_secret_tls/</a></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_create/kubectl_create_secret/">kubernetes.io/docs/reference/kubectl/generated/kubectl_create/kubectl_create_secret/</a></p><p><a href="http://kubernetes.io/docs/concepts/configuration/secret/">kubernetes.io/docs/concepts/configuration/secret/</a></p><p><strong>Q28.</strong></p><p><strong>At the fictional company Orivanta Service Discovery, for a normal non-headless Kubernetes Service that has a cluster IP, how does DNS resolution work inside the cluster?</strong></p><p><strong>A. DNS resolution for Services uses the IP addresses of the Pods directly.</strong></p><p><strong>B. DNS resolution for Services is handled by an external DNS provider that manages all Service names.</strong></p><p><strong>C. Kubernetes creates Service DNS names that resolve to stable cluster IPs.</strong></p><p><strong>D. Service names in Kubernetes are resolved using environment variables set in each Pod.</strong></p><p>A is incorrect because a normal Service DNS name does not resolve directly to the individual Pod IP addresses behind that Service. For a standard non-headless ClusterIP-style Service, Kubernetes cluster DNS publishes an A and/or AAAA record whose result is the Service’s cluster IP. Kubernetes Service networking then directs traffic sent to that Service virtual IP toward eligible backend endpoints. Headless Services are the important exception: their DNS names can resolve to backend addresses rather than to a cluster IP. The question explicitly asks about a normal non-headless Service.</p><p>B is incorrect because Kubernetes normally provides in-cluster DNS through a cluster DNS add-on such as CoreDNS. External DNS providers can integrate with Kubernetes and can publish records for externally reachable services, but they are not required to resolve ordinary Service names inside the cluster. CoreDNS can watch Kubernetes API state and provide DNS responses for the cluster’s Service and Pod naming domains.</p><blockquote><strong><em>C is correct because Kubernetes assigns DNS names to Services, and a normal non-headless Service receives DNS A and/or AAAA records that resolve to its cluster IP address or addresses. A workload can therefore call a Service using a stable name such as service-name.namespace.svc.cluster-domain rather than tracking the addresses of changing backend Pods. The Service cluster IP and DNS name remain stable across ordinary Pod replacement, while EndpointSlice information tracks the current backend endpoints. This stable discovery model is one of the core benefits of Kubernetes Services.</em></strong></blockquote><p>D is incorrect because environment variables are a separate built-in Service discovery mechanism and are not how Kubernetes DNS names are resolved. kubelet can inject environment variables for Services that exist when a Pod starts, but those variables have creation-order limitations and do not constitute DNS resolution. Cluster DNS is the name-resolution mechanism, while Service environment variables are an alternative discovery method.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/dns-pod-service/">kubernetes.io/docs/concepts/services-networking/dns-pod-service/</a></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/service/">kubernetes.io/docs/concepts/services-networking/service/</a></p><p><a href="http://kubernetes.io/docs/concepts/services-networking/endpoint-slices/">kubernetes.io/docs/concepts/services-networking/endpoint-slices/</a></p><p><strong>Q29.</strong></p><p><strong>At the fictional company Velunex Community Labs, CNCF supports regional, community-organized events known as KCDs. What does KCD stand for?</strong></p><p><strong>A. Kubernetes Cloud Developers</strong></p><p><strong>B. Kubernetes Cloud Days</strong></p><p><strong>C. Kubernetes Community Days</strong></p><p><strong>D. Kubernetes Community Developers</strong></p><p>A is incorrect because Kubernetes Cloud Developers is not the official expansion of KCD. Developers are an important part of the Kubernetes and cloud-native community, but the KCD program is broader and brings together adopters, contributors, operators, developers, students, and other technologists. The program name emphasizes the local community rather than one professional role.</p><p>B is incorrect because Kubernetes Cloud Days is not the official name of the CNCF-supported event program. The events focus on Kubernetes and the wider cloud-native ecosystem, but the letter C in KCD stands for Community rather than Cloud. Current CNCF event materials consistently refer to these gatherings as Kubernetes Community Days.</p><blockquote><strong><em>C is correct because KCD stands for Kubernetes Community Days. KCDs are community-organized, CNCF-supported regional events that bring open-source adopters and cloud-native technologists together for technical education, collaboration, networking, and local community building. The program is designed to strengthen regional Kubernetes and cloud-native communities and provides opportunities for attendees, speakers, organizers, and sponsors to participate locally.</em></strong></blockquote><p>D is incorrect because Kubernetes Community Developers is not the program’s official name. KCD events are not limited to software developers and commonly include platform engineers, operators, security practitioners, students, project contributors, end users, and other cloud-native community members. The correct expansion is Kubernetes Community Days.</p><p><strong>Reference links:</strong></p><p><a href="http://cncf.io/kcds/">cncf.io/kcds/</a></p><p><a href="http://cncf.io/events/">cncf.io/events/</a></p><p><strong>Q30.</strong></p><p><strong>At the fictional company Ceryvanta Node Operations, a DevOps engineer is preparing worker-1 for maintenance. The Node is running multiple Pods, but eligible workloads must be safely evicted so their controllers can recreate them on other Nodes. Which command should the engineer use?</strong></p><p><strong>A. kubectl delete node worker-1 — delete-emptydir-data</strong></p><p><strong>B. kubectl drain worker-1 — ignore-daemonsets</strong></p><p><strong>C. kubectl scale deployment — replicas=0 — namespace=production</strong></p><p><strong>D. kubectl cordon worker-1 — daemonsets=off</strong></p><p>A is incorrect because deleting the Node object is not the standard maintenance workflow for safely evicting workloads. A physical or virtual Node can continue to exist even if its Kubernetes API object is deleted, and deleting the object does not provide the controlled policy-aware eviction sequence intended for planned maintenance. The — delete-emptydir-data option belongs to kubectl drain rather than kubectl delete node. Administrators should normally drain the Node before shutting down or servicing it.</p><blockquote><strong><em>B is correct because kubectl drain marks worker-1 unschedulable and safely evicts eligible Pods in preparation for maintenance. The eviction process respects graceful termination and PodDisruptionBudgets when the Eviction API is used. If DaemonSet-managed Pods exist, drain requires — ignore-daemonsets because those Pods are expected to remain managed by the DaemonSet controller and would otherwise be recreated on the Node. Depending on the workloads, additional options such as — delete-emptydir-data or — force may be necessary, but the command shown is the standard starting point for safely draining a Node that contains DaemonSet Pods.</em></strong></blockquote><p>C is incorrect because scaling a Deployment in one namespace to zero addresses only that particular workload and does not safely evacuate all eligible Pods from worker-1. The Node can host Pods from many controllers and namespaces, including StatefulSets, Jobs, ReplicaSets, and other Deployments. Scaling application workloads to zero can also cause unnecessary service disruption. kubectl drain is specifically designed to coordinate Node maintenance across the Pods running on that Node.</p><p>D is incorrect because kubectl cordon is a valid command for marking a Node unschedulable, but — daemonsets=off is not a supported cordon option. More importantly, cordon alone leaves existing Pods running on the Node. That is appropriate when the administrator only wants to stop new scheduling, but the scenario explicitly requires workloads to be safely evicted and rescheduled before maintenance. kubectl drain performs both the unschedulable step and the eviction process.</p><p><strong>Reference links:</strong></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_drain/">kubernetes.io/docs/reference/kubectl/generated/kubectl_drain/</a></p><p><a href="http://kubernetes.io/docs/tasks/administer-cluster/safely-drain-node/">kubernetes.io/docs/tasks/administer-cluster/safely-drain-node/</a></p><p><a href="http://kubernetes.io/docs/reference/kubectl/generated/kubectl_cordon/">kubernetes.io/docs/reference/kubectl/generated/kubectl_cordon/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e79560e06157" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Professional Cloud Database Engineer Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-professional-cloud-database-engineer-practice-questions-85a91e38c490?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/85a91e38c490</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[google-cloud]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 08:10:39 GMT</pubDate>
            <atom:updated>2026-10-06T08:10:39.647Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w4lfNQON21dgVtBUsvmhIw.jpeg" /><figcaption>KloudExams GCP Google Cloud Professional Cloud Database Engineer Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Professional Cloud Database Engineer certification</strong>? Knowing what database services do is a starting point — understanding how to choose, configure, migrate, and manage them to meet application requirements is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing database solutions and operational approaches that fit specific requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on database practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>Zorvella Retail is a large retail and ecommerce company expanding globally and migrating to Google Cloud. You want a platform that scales easily, handles transactions with low latency, and provides a reliable customer experience. You need a storage layer for sales transactions and current inventory levels, and you want to retain a relational schema. What should you do?</strong></p><p><strong>A. Store the data in Firestore in a multi-region location, and place compute resources in one of the constituent regions.</strong></p><p><strong>B. Deploy Spanner using a multi-region instance, and place compute resources close to the default leader region.</strong></p><p><strong>C. Build an in-memory cache in Memorystore, and deploy it to the geographic regions where the application resides.</strong></p><p><strong>D. Deploy a Bigtable instance with a cluster in one region and a replica cluster in another geographic region.</strong></p><p>A is incorrect because Firestore is a NoSQL document database. It stores data as documents organized into collections rather than as relational tables with the same relational schema used by a traditional transactional system. Firestore can provide highly available multi-region storage and is appropriate for many globally distributed application workloads, but choosing it here would require a data-model redesign and would not satisfy the explicit requirement to retain a relational schema for sales transactions and inventory.</p><blockquote><strong><em>B is correct because Spanner is a horizontally scalable relational database that supports SQL, schemas, transactions, and strong consistency while distributing data across regions. Multi-region Spanner configurations can provide up to 99.999% availability with the appropriate edition and preserve transactional semantics across geographically distributed replicas. For write-heavy traffic, Google recommends placing application compute within or close to the default leader region to reduce write latency, while replicas can provide read locality in other regions. This combination directly matches the requirements for a relational global transaction system, scalable inventory and sales data, low latency, and high reliability.</em></strong></blockquote><p>C is incorrect because Memorystore is designed as a managed in-memory data store and cache rather than as the durable relational system of record for globally important sales transactions and inventory balances. A cache can reduce read latency for selected application data, but it should not replace the authoritative transactional database in this scenario. It also does not let the company retain the existing relational schema as its primary persistent data model.</p><p>D is incorrect because Bigtable is a massively scalable NoSQL wide-column database organized around row keys, not a general-purpose relational database. It does not support joins, and its transactional guarantees are scoped differently from those expected in a relational sales and inventory system. Bigtable can deliver very high throughput and low latency for suitable key-based workloads, but adopting it would require redesigning the relational schema and access patterns. That conflicts with the requirement to retain the existing relational model.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/instance-configurations">docs.cloud.google.com/spanner/docs/instance-configurations</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/editions-overview">docs.cloud.google.com/spanner/docs/editions-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/true-time-external-consistency">docs.cloud.google.com/spanner/docs/true-time-external-consistency</a></p><p><strong>Q2.</strong></p><p><strong>Brinovia Financial operates in a highly regulated industry where separation of concerns and the principle of least privilege are critical. The operations team consists of:</strong></p><p><strong>Person A, a database administrator.</strong></p><p><strong>Person B, an analyst who generates metric reports.</strong></p><p><strong>Application C, which is responsible for automated backups.</strong></p><p><strong>You need to assign IAM roles for Spanner. Which roles should you assign?</strong></p><p><strong>A. roles/spanner.databaseAdmin for Person A</strong></p><p><strong>roles/spanner.databaseReader for Person B</strong></p><p><strong>roles/spanner.backupWriter for Application C</strong></p><p><strong>B. roles/spanner.databaseAdmin for Person A</strong></p><p><strong>roles/spanner.databaseReader for Person B</strong></p><p><strong>roles/spanner.backupAdmin for Application C</strong></p><p><strong>C. roles/spanner.databaseAdmin for Person A</strong></p><p><strong>roles/spanner.databaseUser for Person B</strong></p><p><strong>roles/spanner.databaseReader for Application C</strong></p><p><strong>D. roles/spanner.databaseAdmin for Person A</strong></p><p><strong>roles/spanner.databaseUser for Person B</strong></p><p><strong>roles/spanner.backupWriter for Application C</strong></p><blockquote><strong><em>A is correct because the three roles align closely with the stated duties and therefore best implement least privilege. roles/spanner.databaseAdmin gives Person A the database-administration capabilities expected of a DBA. roles/spanner.databaseReader lets Person B read database data for reporting without granting write privileges. roles/spanner.backupWriter is specifically intended for principals or scripts that automate backup creation; it can create and view the relevant backup resources but does not grant the broader update and delete authority of a backup administrator. This mapping also supports separation of duties by keeping database administration, reporting, and backup creation distinct.</em></strong></blockquote><p>B is incorrect because the role assignments for Persons A and B are appropriate, but roles/spanner.backupAdmin grants Application C more backup authority than an automated backup creator needs. The backup administrator role can create, view, update, and delete backups and manage additional backup-related capabilities. The backup writer role is intentionally narrower and is documented for scripts that automate backup creation. Under least privilege, the narrower backupWriter role is preferable.</p><p>C is incorrect because roles/spanner.databaseUser grants Person B read and write access that exceeds the needs of an analyst who only generates reports, while assigning roles/spanner.databaseReader to Application C does not give the application the permissions required to create backups. Even after correcting the role-ID punctuation, the responsibilities are mismatched. The analyst should receive the read-only database role, and the automated backup application should receive backupWriter.</p><p>D is incorrect because roles/spanner.backupWriter is appropriate for Application C, and roles/spanner.databaseAdmin is appropriate for the DBA, but roles/spanner.databaseUser gives the reporting analyst unnecessary write privileges. Least privilege requires granting only the permissions needed for the job. Since Person B needs to read data to generate metric reports, roles/spanner.databaseReader is the better fit.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/iam">docs.cloud.google.com/spanner/docs/iam</a></p><p><a href="http://docs.cloud.google.com/iam/docs/roles-permissions/spanner">docs.cloud.google.com/iam/docs/roles-permissions/spanner</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/backup/create-backups">docs.cloud.google.com/spanner/docs/backup/create-backups</a></p><p><strong>Q3.</strong></p><p><strong>Vantrelix Supply uses Spanner for a mission-critical, globally available inventory-management system. The company recently loaded stock keeping unit (SKU) and product-catalog data from an acquisition and observed hotspots in the Spanner database. You want to follow Google-recommended schema-design practices to avoid performance degradation. What should you do? (Choose two.)</strong></p><p><strong>A. Use an auto-incrementing value as the primary key.</strong></p><p><strong>B. Normalize the data model.</strong></p><p><strong>C. Promote low-cardinality attributes in multi-attribute primary keys.</strong></p><p><strong>D. Promote high-cardinality attributes in multi-attribute primary keys.</strong></p><p><strong>E. Use a bit-reversed sequential value as the primary key.</strong></p><p>A is incorrect because an ordinary auto-incrementing primary key produces monotonically increasing values. Spanner distributes data by key ranges, so a stream of new rows with increasing keys tends to direct inserts toward the same end of the key space. At scale, this creates a hotspot in which a limited set of splits or servers receives a disproportionate share of write traffic. Spanner documentation specifically warns against monotonically increasing keys for high-rate inserts unless the values are transformed or otherwise distributed.</p><p>B is incorrect because normalization is a relational modeling technique that can reduce redundancy and improve data integrity, but it is not the direct remedy for a Spanner key-range hotspot. Hotspots are primarily driven by skewed access patterns and key designs that concentrate traffic into a narrow portion of the key space. Depending on the workload, excessive normalization can also increase joins or distributed access. The schema change needed here should focus on distributing primary-key values and traffic more evenly.</p><p>C is incorrect because putting a low-cardinality attribute at the leading part of a composite primary key creates only a small number of key prefixes. If incoming rows or frequently accessed data are concentrated in one or a few of those values, traffic can still collect into a limited portion of the key space. That works against the goal of spreading writes broadly. For hotspot avoidance, the leading key components should provide enough distribution so that traffic is not funneled into a small number of contiguous ranges.</p><blockquote><strong><em>D is correct because high-cardinality leading attributes in a multi-column primary key can spread rows and requests across a much larger number of key ranges than low-cardinality prefixes. The important principle is to design the primary key so that frequently inserted or accessed rows do not all land in a narrow contiguous range. High-cardinality attributes are therefore preferable among these two composite-key choices when they distribute the workload effectively and still support the application’s access patterns.</em></strong></blockquote><blockquote><strong><em>E is correct because Spanner supports bit-reversed sequences specifically to preserve unique integer-style identifiers while distributing generated values across the positive integer key space. Bit reversal prevents successive logical sequence values from remaining adjacent in the high-order key space, which helps spread inserts across splits and avoid the hotspot created by monotonically increasing IDs. For workloads that require numeric generated keys, a bit-reversed sequence is a current Spanner-supported technique for hotspot avoidance.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/schema-design">docs.cloud.google.com/spanner/docs/schema-design</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/primary-keys-overview">docs.cloud.google.com/spanner/docs/primary-keys-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/primary-key-default-value">docs.cloud.google.com/spanner/docs/primary-key-default-value</a></p><p><strong>Q4.</strong></p><p><strong>Marovex Industries operates MySQL databases both on Compute Engine in Google Cloud and in its own data center as a hybrid environment. You need to create replicas in Cloud SQL to scale read workloads while offloading part of replica administration to Google Cloud. What should you do?</strong></p><p><strong>A. Use replication from an external server to create Cloud SQL replicas of the self-managed MySQL source servers.</strong></p><p><strong>B. Use Database Migration Service.</strong></p><p><strong>C. Use a Cloud SQL for MySQL external read replica.</strong></p><p><strong>D. Use mysqldump together with binary logs to build and maintain the replicas manually.</strong></p><blockquote><strong><em>A is correct because current Cloud SQL terminology defines replication from an external server as a topology in which a self-managed MySQL server, including one on-premises or on Compute Engine, is the source and one or more Cloud SQL instances are Cloud SQL replicas. Google Cloud specifically documents hybrid cloud and offloading administration of replicas to Cloud SQL as use cases for this configuration. It lets the organization keep control of the existing source databases while using managed Cloud SQL replicas for additional read capacity and reduced replica-management overhead.</em></strong></blockquote><p>B is incorrect because Database Migration Service is primarily designed to migrate databases into supported Google Cloud database destinations, including continuous migrations that keep a destination synchronized until cutover. It can be useful when the eventual goal is to promote Cloud SQL and move production ownership there, but the requirement here is to maintain a hybrid configuration and create managed replicas for ongoing read scaling. External-server replication is the Cloud SQL feature explicitly designed for that persistent hybrid-replica use case.</p><p>C is incorrect because an external read replica has the opposite replication direction under current Cloud SQL terminology. An external read replica is a MySQL server outside Cloud SQL that replicates from a Cloud SQL primary. The question instead starts with self-managed sources in the data center or on Compute Engine and asks for replicas managed in Cloud SQL. Those destinations are called Cloud SQL replicas in a replication-from-an-external-server configuration, not external read replicas.</p><p>D is incorrect because mysqldump and binary logs can be assembled into a manual replication process, but doing so leaves the organization responsible for seeding, replication setup, monitoring, failure recovery, patching, and lifecycle management of the replica environment. That conflicts with the requirement to offload part of the management operation. Cloud SQL’s external-server replication provides a managed destination and a documented hybrid-cloud topology without requiring the team to construct the entire replica lifecycle manually.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/replication">docs.cloud.google.com/sql/docs/mysql/replication</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/replication/external-server">docs.cloud.google.com/sql/docs/mysql/replication/external-server</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/replication/configure-replication-from-external">docs.cloud.google.com/sql/docs/mysql/replication/configure-replication-from-external</a></p><p><strong>Q5.</strong></p><p><strong>Trivanta Markets is moving the infrastructure for a highly transactional application to Google Cloud. A MySQL instance contains several databases, and you need to transfer approximately 500 GB of data to Cloud SQL for MySQL while minimizing downtime. What should you do?</strong></p><p><strong>A. 1. Create a Cloud SQL for MySQL instance for the databases, and configure Datastream to stream database changes directly to Cloud SQL.</strong></p><p><strong>2. Select Backfill historical data in the stream configuration so Datastream backfills data that is out of sync between the source and destination.</strong></p><p><strong>3. Delete the stream after all changes reach Cloud SQL, and update the application to use the new instance.</strong></p><p><strong>B. 1. Create a migration job in Database Migration Service.</strong></p><p><strong>2. Set the migration job type to Continuous, let the databases complete the full dump phase, and then let the job process ongoing changes in change data capture (CDC) mode.</strong></p><p><strong>3. Wait for replication delay to become minimal, stop source writes as required for cutover, promote the Cloud SQL destination, and wait for the migration job to complete.</strong></p><p><strong>4. Update the application connections to the new instance.</strong></p><p><strong>C. 1. Create a migration job in Database Migration Service.</strong></p><p><strong>2. Set the migration job type to One-time, and schedule the migration during a maintenance window.</strong></p><p><strong>3. Stop all writes to the source database and initiate the dump. Wait for the dump to load into Cloud SQL and for the destination to become available.</strong></p><p><strong>4. Update the application connections to the new instance.</strong></p><p><strong>D. 1. Use mysqldump to initiate a manual backup during an application maintenance window.</strong></p><p><strong>2. Move the files to Cloud Storage and import each database into Cloud SQL.</strong></p><p><strong>3. Continue creating dumps until all databases are migrated.</strong></p><p><strong>4. Update the application connections to the new instance.</strong></p><p>A is incorrect because Datastream is a serverless change-data-capture service, but Cloud SQL is not a direct Datastream destination in the way described here. Current Datastream destination patterns include services such as BigQuery, Cloud Storage, and supported data-lake destinations rather than using a Cloud SQL instance as the direct target for this migration workflow. Database Migration Service is the purpose-built managed service for homogeneous MySQL-to-Cloud-SQL migration and includes the full-copy, ongoing-replication, and promotion workflow required here.</p><blockquote><strong><em>B is correct because a continuous Database Migration Service migration is specifically designed to minimize downtime. It begins with a full dump and load while the source can continue receiving writes, then applies ongoing changes through CDC. Once the 500 GB destination has loaded and replication delay has fallen sufficiently, the team can stop writes briefly, let the destination fully catch up, promote the Cloud SQL instance, and redirect the application. Most of the time-consuming data transfer therefore occurs while production remains online, limiting outage time to the controlled final cutover.</em></strong></blockquote><p>C is incorrect because a one-time migration transfers a snapshot rather than maintaining continuous synchronization with new source writes. To guarantee consistency, the application would need to stop writes for a substantial portion of the full dump and load of a 500 GB transactional instance. That produces much more downtime than a continuous migration. One-time jobs are appropriate when downtime is acceptable or the source can remain static, which is not the priority stated here.</p><p>D is incorrect because repeated manual mysqldump operations create significant operational work and do not by themselves provide a clean, continuously synchronized cutover point. A large dump and import can take considerable time, during which an active transactional system continues to change. The team would need extra binary-log handling or repeated reconciliation to avoid losing those changes. Database Migration Service automates the initial copy, ongoing change capture, monitoring, and promotion workflow and is therefore the better minimal-downtime approach.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/database-migration/docs/mysql/migration-types">docs.cloud.google.com/database-migration/docs/mysql/migration-types</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/mysql/promote-migration">docs.cloud.google.com/database-migration/docs/mysql/promote-migration</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/mysql/quickstart">docs.cloud.google.com/database-migration/docs/mysql/quickstart</a></p><p><strong>Q6.</strong></p><p><strong>At Brivanta Media, your company uses Bigtable for a user-facing application that displays a low-latency, real-time dashboard. You need to recommend the optimal storage type for this read-intensive database. What should you do?</strong></p><p><strong>A. Recommend solid-state drive (SSD) storage.</strong></p><p><strong>B. Recommend splitting the Bigtable instance into two instances to load balance the concurrent reads.</strong></p><p><strong>C. Recommend hard disk drive (HDD) storage.</strong></p><p><strong>D. Recommend mixed storage types.</strong></p><blockquote><strong><em>A is correct because SSD storage is the recommended Bigtable storage choice for most latency-sensitive workloads and is particularly appropriate for a user-facing, read-intensive real-time dashboard. Bigtable documentation states that SSD provides significantly lower and more predictable read and write latency than HDD and much higher throughput per node. Point reads on HDD are especially disadvantaged because of disk seek latency, while SSD is designed for active data that requires frequent, low-latency access. This directly matches the stated requirement.</em></strong></blockquote><p>B is incorrect because creating two independent Bigtable instances does not provide transparent read load balancing for one logical database and would require the application to manage data duplication and routing between separate instances. If geographic replication, availability, or read distribution is needed, Bigtable supports multiple clusters within an instance. However, the question asks specifically for the optimal storage type, and SSD is the appropriate choice for low-latency read-heavy access.</p><p>C is incorrect because HDD storage is best suited to large, latency-insensitive datasets that are accessed infrequently or primarily through batch scans and writes. Current Bigtable documentation specifically notes that HDD is generally not appropriate for a real-time dashboard because individual row reads are substantially slower than on SSD. The lower storage cost does not compensate for the latency penalty in this user-facing workload.</p><p>D is incorrect because a Bigtable instance does not use an arbitrary mixture of SSD and HDD clusters as a general strategy for serving one low-latency workload. The storage choice is associated with the instance design, and current Bigtable also offers tiered storage capabilities on eligible SSD deployments for infrequently accessed data. For the active read-intensive data in this scenario, SSD remains the correct primary storage type.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigtable/docs/choosing-ssd-hdd">docs.cloud.google.com/bigtable/docs/choosing-ssd-hdd</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/performance">docs.cloud.google.com/bigtable/docs/performance</a></p><p><strong>Q7.</strong></p><p><strong>At Orbinexa Manufacturing, you need to migrate databases from Microsoft SQL Server 2016 Standard Edition on a Windows Server 2019 Datacenter server to a single Cloud SQL for SQL Server instance. During discovery, you observe peaks of approximately 25,000 read IOPS. You need to size the Cloud SQL instance to maximize read performance. What should you do?</strong></p><p><strong>A. Create SQL Server 2019 Standard on a Cloud SQL Enterprise edition general purpose dedicated-core configuration with 4 vCPUs, 15 GB of RAM, and 800 GB of SSD storage.</strong></p><p><strong>B. Create SQL Server 2019 Standard on a Cloud SQL Enterprise edition general purpose dedicated-core configuration with at least 16 vCPUs, 104 GB of RAM, and 200 GB of SSD storage.</strong></p><p><strong>C. Create SQL Server 2019 Standard on a Cloud SQL Enterprise edition general purpose dedicated-core configuration with 16 vCPUs, 104 GB of RAM, and 4 TB of SSD storage.</strong></p><p><strong>D. Create SQL Server 2019 Enterprise on a Cloud SQL Enterprise edition general purpose dedicated-core configuration with 16 vCPUs, 104 GB of RAM, and 500 GB of SSD storage.</strong></p><p>A is incorrect because 4 vCPUs do not provide the same IOPS ceiling as a 16-vCPU configuration, and the 800 GB disk is also smaller than the storage configuration in option C. Current Cloud SQL storage guidance shows that SSD IOPS and throughput are constrained by both machine configuration and storage capacity. A workload already peaking near 25,000 read IOPS should not be placed on the smallest compute option in the list when the objective is to maximize read performance and maintain appropriate headroom.</p><p>B is incorrect because 16 vCPUs provide a much better IOPS ceiling than 4 vCPUs, but only 200 GB of SSD storage provides significantly less storage performance headroom than the much larger SSD configuration in option C. Cloud SQL determines SSD performance from both the machine type and configured storage capacity. Therefore, selecting sufficient vCPU without sufficient storage capacity can still prevent the workload from realizing the maximum available disk performance.</p><blockquote><strong><em>C is correct because a 16-vCPU Cloud SQL for SQL Server configuration is in the range that supports up to approximately 25,000 read IOPS for the relevant SSD-backed machine family, and the 4 TB SSD option provides the largest storage-performance envelope among the answer choices. The source workload already peaks at about 25,000 read IOPS, so the destination needs both enough compute to reach that IOPS class and enough SSD capacity to avoid storage size becoming the limiting factor. SQL Server 2019 Standard also preserves the Standard database edition rather than unnecessarily changing to Enterprise solely for disk performance. In an actual migration, you should still load-test the final Cloud SQL edition, machine series, storage option, throughput, and working-set behavior before production cutover.</em></strong></blockquote><p>D is incorrect because changing the database engine edition from SQL Server Standard to Enterprise is not required merely to obtain the requested storage performance, and the 500 GB SSD is materially smaller than the 4 TB option. Database edition licensing features are separate from Cloud SQL storage IOPS. The larger SSD configuration in option C is the more relevant choice for a workload whose sizing requirement is explicitly driven by read IOPS.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/sqlserver/storage-options-overview">docs.cloud.google.com/sql/docs/sqlserver/storage-options-overview</a></p><p><a href="http://docs.cloud.google.com/sql/docs/sqlserver/machine-series-overview">docs.cloud.google.com/sql/docs/sqlserver/machine-series-overview</a></p><p><a href="http://docs.cloud.google.com/sql/docs/sqlserver/editions-intro">docs.cloud.google.com/sql/docs/sqlserver/editions-intro</a></p><p><strong>Q8.</strong></p><p><strong>At Nexalith Payments, you are designing a payment-processing application on Google Cloud. The application must continue serving requests without user disruption if a regional failure occurs. Data at rest must be protected using AES-256 encryption, and your organization must control the location and lifecycle of the encryption key. What should you do?</strong></p><p><strong>A. Use Spanner in a multi-region configuration with a customer-managed encryption key (CMEK).</strong></p><p><strong>B. Use Spanner in a multi-region configuration with Google default encryption.</strong></p><p><strong>C. Use Cloud SQL with a customer-managed encryption key (CMEK).</strong></p><p><strong>D. Use Bigtable with Google default encryption.</strong></p><blockquote><strong><em>A is correct because Spanner multi-region configurations are designed to maintain availability through a regional failure and provide up to 99.999% availability in the applicable Spanner edition. Spanner supports CMEK through Cloud KMS, giving the organization control over key location, rotation, access permissions, disablement, destruction, and auditing. Cloud KMS symmetric encryption keys use the GOOGLE_SYMMETRIC_ENCRYPTION algorithm, which uses AES-256 in Galois/Counter Mode. For current Spanner deployments, dual-region and multi-region configurations require the appropriate Enterprise Plus edition. The exact CMEK topology must also follow Spanner key-location rules, including regional-key requirements where multiple regional keys are used.</em></strong></blockquote><p>B is incorrect because a multi-region Spanner configuration satisfies the regional-failure and availability requirements, and Google default encryption protects Spanner data at rest, but Google manages the encryption keys. The question explicitly requires the organization to control where the encryption key is stored and to manage its lifecycle. That requirement calls for customer-managed encryption keys rather than Google-managed default encryption.</p><p>C is incorrect because Cloud SQL supports CMEK and can therefore satisfy the customer-controlled encryption-key requirement, but a standard Cloud SQL deployment does not provide the same transparent multi-region database availability model as Spanner. Cloud SQL high availability is regional, while cross-region read replicas or designated disaster-recovery replicas can support regional DR with additional failover and application considerations. The requirement to continue serving through a regional failure without user disruption aligns more directly with a multi-region Spanner architecture.</p><p>D is incorrect because Bigtable can be configured with replication across clusters and is highly scalable, but Google default encryption does not satisfy the requirement for the organization to control encryption-key location and lifecycle. In addition, the scenario describes a payment-processing workload for which strongly consistent relational transactions are commonly important. Spanner provides a more direct combination of transactional semantics, multi-region resilience, and CMEK control.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/instance-configurations">docs.cloud.google.com/spanner/docs/instance-configurations</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/cmek">docs.cloud.google.com/spanner/docs/cmek</a></p><p><a href="http://docs.cloud.google.com/kms/docs/algorithms">docs.cloud.google.com/kms/docs/algorithms</a></p><p><strong>Q9.</strong></p><p><strong>At Kestrelune Services, you are migrating an on-premises application to Google Cloud. The application requires a highly available PostgreSQL database for business-critical functions. The disaster-recovery policy requires both recovery time objective (RTO) and recovery point objective (RPO) to be within 30 minutes of a failure. You want to use a Google Cloud managed service and maximize uptime. What should you do?</strong></p><p><strong>A. Deploy Cloud SQL for PostgreSQL as a standalone regional instance. Create a read replica in another zone in the same region and another read replica in a different region for disaster recovery.</strong></p><p><strong>B. Deploy Cloud SQL for PostgreSQL with high availability enabled. Take periodic backups and restore a backup into a new Cloud SQL for PostgreSQL instance in another region during a disaster.</strong></p><p><strong>C. Deploy Cloud SQL for PostgreSQL with high availability enabled. Create a cross-region read replica in the disaster-recovery region and promote that replica when regional disaster recovery is required.</strong></p><p><strong>D. Migrate the PostgreSQL database to a multi-region Spanner database so that a single-region outage does not affect the application. Convert the schema and refactor the application for Spanner.</strong></p><p>A is incorrect because ordinary read replicas do not provide the same automatic zonal failover capability as a Cloud SQL high-availability configuration. A standalone primary can remain unavailable during a zonal failure until an administrator performs recovery or promotes a replica. The additional cross-region replica helps with disaster recovery, but the absence of managed HA for the primary reduces uptime during the more common zonal or instance-level failure scenarios.</p><p>B is incorrect because enabling Cloud SQL HA correctly protects the primary region from a zonal failure, but relying on periodic backups for regional disaster recovery generally produces a longer RTO and coarser RPO. During a regional outage, the team would need to locate a usable backup, create or select a destination instance in another region, restore the database, validate it, and redirect the application. That is much slower than keeping an asynchronously replicated cross-region destination ready for promotion.</p><blockquote><strong><em>C is correct because Cloud SQL HA provides a primary and standby in separate zones of the same region for managed zonal failover, while a cross-region read replica continuously receives changes asynchronously in a second region. If the primary region is lost, the cross-region replica can be promoted and the application can be redirected to it. This architecture is the standard Cloud SQL DR pattern for combining zonal HA with regional recovery. Because cross-region replication is asynchronous, the exact RPO depends on replication lag, so the team must monitor lag and size the replica appropriately to keep the required RPO within 30 minutes. Current Cloud SQL Enterprise Plus designs can also use the advanced disaster-recovery capabilities for eligible designated DR replicas.</em></strong></blockquote><p>D is incorrect because multi-region Spanner can provide stronger transparent regional resilience, but migrating an existing PostgreSQL application to Spanner is not a simple database relocation. It can require schema redesign, SQL and transaction changes, application refactoring, testing, and operational changes. The question asks for a managed PostgreSQL solution and gives a 30-minute DR target that Cloud SQL HA plus a cross-region replica can satisfy when properly designed and monitored.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/high-availability">docs.cloud.google.com/sql/docs/postgres/high-availability</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/intro-to-cloud-sql-disaster-recovery">docs.cloud.google.com/sql/docs/postgres/intro-to-cloud-sql-disaster-recovery</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/replication/cross-region-replicas">docs.cloud.google.com/sql/docs/postgres/replication/cross-region-replicas</a></p><p><strong>Q10.</strong></p><p><strong>At Velmora Systems, you are running a transactional application on Cloud SQL for PostgreSQL in Google Cloud. The database uses a high-availability configuration within one region. You encounter a logical data problem and want to restore the database to the last known pristine state from approximately two hours ago. What should you do?</strong></p><p><strong>A. Create a clone database from a read replica database, and restore the clone in the same region.</strong></p><p><strong>B. Create a clone database from a read replica database, and restore the clone into a different zone.</strong></p><p><strong>C. Use Cloud SQL point-in-time recovery (PITR). Restore the database to its state from two hours ago in a new Cloud SQL instance.</strong></p><p><strong>D. Use the Cloud SQL database import feature. Import last week’s dump file from Cloud Storage.</strong></p><p>A is incorrect because cloning a read replica does not provide a precise recovery point immediately before the logical data problem. A clone represents a copy of database state associated with the source at the cloning point rather than a time-travel operation to an arbitrary earlier timestamp. If the unwanted changes have already replicated to the read replica, cloning that replica preserves the same logical corruption. The requirement is to return to the last known pristine state from about two hours earlier, so a time-based recovery mechanism is required.</p><p>B is incorrect because changing the zone of a clone does not solve logical data corruption. Zonal placement is primarily an availability and locality concern, while the recovery problem is about restoring an earlier database state. A replica can already contain the corrupted changes because read-replica replication propagates committed database changes. Creating a clone in another zone therefore does not inherently recover the clean version of the data.</p><blockquote><strong><em>C is correct because Cloud SQL PITR is specifically designed to recover a database to a selected timestamp within the configured transaction-log retention window. For Cloud SQL for PostgreSQL, PITR reconstructs the database from backup and transaction-log information and always creates a new instance rather than overwriting the existing source instance. Restoring to a timestamp immediately before the data issue minimizes data loss compared with restoring an older backup or dump and allows the team to validate the recovered instance before redirecting production traffic.</em></strong></blockquote><p>D is incorrect because importing last week’s dump would discard all legitimate transactions committed between the dump and the incident. A logical dump can be useful when PITR is unavailable or when a particular archival copy is required, but it provides a much coarser recovery point in this scenario. Because PITR is available and the desired recovery point is only about two hours old, PITR minimizes data loss and operational disruption.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/pitr">docs.cloud.google.com/sql/docs/postgres/backup-recovery/pitr</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/restore">docs.cloud.google.com/sql/docs/postgres/backup-recovery/restore</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/backups">docs.cloud.google.com/sql/docs/postgres/backup-recovery/backups</a></p><p><strong>Q11.</strong></p><p><strong>At Quorilume Analytics, you are configuring a new application that reads an existing Spanner database to gather statistics for a dashboard. The application does not need to modify data. You want to follow Google-recommended IAM practices and apply least privilege. What should you do?</strong></p><p><strong>A. Reuse the existing service account that populates the database.</strong></p><p><strong>B. Create a new service account, and grant it the Spanner Database Admin role.</strong></p><p><strong>C. Create a new service account, and grant it the Spanner Database Reader role (roles/spanner.databaseReader).</strong></p><p><strong>D. Create a new service account, and grant it only the spanner.databases.select permission directly.</strong></p><p>A is incorrect because reusing a service account that writes data gives the read-only dashboard application unnecessary privileges and couples two independent workloads to the same identity. If the service account is compromised, both the data-population workload and the dashboard workload are affected. Google Cloud IAM best practices favor separate workload identities and permissions that match each application’s actual responsibilities.</p><p>B is incorrect because roles/spanner.databaseAdmin grants far more authority than a dashboard reader requires. A database administrator can perform administrative actions and read or write database data, creating unnecessary risk for an application whose only responsibility is to query statistics. Granting broad administrative access violates the principle of least privilege.</p><blockquote><strong><em>C is correct because roles/spanner.databaseReader is the predefined read-only role designed for principals that need to read and query Spanner data without modifying it. The role includes permissions such as spanner.databases.read, spanner.databases.select, read-only transaction permissions, and session access needed to execute queries. Creating a dedicated service account also isolates the dashboard workload from the service account that populates the database and makes auditing and permission management clearer.</em></strong></blockquote><p>D is incorrect because IAM permissions are granted through IAM roles rather than by attaching an individual permission directly to a principal. In addition, a practical Spanner client needs more than spanner.databases.select alone, including permissions for sessions and read-only transactions. The predefined databaseReader role packages the required read-only permissions and is the appropriate least-privilege choice.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/iam">docs.cloud.google.com/spanner/docs/iam</a></p><p><a href="http://docs.cloud.google.com/iam/docs/roles-permissions/spanner">docs.cloud.google.com/iam/docs/roles-permissions/spanner</a></p><p><strong>Q12.</strong></p><p><strong>At Lumivora Financial, your company uses Spanner for a fully managed transactional database. Traffic grows gradually over time and also experiences occasional large holiday spikes. You plan to use the open-source Spanner Autoscaler tool and need a scaling method that automatically adjusts compute capacity according to observed utilization rather than only at predetermined times. What should you do?</strong></p><p><strong>A. Use linear scaling with the Spanner Autoscaler tool.</strong></p><p><strong>B. Use direct scaling with the Spanner Autoscaler tool.</strong></p><p><strong>C. Increase Spanner compute capacity manually on a periodic maintenance schedule.</strong></p><p><strong>D. Set alerts for Spanner utilization thresholds and then manually schedule capacity increases.</strong></p><blockquote><strong><em>A is correct because the open-source Spanner Autoscaler tool’s linear method calculates the required capacity from the ratio of observed utilization to the configured utilization threshold and can add or remove multiple nodes or processing units in one scaling action. Google documents linear scaling as suitable for workloads that change gradually or have a few large peaks, which closely matches steady annual growth with occasional holiday spikes. In current Spanner Enterprise and Enterprise Plus editions, managed autoscaling is also available and is generally the simpler Google-managed option when its capabilities fit the workload, but among the scaling methods listed here, linear scaling is the best match.</em></strong></blockquote><p>B is incorrect because direct scaling is intended primarily for workloads whose higher capacity requirement is known in advance and occurs on a predictable schedule, such as a recurring batch job. It immediately scales to a predefined capacity at the scheduled time. Holiday spikes that vary with consumer demand and long-term growth are better handled by utilization-driven scaling, while direct scaling can be combined with another method when a specific event has a known start time.</p><p>C is incorrect because manual periodic scaling requires an administrator to predict demand and change capacity, which removes the main benefit of autoscaling. Spanner compute capacity can be changed without a traditional database maintenance window, so tying capacity changes to scheduled maintenance is also unnecessary.</p><p>D is incorrect because alerts are useful for observability, but an alert followed by a manually scheduled resize is not an automated scaling architecture. It increases operational workload and can react too slowly to sudden demand. The Autoscaler tool or Spanner managed autoscaling can adjust capacity automatically based on utilization and storage targets.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/autoscaler-tool-overview">docs.cloud.google.com/spanner/docs/autoscaler-tool-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/autoscaling-overview">docs.cloud.google.com/spanner/docs/autoscaling-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/managed-autoscaler">docs.cloud.google.com/spanner/docs/managed-autoscaler</a></p><p><strong>Q13.</strong></p><p><strong>At Cerynexa Technologies, your company is developing a new global relational transactional application that must be ACID-compliant, use SQL semantics, and provide a 99.999% availability SLA across regions. You are responsible for selecting the appropriate Google Cloud database. What should you do?</strong></p><p><strong>A. Use Firestore.</strong></p><p><strong>B. Use Spanner with an Enterprise Plus dual-region or multi-region configuration.</strong></p><p><strong>C. Use Cloud SQL.</strong></p><p><strong>D. Use Bigtable.</strong></p><p>A is incorrect because current Firestore multi-region deployments do provide ACID transactions, strong consistency, and a 99.999% availability SLA, so Firestore is a strong option for document-oriented global applications. However, Firestore is a NoSQL document database rather than a relational SQL database. The clarified requirement calls for global relational transactions with SQL semantics, which is the area Spanner is designed to serve.</p><blockquote><strong><em>B is correct because Spanner provides relational schemas, SQL, ACID transactions, external consistency, and horizontal scalability. Current Spanner Enterprise Plus supports dual-region and multi-region configurations with availability up to 99.999%, and those configurations maintain service through regional failures using synchronous replication and automatic leader management. This directly matches the requirement for a globally distributed relational transactional database with five-nines availability.</em></strong></blockquote><p>C is incorrect because Cloud SQL is an excellent managed relational service for MySQL, PostgreSQL, and SQL Server, but Cloud SQL high availability is regional rather than a single synchronously consistent multi-region database architecture. Cross-region read replicas provide disaster recovery and read locality through asynchronous replication, so Cloud SQL does not provide the same transparent global relational availability model as Spanner.</p><p>D is incorrect because Bigtable provides massive scalability, low latency, and multi-cluster replication, but it is a NoSQL wide-column database rather than a relational SQL transaction-processing service. Its transaction model does not provide the same general-purpose multi-row relational ACID semantics required by the application.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/instance-configurations">docs.cloud.google.com/spanner/docs/instance-configurations</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/editions-overview">docs.cloud.google.com/spanner/docs/editions-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/true-time-external-consistency">docs.cloud.google.com/spanner/docs/true-time-external-consistency</a></p><p><strong>Q14.</strong></p><p><strong>At Novarixa Media, you have a large Cloud SQL for PostgreSQL instance. The database is not mission-critical, and you want to minimize operational costs while retaining automated backups. What should you do to lower backup-storage cost?</strong></p><p><strong>A. Set the automated backups to occur every other day to lower the frequency of backups.</strong></p><p><strong>B. Change the storage tier of the automated backups from solid-state drive (SSD) to hard disk drive (HDD).</strong></p><p><strong>C. Select a different region to store your backups.</strong></p><p><strong>D. Reduce the number of automated backups that are retained to two.</strong></p><p>A is incorrect because standard Cloud SQL automated backups are managed by the Cloud SQL backup service and are not configured as an arbitrary every-other-day schedule in the traditional standard-backup model. Current enhanced backups can support more granular schedules, but merely reducing schedule frequency is not the direct standard-backup retention control represented by this question. The straightforward way to reduce the amount of backup data retained is to shorten the retention policy while keeping automated recovery points.</p><p>B is incorrect because Cloud SQL does not expose an SSD-versus-HDD storage-tier selector for an individual automated backup in the manner described. Backup storage is managed by the service according to the selected Cloud SQL backup option and storage location. You cannot lower backup cost by editing each automated backup to use HDD storage.</p><p>C is incorrect because changing backup location is primarily a residency, regulatory, resilience, and disaster-recovery decision rather than the first cost-optimization mechanism for this scenario. Backup cost can depend on storage location, but moving backups solely to reduce price may weaken locality or regulatory requirements and does not reduce the amount of retained backup data. For a non-mission-critical database, shortening retention is the more direct lever.</p><blockquote><strong><em>D is correct because Cloud SQL lets you configure how long or how many automated standard backups are retained, subject to the capabilities and limits of the selected backup option and management interface. Shorter retention reduces the amount of historical backup data stored and therefore reduces backup-storage cost. Because the database is explicitly not mission-critical, retaining only a very small number of recovery points can be appropriate if that still satisfies the organization’s recovery requirements. In a real production policy, verify that the chosen retention value is supported by the selected standard or enhanced backup configuration and is sufficient for the required RPO and recovery window.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/backups">docs.cloud.google.com/sql/docs/postgres/backup-recovery/backups</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/backup-options">docs.cloud.google.com/sql/docs/postgres/backup-recovery/backup-options</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/manage-standard-backups">docs.cloud.google.com/sql/docs/postgres/backup-recovery/manage-standard-backups</a></p><p><strong>Q15.</strong></p><p><strong>At Pelorune Analytics, you use Python scripts to generate weekly SQL reports that assess database health and determine whether tables need maintenance or updated statistics. You want to automate the report while minimizing operational cost and administration. What should you do?</strong></p><p><strong>A. Create a VM in Compute Engine, and run a cron job.</strong></p><p><strong>B. Create a Managed Service for Apache Airflow environment, and create a directed acyclic graph (DAG).</strong></p><p><strong>C. Create a Cloud Run function, and invoke it on a weekly schedule with Cloud Scheduler.</strong></p><p><strong>D. Create a Cloud Run function, and call the function from a Cloud Tasks queue.</strong></p><p>A is incorrect because a Compute Engine VM would need to remain provisioned, patched, secured, monitored, and paid for even though the workload runs only once per week. A cron job can schedule the script, but maintaining a VM for a short periodic task creates unnecessary cost and operational overhead compared with a serverless execution model.</p><p>B is incorrect because Managed Service for Apache Airflow is intended for orchestrating workflows with multiple tasks, dependencies, retries, branching, data pipelines, and more complex schedules. It can certainly run a weekly job, but provisioning a managed Airflow environment solely for one Python reporting script is much heavier and more expensive than necessary.</p><blockquote><strong><em>C is correct because a Cloud Run function provides serverless execution for the Python code and incurs compute usage primarily when invoked. Cloud Scheduler is purpose-built for cron-style schedules and can invoke an HTTP Cloud Run function directly or publish to Pub/Sub for an event-driven function. A weekly function therefore avoids continuously running infrastructure while providing a simple managed schedule with minimal administration.</em></strong></blockquote><p>D is incorrect because Cloud Tasks is primarily a task-dispatch and rate-control service for asynchronous work. It can schedule individual tasks for future execution, but it is not the natural recurring cron scheduler for a weekly job. Cloud Scheduler is the direct managed service for recurring schedules and can invoke the function without requiring a task queue.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/scheduler/docs/tut-gcf-http">docs.cloud.google.com/scheduler/docs/tut-gcf-http</a></p><p><a href="http://docs.cloud.google.com/scheduler/docs/tut-gcf-pub-sub">docs.cloud.google.com/scheduler/docs/tut-gcf-pub-sub</a></p><p><a href="http://docs.cloud.google.com/run/docs/function-triggers">docs.cloud.google.com/run/docs/function-triggers</a></p><p><strong>Q16.</strong></p><p><strong>At Velmora Systems, your project uses Bigtable to store sensitive data that must never be accessed from the public internet, even if a requester possesses a valid service account key. You need to secure access to this data. What should you do?</strong></p><p><strong>A. Use Identity and Access Management (IAM) for Bigtable access control.</strong></p><p><strong>B. Use VPC Service Controls to create a trusted service perimeter around Bigtable.</strong></p><p><strong>C. Use customer-managed encryption keys (CMEK).</strong></p><p><strong>D. Use Cloud Armor to add IP addresses to an allowlist.</strong></p><p>A is incorrect because IAM determines which authenticated principals are authorized to perform Bigtable operations, but IAM by itself does not impose a network or service-perimeter boundary. A principal with valid credentials and sufficient IAM permissions could still call the Bigtable API from an external network unless another control restricts the request context. IAM remains necessary for least-privilege authorization, but it does not independently satisfy the requirement that the data must not be accessible from the public internet under any circumstances.</p><blockquote><strong><em>B is correct because VPC Service Controls is designed to create security perimeters around supported Google Cloud managed services, including Bigtable, and to reduce data-exfiltration risk by controlling which identities, projects, networks, and request contexts can access protected resources. Internet-originated access to resources inside a service perimeter is denied unless it explicitly satisfies configured ingress rules or access levels. This provides a network-context and service-boundary control in addition to IAM, so possession of a valid service account key alone is not sufficient to bypass the perimeter.</em></strong></blockquote><p>C is incorrect because CMEK controls how Bigtable data is encrypted at rest and gives the organization control over key permissions, rotation, disablement, and destruction. Encryption at rest does not restrict the network origin from which an authorized principal can call the Bigtable API. A requester who is otherwise authorized can still access decrypted data through the service, so CMEK does not implement the required public-internet access restriction.</p><p>D is incorrect because Cloud Armor protects supported internet-facing applications and services, primarily through policies applied to supported load-balancing architectures. Bigtable is a Google-managed database API and is not protected by placing a Cloud Armor IP allowlist directly in front of the Bigtable service. VPC Service Controls is the Google Cloud control specifically intended to establish a service perimeter around Bigtable and restrict access based on trusted contexts.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/vpc-service-controls/docs/overview">docs.cloud.google.com/vpc-service-controls/docs/overview</a></p><p><a href="http://docs.cloud.google.com/vpc-service-controls/docs/supported-products">docs.cloud.google.com/vpc-service-controls/docs/supported-products</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/access-control">docs.cloud.google.com/bigtable/docs/access-control</a></p><p><strong>Q17.</strong></p><p><strong>At Vantrelix Financial, an internal audit found that an existing Cloud SQL for MySQL instance is not configured for high availability. You want to follow Google-recommended practices and enable HA on the existing instance. What should you do?</strong></p><p><strong>A. Create a new Cloud SQL for MySQL instance with HA enabled, and migrate the data by using export and import.</strong></p><p><strong>B. Create a new Cloud SQL for MySQL instance with HA enabled, and use Cloud Data Fusion to migrate the data.</strong></p><p><strong>C. Use gcloud sql instances patch on the existing instance to set the availability type to REGIONAL and enable the required backup and binary-log settings.</strong></p><p><strong>D. Shut down the existing Cloud SQL for MySQL instance, and enable HA while it is stopped.</strong></p><p>A is incorrect because Cloud SQL supports converting an existing compatible instance from zonal availability to regional high availability in place. Creating a new instance, exporting the database, importing it, validating it, and changing application connections adds unnecessary downtime and operational risk. A migration is not required simply to enable the built-in HA configuration.</p><p>B is incorrect because Cloud Data Fusion is a managed data-integration service for building ETL and ELT pipelines, not the normal mechanism for converting a Cloud SQL instance to high availability. Using it would add an unnecessary data-copy pipeline and still require a new destination instance, even though the existing Cloud SQL instance can be patched directly.</p><blockquote><strong><em>C is correct because Cloud SQL explicitly supports enabling high availability on an existing MySQL instance. The current gcloud command uses gcloud sql instances patch with — availability-type REGIONAL, and MySQL HA also requires automated backups and binary logging to be enabled. Reconfiguring an existing instance can restart it and can take longer for large or heavily loaded instances, so the change should be scheduled appropriately, but it is the direct supported approach and avoids building a replacement database.</em></strong></blockquote><p>D is incorrect because administrators do not need to shut down the instance manually before enabling high availability. Cloud SQL manages the required restart as part of the configuration change. Manually stopping service first adds unnecessary downtime and does not provide any advantage over the supported in-place patch operation.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/configure-ha">docs.cloud.google.com/sql/docs/mysql/configure-ha</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/high-availability">docs.cloud.google.com/sql/docs/mysql/high-availability</a></p><p><a href="http://cloud.google.com/sdk/gcloud/reference/sql/instances/patch">cloud.google.com/sdk/gcloud/reference/sql/instances/patch</a></p><p><strong>Q18.</strong></p><p><strong>At Orbinexa Commerce, your ecommerce application connects to Cloud SQL for SQL Server and is expected to experience heavy holiday-weekend traffic. You want to follow Google-recommended practices and receive text-message notifications when CPU or memory metrics show the first signs of a problem. What should you do?</strong></p><p><strong>A. Use a Cloud Run function to pull CPU and memory metrics from Cloud SQL and call a custom service that sends alerts.</strong></p><p><strong>B. Use Error Reporting to monitor CPU and memory metrics and configure SMS notification channels.</strong></p><p><strong>C. Use Cloud Logging to create a log sink for CPU and memory metrics and send the sink output to Pub/Sub.</strong></p><p><strong>D. Use Cloud Monitoring to create alerting policies for Cloud SQL CPU and memory metrics and configure SMS notification channels.</strong></p><p>A is incorrect because custom code is unnecessary. Cloud SQL exports operational metrics directly to Cloud Monitoring, which already provides threshold-based alerting, incident management, and multiple notification-channel types. A custom polling function would require the team to implement scheduling, metric retrieval, thresholds, retries, notification delivery, and monitoring of the monitoring system itself.</p><p>B is incorrect because Error Reporting groups and analyzes application error events and exceptions. CPU and memory utilization are time-series infrastructure metrics rather than error events, so Error Reporting is not the correct service for threshold monitoring of Cloud SQL resource utilization.</p><p>C is incorrect because Cloud Logging manages log records, while CPU and memory utilization are Cloud Monitoring metrics. A log sink routes log entries to supported destinations but is not the normal mechanism for evaluating metric thresholds. Converting the problem into a logging and Pub/Sub pipeline would add unnecessary complexity and would still require a consumer to decide when to generate alerts.</p><blockquote><strong><em>D is correct because Cloud Monitoring receives Cloud SQL resource metrics and provides native alerting policies that can evaluate thresholds on CPU, memory, disk, connection, and other time-series metrics. SMS is a supported Cloud Monitoring notification channel, so a verified phone number can receive incident notifications when the alerting policy fires. Current Google guidance notes that SMS is not considered fully reliable and might not be available in every region, so production alerting should also include a redundant channel such as email or Pub/Sub.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/sqlserver/monitor-instance">docs.cloud.google.com/sql/docs/sqlserver/monitor-instance</a></p><p><a href="http://docs.cloud.google.com/monitoring/alerts">docs.cloud.google.com/monitoring/alerts</a></p><p><a href="http://docs.cloud.google.com/monitoring/support/notification-options">docs.cloud.google.com/monitoring/support/notification-options</a></p><p><strong>Q19.</strong></p><p><strong>At Lumivora Health, your organization works with sensitive data and requires direct control over the encryption keys used for data at rest. The data is stored in a Cloud SQL for MySQL database. You need to ensure that stored database data is protected with organization-controlled keys. What should you do?</strong></p><p><strong>A. Export the database periodically to a Cloud Storage bucket protected by customer-supplied encryption keys.</strong></p><p><strong>B. Use the Cloud SQL Auth Proxy.</strong></p><p><strong>C. Connect to Cloud SQL by using SSL/TLS encryption.</strong></p><p><strong>D. Configure Cloud SQL to use customer-managed encryption keys (CMEK) in Cloud KMS.</strong></p><p>A is incorrect because encrypting exported copies in Cloud Storage does not change how the active Cloud SQL instance stores and encrypts its database files, logs, temporary data, and managed backups. The requirement concerns the database data at rest in Cloud SQL itself. In addition, Cloud SQL supports CMEK but does not support customer-supplied encryption keys as its database at-rest encryption mechanism.</p><p>B is incorrect because the Cloud SQL Auth Proxy is a connectivity and authentication component. It can establish authorized encrypted connections to Cloud SQL, but it does not control the key encryption key that protects the database’s stored data. Using the proxy therefore addresses secure connectivity rather than customer-managed encryption at rest.</p><p>C is incorrect because SSL/TLS protects data in transit between the client and Cloud SQL. It does not determine which key protects database storage. Transport encryption is an important complementary control, but it cannot satisfy a requirement for organization-controlled encryption keys for stored data.</p><blockquote><strong><em>D is correct because Cloud SQL supports customer-managed encryption keys through Cloud KMS. With CMEK, the organization creates and manages the key encryption key and controls its location, IAM permissions, rotation, disablement, destruction, and auditing. Cloud SQL uses that key to protect the data encryption keys associated with supported Cloud SQL data at rest. The Cloud KMS key must be in a location compatible with the Cloud SQL instance, and administrators must protect key availability because disabling or destroying the key can make the protected Cloud SQL resources inaccessible.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/cmek">docs.cloud.google.com/sql/docs/mysql/cmek</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/configure-cmek">docs.cloud.google.com/sql/docs/mysql/configure-cmek</a></p><p><a href="http://docs.cloud.google.com/kms/docs">docs.cloud.google.com/kms/docs</a></p><p><strong>Q20.</strong></p><p><strong>At Orbinexa Games, you are designing a new gaming application that uses a highly transactional relational database for player authentication and inventory data. You plan to launch the game in multiple regions and need strong transactional consistency across the global player base. What should you do?</strong></p><p><strong>A. Use Spanner in a multi-region configuration.</strong></p><p><strong>B. Use Bigtable with clusters in multiple regions.</strong></p><p><strong>C. Use BigQuery.</strong></p><p><strong>D. Use Cloud SQL with a regional read replica.</strong></p><blockquote><strong><em>A is correct because Spanner is a fully managed relational database that provides SQL, ACID read-write transactions, strong external consistency, horizontal scale, and synchronous replication across supported multi-region configurations. These properties are particularly important for gaming inventory, authentication, currency, and entitlement data, where duplicate items or inconsistent transactions can lead to exploits. Current Spanner guidance for gaming specifically highlights the benefit of a single globally consistent transactional database compared with application-managed regional sharding. Multi-region Spanner therefore matches both the relational and global consistency requirements.</em></strong></blockquote><p>B is incorrect because Bigtable offers massive scale and multi-cluster replication, but it is a NoSQL wide-column database. Multi-cluster routing uses eventual consistency, and Bigtable does not provide arbitrary multi-row relational ACID transactions across an inventory schema. It is excellent for telemetry, leaderboards, event data, and key-based workloads but not the best fit for the transactional relational system described here.</p><p>C is incorrect because BigQuery is a serverless analytical data warehouse rather than an operational transaction-processing database. It is optimized for scanning and aggregating large datasets, not for serving low-latency player authentication and inventory transactions at application request rates.</p><p>D is incorrect because Cloud SQL is a managed relational database and a read replica can improve read capacity or locality, but a read replica is asynchronously updated and is not a globally writable, strongly consistent multi-region primary. Writes still go to one primary region, and regional failure requires a promotion or disaster-recovery process. Spanner provides the transparent global transactional model required by the scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/best-practices-gaming-database">docs.cloud.google.com/spanner/docs/best-practices-gaming-database</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/instance-configurations">docs.cloud.google.com/spanner/docs/instance-configurations</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/true-time-external-consistency">docs.cloud.google.com/spanner/docs/true-time-external-consistency</a></p><p><strong>Q21.</strong></p><p><strong>At Merovexa Insights, you are building a BigQuery data warehouse. Several source MySQL databases remain on-premises. You need low-latency replication into BigQuery for analytics and want to use a managed solution that is straightforward to operate. Among the following choices, what should you do?</strong></p><p><strong>A. Use Datastream to connect to the on-premises MySQL databases and create streams to Cloud Storage, then use a managed Dataflow Datastream-to-BigQuery pipeline to apply the changes to BigQuery.</strong></p><p><strong>B. Use Cloud Data Fusion and scheduled workflows to extract MySQL data, transform it, and load it into BigQuery.</strong></p><p><strong>C. Use Database Migration Service to replicate the databases to Cloud SQL for MySQL, and then create BigQuery federated queries over the Cloud SQL replicas.</strong></p><p><strong>D. Periodically create database extracts, upload them to Cloud Storage, and manually merge each batch into BigQuery.</strong></p><blockquote><strong><em>A is correct because Datastream is a serverless change data capture and replication service designed to capture ongoing changes from supported MySQL sources with low latency. The documented Datastream and Dataflow analytics pattern writes change events to Cloud Storage and uses a managed Dataflow template to materialize those changes in BigQuery. This provides continuous CDC without the organization building its own binlog reader, checkpointing, retry, schema, and merge logic. Current Datastream also supports BigQuery as a direct destination by using BigQuery CDC and the Storage Write API, which is generally the simpler current design for a new deployment, but option A remains the strongest managed low-latency approach among the choices presented.</em></strong></blockquote><p>B is incorrect because Cloud Data Fusion is a capable managed data-integration platform and can ingest data from MySQL, but a scheduled batch workflow is not the lowest-latency approach for keeping BigQuery synchronized with frequently changing transactional databases. It is more appropriate when complex ETL transformations and scheduled batch orchestration are the dominant requirements.</p><p>C is incorrect because Database Migration Service is designed to migrate databases into supported managed database destinations rather than to serve as the normal ongoing analytics ingestion layer for BigQuery. Adding a Cloud SQL copy and then querying it through federation introduces an unnecessary database tier, still exposes an operational database to analytical query pressure, and does not provide the same native BigQuery ingestion model as Datastream.</p><p>D is incorrect because periodic manual extracts create batch latency and require custom scheduling, transfer, merge, failure handling, and incremental-change logic. The requirement specifically calls for a managed, low-latency solution, which is exactly the use case for Datastream CDC.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/datastream/docs/configure-bigquery-destination">docs.cloud.google.com/datastream/docs/configure-bigquery-destination</a></p><p><a href="http://docs.cloud.google.com/datastream/docs/implementing-datastream-dataflow-analytics">docs.cloud.google.com/datastream/docs/implementing-datastream-dataflow-analytics</a></p><p><a href="http://docs.cloud.google.com/datastream/docs/destinations">docs.cloud.google.com/datastream/docs/destinations</a></p><p><strong>Q22.</strong></p><p><strong>At Avenora Games, your company is launching a gaming application that uses Firestore. You need an easy-to-manage and cost-effective way to schedule recurring Firestore managed exports. What should you do?</strong></p><p><strong>A. Create a Compute Engine VM, and configure a cron job that runs the gcloud firestore export command.</strong></p><p><strong>B. Use Dataflow to build a custom pipeline that extracts Firestore data, transforms it, and writes it to Cloud Storage on a schedule.</strong></p><p><strong>C. Use Cloud Scheduler to trigger a Cloud Run function that starts the Firestore managed export process.</strong></p><p><strong>D. Use the Firebase Admin SDK alone to programmatically schedule and manage exports.</strong></p><p>A is incorrect because a Compute Engine VM can run cron and invoke a Firestore export command, but the organization must provision, patch, secure, monitor, and pay for that VM continuously even though the job runs only periodically. This creates unnecessary infrastructure and operational overhead for a simple scheduled control-plane task.</p><p>B is incorrect because Dataflow is intended for distributed batch and stream data processing. A custom pipeline would be much more complex than necessary simply to start Firestore’s existing managed export operation. It would also require the organization to design and maintain extraction, transformation, scheduling, and storage behavior that the managed Firestore export feature already provides.</p><blockquote><strong><em>C is correct because current Google Cloud documentation explicitly recommends using Cloud Scheduler together with Cloud Run functions to schedule Firestore data exports. Cloud Scheduler provides the recurring cron schedule, while the function performs the lightweight API call that starts a managed Firestore export to Cloud Storage. The function runs only when triggered, so the design has little idle infrastructure cost and is straightforward to maintain. Exported documents incur the documented Firestore export read charges, so the schedule and retention policy should still be chosen carefully.</em></strong></blockquote><p>D is incorrect because the Firebase Admin SDK can be useful for privileged server-side application logic, but it is not itself a managed recurring scheduling service. The team would still need somewhere to run the scheduling logic reliably. Cloud Scheduler provides the managed cron capability, and a Cloud Run function is the documented lightweight execution target for the export call.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/firestore/native/docs/solutions/schedule-export">docs.cloud.google.com/firestore/native/docs/solutions/schedule-export</a></p><p><a href="http://docs.cloud.google.com/scheduler/docs/firestore-data-exports">docs.cloud.google.com/scheduler/docs/firestore-data-exports</a></p><p><a href="http://docs.cloud.google.com/firestore/native/docs/manage-data/export-import">docs.cloud.google.com/firestore/native/docs/manage-data/export-import</a></p><p><strong>Q23.</strong></p><p><strong>At Braventra Media, you manage a 10 GB Cloud SQL for MySQL database for a noncritical application. The database is refreshed nightly from data in Cloud Storage and is read-only during the day. Regulations require the database and its backups to remain in one specified region. You want to minimize cost, and an RTO of one day is acceptable. What should you do?</strong></p><p><strong>A. Enable Cloud SQL high availability and use the default automated-backup location.</strong></p><p><strong>B. Disable high availability and create a read replica in the same zone.</strong></p><p><strong>C. Disable high availability and create a read replica in a second region.</strong></p><p><strong>D. Disable high availability, enable automated backups, and configure a custom backup location in the same required region.</strong></p><p>A is incorrect because high availability adds a standby infrastructure footprint and additional cost that is unnecessary for a noncritical workload with an acceptable RTO of one day. In addition, the default Cloud SQL standard-backup location is commonly the geographically closest multi-region rather than necessarily the same individual region as the database. That can conflict with a regulation requiring backups to remain in one specified region.</p><p>B is incorrect because a read replica adds ongoing compute and storage cost but does not provide the same managed zonal failover behavior as Cloud SQL high availability. A replica in the same zone also shares the same zonal failure domain as the primary. Given that the database is only 10 GB, refreshed nightly, and can tolerate a one-day recovery objective, an additional continuously running replica is unnecessary.</p><p>C is incorrect because a cross-region read replica adds even more cost and explicitly places a copy of the data in another region, which violates the stated residency requirement. Cross-region replicas are useful for disaster recovery and geographic read locality, but neither capability is required strongly enough to justify breaking the single-region constraint.</p><blockquote><strong><em>D is correct because disabling high availability minimizes ongoing database infrastructure cost, while automated backups provide a managed recovery mechanism appropriate for a noncritical workload with a one-day RTO. Cloud SQL lets administrators configure a custom backup location so standard backups can remain in the required region instead of using the default nearby multi-region. Cloud SQL manages the backup storage itself; the custom backup location is a Cloud SQL backup setting rather than a customer-managed Cloud Storage bucket.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/backup-recovery/manage-standard-backups">docs.cloud.google.com/sql/docs/mysql/backup-recovery/manage-standard-backups</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/backup-recovery/backup-options">docs.cloud.google.com/sql/docs/mysql/backup-recovery/backup-options</a></p><p><a href="http://docs.cloud.google.com/sql/docs/mysql/high-availability">docs.cloud.google.com/sql/docs/mysql/high-availability</a></p><p><strong>Q24.</strong></p><p><strong>At Lumorix Systems, you run Cloud SQL for PostgreSQL 13 on Cloud SQL Enterprise edition with point-in-time recovery enabled. An audit finds that the retained transaction logs used for PITR are stored on disk. Going forward, you need those PITR logs stored in Cloud Storage while preserving recoverability and avoiding loss of the currently retained recovery history. Which available option should you choose?</strong></p><p><strong>A. Clone the instance, and configure the clone with a 30-day point-in-time recovery retention period.</strong></p><p><strong>B. Change only the point-in-time recovery retention period.</strong></p><p><strong>C. Upgrade the instance in place to Cloud SQL Enterprise Plus edition.</strong></p><p><strong>D. Disable point-in-time recovery, and then enable point-in-time recovery again.</strong></p><p>A is incorrect because cloning the instance creates a separate Cloud SQL instance and does not change how the original instance stores its PITR transaction logs. A clone also creates additional cost and operational work, and changing the retention period affects how long recovery information is preserved rather than converting the transaction-log storage mechanism for the original instance.</p><p>B is incorrect because PITR retention determines how far back the database can be recovered, but it does not by itself switch transaction-log storage from disk to Cloud Storage. Retention and log-storage location are separate settings. Adjusting the retention period therefore does not satisfy the audit requirement.</p><blockquote><strong><em>C is correct because current Cloud SQL documentation states that an in-place upgrade from Cloud SQL Enterprise edition to Cloud SQL Enterprise Plus edition transitions PITR transaction-log storage to Cloud Storage and preserves the existing PITR recovery history during the transition. Enterprise Plus also provides additional availability and performance capabilities. Current Cloud SQL now also supports directly switching transaction-log storage from disk to Cloud Storage through the supported configuration interfaces without disabling PITR; if an edition upgrade were not otherwise desired, that direct switch would normally be the simpler modern approach. Among the choices provided, however, the Enterprise Plus in-place upgrade is the option that accomplishes the required transition without discarding the retained logs.</em></strong></blockquote><p>D is incorrect because disabling PITR causes the currently retained transaction logs used for point-in-time recovery to be lost. Re-enabling PITR starts building a new recovery history rather than preserving the existing one. That directly violates the requirement not to lose the current PITR logs and temporarily weakens recoverability.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/backup-recovery/configure-pitr">docs.cloud.google.com/sql/docs/postgres/backup-recovery/configure-pitr</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/upgrade-cloud-sql-instance-to-enterprise-plus-in-place">docs.cloud.google.com/sql/docs/postgres/upgrade-cloud-sql-instance-to-enterprise-plus-in-place</a></p><p><a href="http://docs.cloud.google.com/sql/docs/postgres/upgrade-overview">docs.cloud.google.com/sql/docs/postgres/upgrade-overview</a></p><p><strong>Q25.</strong></p><p><strong>At Ceryntra Applications, a production application performs continuous reads and writes against a single Bigtable cluster. You are adding a large batch-analytics workload that needs to read the same data, and the new workload must not degrade the existing application’s performance. What should you do?</strong></p><p><strong>A. Manually add more nodes to the existing Bigtable cluster before each batch job.</strong></p><p><strong>B. Add a second Bigtable cluster, and configure one multi-cluster routing application profile for both production traffic and batch analytics.</strong></p><p><strong>C. Add a second Bigtable cluster, and create separate application profiles so the production workload and batch workload are routed to different clusters.</strong></p><p><strong>D. Enable autoscaling on the existing single cluster and run both workloads through the same application profile.</strong></p><p>A is incorrect because adding nodes can increase total cluster capacity, but both the transactional application and batch analytics would still compete for the same cluster resources, tablets, and hot row ranges. Manual resizing also requires forecasting the batch workload and does not provide strong workload isolation. Capacity scaling can complement isolation, but it is not a substitute for placing the workloads on different clusters.</p><p>B is incorrect because multi-cluster routing lets Bigtable send requests to multiple eligible clusters and can provide automatic failover, but it does not guarantee that the production and batch workloads remain isolated from one another. Both workloads could be routed to the same cluster depending on routing behavior and availability, allowing the analytics workload to affect the live application. That conflicts with the explicit isolation requirement.</p><blockquote><strong><em>C is correct because Bigtable replication can maintain the same table data in multiple clusters, while separate application profiles can direct different workloads to different clusters. The production application can use a single-cluster routing profile for its serving cluster, and the analytics process can use another profile routed to the batch cluster. This isolates CPU and node capacity for the two workloads while Bigtable replication keeps the data available to both. Current Bigtable Data Boost can also offload certain analytical reads without provisioning additional nodes, but among the listed choices, separate clusters and application profiles are the appropriate isolation design.</em></strong></blockquote><p>D is incorrect because autoscaling helps a cluster adjust node capacity as utilization changes, but both workloads would still share the same cluster and potentially contend for the same tablets and row-key ranges. A sudden batch scan can affect latency before autoscaling adds sufficient capacity, and a schema hotspot cannot necessarily be solved by adding nodes. Dedicated cluster routing provides stronger workload separation.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigtable/docs/app-profiles">docs.cloud.google.com/bigtable/docs/app-profiles</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/replication-settings">docs.cloud.google.com/bigtable/docs/replication-settings</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/replication-overview">docs.cloud.google.com/bigtable/docs/replication-overview</a></p><p><strong>Q26.</strong></p><p><strong>At Vantressa Analytics, you are designing a global 24/7 real-time analytics platform that stores large volumes of versioned time-series data. The service is mission-critical, must tolerate cluster failures, and must scale through unpredictable traffic spikes without creating row-key hotspots. What should you do?</strong></p><p><strong>A. Use a single-cluster Bigtable instance with autoscaling and a well-designed time-series row key.</strong></p><p><strong>B. Use a multi-cluster Bigtable instance with autoscaling and an optimized time-series schema, placing clusters across appropriate regions and using a routing profile that supports failover.</strong></p><p><strong>C. Use a multi-cluster Bigtable instance across multiple regions with replication, but use fixed cluster sizes and leave routing and schema behavior at their defaults.</strong></p><p><strong>D. Use AlloyDB for PostgreSQL and send analytical reads to a read pool.</strong></p><p>A is incorrect because autoscaling and good row-key design address throughput and hotspot risk, but a single cluster remains a single cluster-level failure domain. A mission-critical global service requires replication so traffic can continue if a cluster or region becomes unavailable. One cluster therefore does not satisfy the high-availability requirement even if it scales well during normal operation.</p><blockquote><strong><em>B is correct because Bigtable is purpose-built for large time-series and versioned key-value datasets requiring low latency and very high throughput. Multi-cluster replication provides redundant copies of the data, and multi-cluster routing can automatically route requests to healthy clusters when one becomes unavailable. For a global workload, clusters should be placed in regions that provide the required availability and user locality. Autoscaling on each cluster lets capacity grow and shrink with traffic, while a well-designed time-series row key avoids sequential hot ranges. This combination addresses availability, traffic spikes, scale, and hotspot avoidance together.</em></strong></blockquote><p>C is incorrect because multi-region replication is an important part of the solution, but replication alone does not guarantee cost-effective capacity management, automatic failover for every workload, or good performance. Bigtable application profiles determine routing behavior, and poorly designed time-series keys can still create hotspots regardless of the number of clusters. Fixed sizing also requires manual capacity planning for unpredictable spikes. The more complete architecture in option B includes the scaling, routing, and schema controls needed by the scenario.</p><p>D is incorrect because AlloyDB for PostgreSQL is a high-performance relational database with read pools and analytical acceleration, but the workload described is a massive, versioned time-series serving platform with unpredictable spikes and explicit hotspot concerns. Bigtable’s distributed row-key model, cluster replication, and autoscaling are more directly suited to that workload when general relational transactions are not required.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigtable/docs/autoscaling">docs.cloud.google.com/bigtable/docs/autoscaling</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/replication-overview">docs.cloud.google.com/bigtable/docs/replication-overview</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/schema-design-time-series">docs.cloud.google.com/bigtable/docs/schema-design-time-series</a></p><p><strong>Q27.</strong></p><p><strong>At Velquoria Systems, you are planning the migration of a large Oracle database to AlloyDB for PostgreSQL. The database contains a large amount of application logic implemented as stored procedures, functions, packages, triggers, and related database objects that must be migrated to PostgreSQL. You want to minimize both migration effort and application downtime. What should you do?</strong></p><p><strong>A. Use Ora2Pg for schema and code conversion and for the complete data migration.</strong></p><p><strong>B. Use Ora2Pg for schema and code conversion. Use the oracle_fdw extension in AlloyDB for PostgreSQL and replicate data from source to destination by using CREATE TABLE AS SELECT statements.</strong></p><p><strong>C. Use Database Migration Service. Set up a conversion workspace for schema and code conversion, and create a continuous migration job to perform the initial backfill and change data capture to replicate data from source to destination.</strong></p><p><strong>D. Use Database Migration Service. Set up a legacy conversion workspace for schema and code conversion, and create a continuous migration job to perform the initial backfill and change data capture to replicate data from source to destination.</strong></p><p>A is incorrect because Ora2Pg can be useful for assessing and converting Oracle schemas and PL/SQL code to PostgreSQL-compatible syntax, but using it as the complete migration mechanism does not provide the same integrated managed continuous-migration workflow as Database Migration Service. For a large production database, a one-time export and import can require an extended write freeze unless the team separately designs change capture, synchronization, monitoring, cutover, and error handling. The requirement is to reduce both code-conversion effort and downtime, so a managed conversion workspace combined with continuous replication is a better fit.</p><p>B is incorrect because oracle_fdw can access Oracle data from PostgreSQL and can be useful in specialized integration or migration patterns, but using CREATE TABLE AS SELECT statements is a custom bulk-copy approach rather than a complete low-downtime migration workflow. The team would still need to implement change data capture for transactions that occur after the initial copy, handle restart and consistency logic, monitor migration state, and coordinate cutover. This introduces significant manual engineering and operational risk for a large database.</p><blockquote><strong><em>C is correct because current Database Migration Service supports heterogeneous Oracle-to-AlloyDB for PostgreSQL migrations with conversion workspaces and migration jobs. The conversion workspace can convert Oracle schema and supported code objects such as packages, functions, stored procedures, and table triggers to PostgreSQL syntax and identify objects that require review or refactoring. A continuous migration job then performs the historical data backfill and captures ongoing source changes until the destination is ready for promotion. This allows most of the migration to occur while Oracle remains online, reducing final downtime to the controlled cutover period. Current conversion workspaces also provide interactive editing and optional Gemini-powered conversion assistance, further reducing manual conversion effort.</em></strong></blockquote><p>D is incorrect because legacy conversion workspaces remain available for compatibility, but they rely on older conversion-workspace behavior and can require more manual configuration, such as conversion mapping files and additional remediation for certain objects. The current built-in conversion workspace provides the modern interactive conversion experience, deterministic conversion rules, conversion-issue analysis, and optional Gemini-powered assistance. Therefore, a legacy workspace is not the preferred Google-recommended approach when the objective is to minimize conversion effort.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/guide">docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/guide</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/create-conversion-workspace">docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/create-conversion-workspace</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/create-migration-job">docs.cloud.google.com/database-migration/docs/oracle-to-alloydb/create-migration-job</a></p><p><strong>Q28.</strong></p><p><strong>At Pelorvanta Systems, your company is migrating a legacy on-premises Oracle database to Cloud SQL for PostgreSQL. You need to convert the Oracle schema and database code objects, including procedures, functions, packages, and triggers, to PostgreSQL with minimal application-code changes and manual intervention. What should you do?</strong></p><p><strong>A. Manually rewrite both the Oracle schema and all application queries to use PostgreSQL syntax and semantics.</strong></p><p><strong>B. Use a generic schema-conversion utility to convert only the schema automatically, and then manually rewrite all database code and application queries.</strong></p><p><strong>C. Migrate the Oracle data directly to PostgreSQL without code or schema conversion, and rely on an application compatibility layer to interpret Oracle syntax.</strong></p><p><strong>D. Use a Database Migration Service conversion workspace to convert the Oracle schema and supported code objects, review and refactor identified conversion issues, and use Database Migration Service for the migration workflow.</strong></p><p>A is incorrect because manually rewriting all schema objects and queries creates the highest migration effort and risk. Oracle and PostgreSQL differ in data types, built-in functions, packages, stored-program behavior, exception handling, sequences, transaction semantics, and other features. Reimplementing all of these manually is slow and makes it more difficult to validate that converted behavior matches the original application.</p><p>B is incorrect because schema-only conversion addresses tables and basic database objects but leaves procedures, functions, packages, triggers, and application dependencies for manual conversion. The question specifically emphasizes substantial database code and minimal manual intervention. Current Database Migration Service conversion workspaces can analyze and convert a broader set of Oracle schema and code objects and identify exactly where manual review is required.</p><p>C is incorrect because Oracle PL/SQL and PostgreSQL SQL/PLpgSQL are not transparently interchangeable. PostgreSQL cannot natively execute arbitrary Oracle packages, procedures, triggers, or Oracle-specific SQL syntax simply because an application compatibility layer is present. A heterogeneous migration requires explicit schema and code conversion, testing, and remediation.</p><blockquote><strong><em>D is correct because Database Migration Service conversion workspaces support deterministic Oracle-to-PostgreSQL schema and code conversion for objects including functions, stored procedures, packages, and table-level triggers. The workspace groups conversion issues by whether they require review, data-type adjustment, functional review, or refactoring, and provides an interactive editor for remediation. Current workspaces can also use Gemini-powered conversion assistance for supported scenarios. This automates as much conversion as possible while making the remaining Oracle-specific incompatibilities explicit, thereby minimizing manual effort and application changes.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/about-conversion-workspaces">docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/about-conversion-workspaces</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/conversion-issues-reference">docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/conversion-issues-reference</a></p><p><a href="http://docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/create-conversion-workspace">docs.cloud.google.com/database-migration/docs/oracle-to-postgresql/create-conversion-workspace</a></p><p><strong>Q29.</strong></p><p><strong>At Zorivexa Media, your organization runs a high-volume, low-latency application on Bigtable. You need to reduce Bigtable cost without compromising application performance. What should you do?</strong></p><p><strong>A. Reduce the number of replicated clusters regardless of the application’s availability requirements.</strong></p><p><strong>B. Consolidate multiple small workloads into one large cluster without first analyzing access patterns or isolation requirements.</strong></p><p><strong>C. Analyze application access patterns and row-key distribution, optimize the schema where necessary, and right-size or autoscale cluster node capacity based on CPU, storage, and latency metrics.</strong></p><p><strong>D. Change every cluster to a lower-performance storage type.</strong></p><p>A is incorrect because reducing replication can lower infrastructure cost, but it can also reduce availability, durability, geographic locality, and failover capability. If the current replication topology exists to satisfy an SLA or disaster-recovery requirement, removing replicas directly compromises non-performance requirements and may force more traffic onto fewer clusters. Cost optimization should start by eliminating inefficient capacity or schema behavior without weakening required resilience.</p><p>B is incorrect because consolidating clusters without analyzing workloads can create noisy-neighbor effects, hotspot interactions, capacity contention, and loss of workload isolation. Bigtable clusters and application profiles can intentionally separate workloads that have different performance characteristics. Consolidation can be considered after analysis, but it is not a safe universal cost optimization.</p><blockquote><strong><em>C is correct because Bigtable cost is strongly influenced by provisioned node capacity, while performance depends on whether reads and writes are evenly distributed and whether each cluster has sufficient CPU and storage headroom. Poor row-key design can create hotspots that adding nodes does not solve, so schema and access-pattern analysis should come first. After the workload is well distributed, node count can be right-sized or Bigtable autoscaling can automatically remove excess nodes during quieter periods and add capacity when demand increases. This reduces overprovisioning while preserving latency and throughput objectives.</em></strong></blockquote><p>D is incorrect because lower-performance storage can materially increase read latency and reduce throughput. Bigtable HDD is suited to large, latency-insensitive datasets and batch-oriented access rather than high-volume low-latency serving workloads. Changing storage type simply to save money can therefore violate the requirement not to compromise performance.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigtable/docs/scaling">docs.cloud.google.com/bigtable/docs/scaling</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/performance">docs.cloud.google.com/bigtable/docs/performance</a></p><p><a href="http://docs.cloud.google.com/bigtable/docs/autoscaling">docs.cloud.google.com/bigtable/docs/autoscaling</a></p><p><strong>Q30.</strong></p><p><strong>At Quoralyx Commerce, you are designing a new ecommerce application on Google Cloud. You expect high and bursty write volumes during major sales events. You need Spanner to maintain consistent performance while scaling capacity automatically. What should you do?</strong></p><p><strong>A. Manually shard the application across multiple independent Spanner instances.</strong></p><p><strong>B. Configure read-only transactions to improve write performance during peak periods.</strong></p><p><strong>C. Use interleaved tables as the primary scaling mechanism for bursty write traffic.</strong></p><p><strong>D. Enable Spanner managed autoscaling, and configure the maximum compute capacity to handle the expected peak write volume while setting an appropriate minimum capacity for the workload’s burstiness.</strong></p><p>A is incorrect because Spanner already distributes data internally into splits and horizontally scales those splits across provisioned compute capacity. Manually sharding one logical application across independent Spanner instances adds application routing, cross-shard transaction, operational, and rebalancing complexity and gives up much of Spanner’s managed distributed-database value. It is not the normal method for handling a temporary sales spike.</p><p>B is incorrect because read-only transactions can optimize consistency and execution behavior for reads, but they do not add compute capacity for a write-heavy event. The requirement specifically concerns bursty writes, so changing read transaction mode does not address the bottleneck.</p><p>C is incorrect because interleaving related tables can improve locality for particular parent-child access patterns, but it is a schema-modeling technique rather than a mechanism that dynamically increases compute capacity. Poorly chosen interleaving or primary keys can also create concentrated access patterns. It does not replace capacity planning or autoscaling for a sudden surge in writes.</p><blockquote><strong><em>D is correct because Spanner managed autoscaling automatically adjusts processing units or nodes according to CPU and storage targets within the configured minimum and maximum limits. Google recommends setting the maximum high enough to serve the heaviest expected traffic, which in this case is the major sales-event write peak. Because Spanner needs time to add compute and rebalance data, current guidance also recommends using a sufficiently high minimum capacity for spiky workloads or raising the minimum before a known peak event. This provides automatic scaling while retaining enough headroom to absorb sudden traffic growth and avoids permanently paying for peak capacity during normal periods.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/managed-autoscaler">docs.cloud.google.com/spanner/docs/managed-autoscaler</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/autoscaling-overview">docs.cloud.google.com/spanner/docs/autoscaling-overview</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/compute-capacity">docs.cloud.google.com/spanner/docs/compute-capacity</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85a91e38c490" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Professional Cloud Architect Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-professional-cloud-architect-practice-questions-0e8795364ab5?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e8795364ab5</guid>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[gcp]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 07:57:36 GMT</pubDate>
            <atom:updated>2026-10-06T07:57:36.284Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DTBimLcOE4Rbhe-OwOeOKA.jpeg" /><figcaption>KloudExams GCP Google Cloud Professional Cloud Architect Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Professional Cloud Architect certification</strong>? Knowing what cloud services do is a starting point — understanding how to combine them into architectures that balance security, reliability, performance, and cost is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise evaluating architectural trade-offs against specific business and technical requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on architecture practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>Solvenyx Retail, a fictional company, is implementing a disaster-recovery plan by replicating a production MySQL database from its private data center to a Google Cloud project over Cloud VPN. The replication is being disrupted by latency variation and a small amount of packet loss on the network path.</strong></p><p><strong>What should the company do?</strong></p><p><strong>A. Configure the database replication to use UDP.</strong></p><p><strong>B. Configure Dedicated Interconnect connectivity between the on-premises network and Google Cloud.</strong></p><p><strong>C. Restore the database to Cloud SQL once per day.</strong></p><p><strong>D. Add additional Cloud VPN connections and load-balance the replication traffic across them.</strong></p><p><strong>E. Send the replicated database transactions through Pub/Sub.</strong></p><p>A is incorrect because MySQL replication relies on reliable, ordered transport semantics and is normally carried over TCP. Switching the replication protocol to UDP is not a valid way to solve packet loss or latency problems and would remove the reliable delivery behavior required by the database protocol. Network quality should be improved rather than attempting to replace the transport with an unreliable datagram protocol.</p><blockquote><strong><em>B is correct because Dedicated Interconnect provides direct physical connectivity between an on-premises network and Google’s network instead of sending the traffic across the public internet. This reduces the number of external network hops and provides high-capacity, predictable private connectivity suited to sustained database replication and other hybrid workloads that are sensitive to network quality. For a production disaster-recovery design, the company should also select an appropriate redundant Interconnect topology according to its availability requirements.</em></strong></blockquote><p>C is incorrect because restoring a database once per day is a batch recovery process, not continuous replication. It creates a recovery point objective that can be as large as approximately a day and does not solve the stated requirement to keep the Google Cloud copy synchronized with production. Cloud SQL can be a target for suitable migration or replication architectures, but a daily restore is not a replacement for reliable network connectivity.</p><p>D is incorrect because multiple HA VPN tunnels can be used to increase aggregate VPN bandwidth and availability, but simply adding tunnels and load-balancing traffic does not inherently eliminate the latency variation or packet loss affecting this replication flow. It also adds routing and tunnel-management complexity. The scenario is asking for a more predictable private connectivity option for production replication, for which Dedicated Interconnect is the stronger choice among the listed answers.</p><p>E is incorrect because Pub/Sub is an asynchronous messaging service, not a transparent transport layer for native MySQL replication. Converting database changes into application messages would require a separate change-data-capture and replay architecture, transaction-ordering considerations, idempotency controls, and custom recovery logic. It does not directly solve the connectivity problem affecting the existing replication mechanism.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/network-connectivity/docs/interconnect/concepts/dedicated-overview">cloud.google.com/network-connectivity/docs/interconnect/concepts/dedicated-overview</a></p><p><a href="http://cloud.google.com/network-connectivity/docs/interconnect/concepts/overview">cloud.google.com/network-connectivity/docs/interconnect/concepts/overview</a></p><p><a href="http://cloud.google.com/network-connectivity/docs/vpn/concepts/topologies-increase-bandwidth">cloud.google.com/network-connectivity/docs/vpn/concepts/topologies-increase-bandwidth</a></p><p><strong>Q2.</strong></p><p><strong>At Virelix Systems, a fictional company, you have an outage in a Compute Engine managed instance group (MIG): all instances keep restarting after 5 seconds. You have an application health check configured for autohealing, but autoscaling is disabled. A colleague who is a Linux expert has offered to investigate the issue. You need to make sure that the colleague can access the VMs.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Grant your colleague the IAM role of Project Viewer.</strong></p><p><strong>B. Perform a rolling restart on the managed instance group.</strong></p><p><strong>C. Disable the health check for the managed instance group. Add the colleague’s SSH key to the project-wide SSH keys.</strong></p><p><strong>D. Disable autoscaling for the managed instance group. Add the colleague’s SSH key to the project-wide SSH keys.</strong></p><p>A is incorrect because the Project Viewer role provides read-only visibility into many project resources but does not, by itself, establish operating-system-level SSH access to a Linux VM. With metadata-managed SSH access, the user’s public key must be made available to the VM through project or instance metadata, unless the environment instead uses OS Login. More importantly, granting a read-only IAM role does nothing to stop the managed instance group from continuously repairing and recreating instances that fail the configured autohealing health check, so the colleague might still be unable to keep an instance alive long enough to diagnose the operating-system problem.</p><p>B is incorrect because a rolling restart deliberately restarts or replaces instances in the managed instance group and therefore does not address the root problem. The instances are already being restarted because the health-check-based autohealing policy considers them unhealthy. Initiating another rolling operation would add more instance churn and would make interactive troubleshooting harder rather than stabilizing a VM for investigation.</p><blockquote><strong><em>C is correct because application health checks used for MIG autohealing are independent of autoscaling. When the health check reports an instance as unhealthy, the managed instance group can repair the VM by recreating it even when autoscaling is disabled. Removing the health check or otherwise turning off the autohealing action stops that health-check-driven repair loop, allowing an affected VM to remain available for diagnosis. Adding the colleague’s SSH key to project-wide SSH metadata then allows the Linux expert to connect to VMs that use metadata-managed SSH keys. In a modern environment where OS Login is enabled, OS Login IAM roles would normally be preferred, but among the listed choices this is the only option that both stops the repeated health-check-driven recreation and provides SSH access.</em></strong></blockquote><p>D is incorrect because autoscaling is already disabled, and autoscaling is not what is causing the five-second restart cycle. Autoscaling changes the desired size of a managed instance group in response to load or other signals; autohealing repairs unhealthy instances based on health checks. Adding the SSH key could provide access, but leaving the failing autohealing configuration active means the VM can continue to be recreated before the colleague can troubleshoot it.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/compute/docs/instance-groups/autohealing-instances-in-migs">cloud.google.com/compute/docs/instance-groups/autohealing-instances-in-migs</a></p><p><a href="http://cloud.google.com/compute/docs/instances/ssh">cloud.google.com/compute/docs/instances/ssh</a></p><p><strong>Q3.</strong></p><p><strong>At Velqora Apps, a fictional company, you are creating an App Engine application that uses Firestore in Datastore mode as its persistence layer. You need to retrieve several root entities for which you already know the entity identifiers. You want to minimize Datastore mode operation overhead.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Create a Key object for each entity and perform a batch lookup.</strong></p><p><strong>B. Create a Key object for each entity and perform a separate lookup operation for every entity.</strong></p><p><strong>C. Use the identifiers to create query filters and run a single query operation.</strong></p><p><strong>D. Use the identifiers to create query filters and run a separate query operation for every entity.</strong></p><blockquote><strong><em>A is correct because an entity key uniquely identifies an entity in Firestore in Datastore mode, so a key lookup is the direct access path when the identifiers are already known. Datastore mode client libraries support batch lookup operations such as get_multi or their language-specific equivalents, allowing multiple known keys to be retrieved together. This avoids unnecessary query evaluation and reduces network and request overhead compared with issuing individual calls.</em></strong></blockquote><p>B is incorrect because individual key lookups would retrieve the correct entities, but performing one request per entity creates more round trips and request overhead than batching the known keys into one multi-lookup operation. The requirement is specifically to minimize operation overhead, so the batch lookup is preferable.</p><p>C is incorrect because a query is unnecessary when exact entity keys are already known. Queries are intended to find entities that match specified conditions and can involve index processing. Direct key lookups are simpler and more efficient for known entities, and a batch lookup is specifically designed for retrieving several keys in one request.</p><p>D is incorrect because it combines both disadvantages: it uses queries even though exact keys are available and it issues multiple separate query operations. This produces more request overhead and more query work than a single batch key lookup.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/datastore/docs/samples/datastore-batch-lookup">cloud.google.com/datastore/docs/samples/datastore-batch-lookup</a></p><p><a href="http://cloud.google.com/datastore/docs/best-practices">cloud.google.com/datastore/docs/best-practices</a></p><p><a href="http://cloud.google.com/datastore/docs/concepts/entities">cloud.google.com/datastore/docs/concepts/entities</a></p><p><strong>Q4.</strong></p><p><strong>At Brinovara Retail, a fictional company, you are moving a MySQL-backed application from an on-premises environment to Google Cloud. The application will run on Compute Engine and use Cloud SQL for MySQL. You want minimal cutover downtime, no data loss, and minimal application modification.</strong></p><p><strong>Which approach should you take?</strong></p><p><strong>A. 1. Create Cloud VPN connectivity between the Compute Engine application and the on-premises MySQL server. 2. Stop the on-premises application. 3. Create a mysqldump of the on-premises MySQL server. 4. Upload the dump to Cloud Storage. 5. Import the dump into Cloud SQL. 6. Modify the application to write to both databases while reading from its local database. 7. Start the Compute Engine application. 8. Stop the on-premises application.</strong></p><p><strong>B. 1. Configure the Cloud SQL Auth Proxy and a separate MySQL proxy. 2. Create a mysqldump of the on-premises MySQL server. 3. Upload the dump to Cloud Storage. 4. Import the dump into Cloud SQL. 5. Stop the on-premises application. 6. Start the Compute Engine application.</strong></p><p><strong>C. 1. Create private hybrid connectivity so that the Compute Engine application can continue using the on-premises MySQL server during migration. 2. Start the Compute Engine application against the existing on-premises database. 3. Configure continuous replication from the external MySQL source to Cloud SQL. 4. Allow the Cloud SQL migration replica to catch up fully. 5. Stop application writes for the final cutover. 6. Wait until replication lag reaches zero. 7. Promote the Cloud SQL destination to a writable primary instance. 8. Reconfigure the Compute Engine application to use Cloud SQL and resume the application.</strong></p><p><strong>D. 1. Stop the on-premises application. 2. Create a mysqldump of the on-premises MySQL server. 3. Upload the dump to Cloud Storage. 4. Import the dump into Cloud SQL. 5. Start the application on Compute Engine.</strong></p><p>A is incorrect because this design introduces application-level dual writes, which violates the requirement to migrate with minimal application modification and creates a difficult consistency problem. The application would need logic to handle partial failures, ordering, retries, and divergence between the two databases. It also stops the on-premises application early even though continuous database replication can keep the source online during most of the migration.</p><p>B is incorrect because a one-time dump and import does not continuously capture changes that occur on the source database while the dump is being exported, transferred, and imported. If the application remains active during that period, those later transactions can be missing from Cloud SQL. If the application is stopped before the dump to avoid data loss, downtime lasts for the entire export, transfer, and import process. Proxies solve connection or authentication concerns but do not provide continuous data synchronization.</p><blockquote><strong><em>C is correct because continuous replication is the appropriate migration pattern when the organization wants minimal downtime and no lost committed transactions. Cloud SQL supports replication from an external MySQL server, and Database Migration Service provides a managed continuous-migration workflow for the same objective. The application can continue using the source database while the initial data is copied and subsequent changes are replicated. At final cutover, the organization stops writes, waits for the replication delay to reach zero, promotes the Cloud SQL destination to a primary instance, changes the application’s database endpoint, and resumes traffic. This limits downtime primarily to the final synchronization and connection switch and avoids implementing application-level dual writes.</em></strong></blockquote><p>D is incorrect because stopping the application before a full dump and import can preserve consistency, but it creates downtime for the entire migration duration. For a substantial database, the dump, upload, import, validation, and application startup process can take a long time. Continuous replication exists specifically to keep the destination nearly synchronized before the final cutover and therefore reduce downtime.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/database-migration/docs/mysql/migration-types">cloud.google.com/database-migration/docs/mysql/migration-types</a></p><p><a href="http://cloud.google.com/database-migration/docs/mysql/promote-migration">cloud.google.com/database-migration/docs/mysql/promote-migration</a></p><p><a href="http://cloud.google.com/sql/docs/mysql/replication/external-server">cloud.google.com/sql/docs/mysql/replication/external-server</a></p><p><strong>Q5.</strong></p><p><strong>At Qorvanta Research, a fictional company, sensitive information is stored in Cloud Storage buckets. Data analysts already have IAM permissions that allow them to read the bucket data. You need to prevent those analysts from retrieving the protected data when their requests originate outside the company’s trusted office network.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Place the projects containing the buckets inside a VPC Service Controls service perimeter, create an access level that identifies the trusted office public IP CIDR ranges, and configure the perimeter so that qualifying requests from that trusted context can access the protected service.</strong></p><p><strong>B. Create VPC firewall rules on Compute Engine instances that allow traffic only when the source IP belongs to the office CIDR ranges.</strong></p><p><strong>C. Create one Cloud Run function that removes analyst IAM permissions from the buckets and another that restores them, and invoke the functions on a daily office-hours schedule.</strong></p><p><strong>D. Create Cloud VPN connectivity to the office and enable Private Google Access for on-premises hosts, but leave the Cloud Storage resources governed only by their existing IAM permissions.</strong></p><blockquote><strong><em>A is correct because IAM answers whether a principal is authorized to perform an operation, while VPC Service Controls adds a context-based data-exfiltration boundary around supported Google Cloud services such as Cloud Storage. An access level can use trusted public source IP CIDR ranges, identities, and other context conditions, and the service perimeter can use that context to permit approved access while rejecting requests originating outside the trusted boundary. This directly addresses the requirement that the same analyst identity must be allowed or denied based on where the request originates. The office public IP ranges used for this control must be trustworthy and tightly controlled.</em></strong></blockquote><p>B is incorrect because VPC firewall rules apply to network traffic involving resources such as VM network interfaces; they do not function as the authorization boundary for requests sent to the Cloud Storage API. An analyst can access Cloud Storage from a browser, client library, workstation, or another network path without that request traversing a Compute Engine VM firewall rule. Therefore, this control does not protect the bucket from off-network API access.</p><p>C is incorrect because repeatedly removing and restoring IAM bindings according to business hours is operationally fragile and does not actually express the stated location-based policy. An analyst inside the office after the scheduled removal would lose access, while an analyst outside the office during the allowed time would still have access. This design also creates IAM churn, propagation concerns, failure modes, and unnecessary custom code.</p><p>D is incorrect because Cloud VPN and Private Google Access can provide a private path from on-premises systems to supported Google APIs, but merely creating that path does not force analysts to use it. If their IAM permissions remain valid and no service-perimeter restriction exists, they can potentially send authorized Cloud Storage requests through other internet paths. The requirement needs an access boundary, not only an additional connectivity option.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/vpc-service-controls/docs/overview">cloud.google.com/vpc-service-controls/docs/overview</a></p><p><a href="http://cloud.google.com/vpc-service-controls/docs/access-level-design">cloud.google.com/vpc-service-controls/docs/access-level-design</a></p><p><a href="http://cloud.google.com/vpc-service-controls/docs/create-service-perimeters">cloud.google.com/vpc-service-controls/docs/create-service-perimeters</a></p><p><strong>Q6.</strong></p><p><strong>At Arvelion Compute, a fictional company, an application runs in a Compute Engine managed instance group (MIG). You have created a new instance template containing a non-critical application update. You do not want any currently running VM to be restarted, refreshed, or replaced solely because of this change, but every new VM that the MIG creates should use the new template.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Start a rolling restart operation.</strong></p><p><strong>B. Start a rolling replace operation.</strong></p><p><strong>C. Update the MIG to the new template and use the proactive update policy.</strong></p><p><strong>D. Update the MIG to the new template and use the selective (opportunistic) update policy.</strong></p><p>A is incorrect because a rolling restart deliberately restarts existing instances. The requirement explicitly says that the current VMs must not be disrupted merely to apply this non-critical change. A restart also does not express the desired policy that the new configuration should apply naturally only as new instances are created.</p><p>B is incorrect because a rolling replace operation deletes and recreates existing managed instances. That would cause the exact disruption the company is trying to avoid. Replacement is appropriate when an update must be forced onto existing instances and cannot be applied by a less disruptive action, but it is unnecessary for an update that can wait until normal instance creation.</p><p>C is incorrect because a proactive MIG update policy automatically applies the new configuration to existing managed instances according to the rollout parameters. Although the update can be controlled by settings such as maximum surge and maximum unavailable, existing instances are still targeted for convergence toward the new template. That conflicts with the requirement to leave every running instance unchanged.</p><blockquote><strong><em>D is correct because the selective, also called opportunistic, update policy lets the MIG’s target configuration change without automatically applying that change to existing VMs. New instances created because of autoscaling, autohealing, manual resize, or later replacement use the current target template, while the existing VMs remain on their present configuration until you explicitly choose to update them or they are naturally replaced. This exactly matches the requirement for a non-disruptive gradual adoption of the new template.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/compute/docs/instance-groups/rolling-out-updates-to-managed-instance-groups">cloud.google.com/compute/docs/instance-groups/rolling-out-updates-to-managed-instance-groups</a></p><p><a href="http://cloud.google.com/compute/docs/instance-groups/set-mig-aic">cloud.google.com/compute/docs/instance-groups/set-mig-aic</a></p><p><strong>Q7.</strong></p><p><strong>At Veltrion Applications, a fictional company, you are developing multiple microservices on Google Kubernetes Engine (GKE). The microservices must remain internal to the cluster. You need to configure a specific replica count for each microservice and give other microservices a stable way to address it even when its Pods are created, removed, or rescheduled.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Deploy each microservice as a Kubernetes Deployment. Expose each Deployment internally by using a Kubernetes Service, and use the Service DNS name when other microservices call it.</strong></p><p><strong>B. Deploy each microservice as a Kubernetes Deployment. Expose each Deployment through an Ingress resource, and use the Ingress IP address for calls between microservices inside the cluster.</strong></p><p><strong>C. Deploy each microservice as a single Pod. Expose that Pod by using a Kubernetes Service, and use the Service DNS name for calls from other microservices.</strong></p><p><strong>D. Deploy each microservice as a single Pod. Expose the Pod through an Ingress resource, and use the Ingress address for calls between microservices.</strong></p><blockquote><strong><em>A is correct because a Deployment manages the desired number of interchangeable Pod replicas and continuously reconciles the actual replica count with that desired state. Individual Pod IP addresses are ephemeral, so clients should not use those addresses directly. A Kubernetes Service selects the Pods belonging to the microservice and provides a stable virtual endpoint and DNS name inside the cluster. Other workloads can therefore call the Service by name while Kubernetes routes requests to currently healthy endpoints, regardless of how many replicas exist or which specific Pod IP addresses change.</em></strong></blockquote><p>B is incorrect because the Deployment portion is appropriate, but Ingress is intended for Layer 7 HTTP(S) traffic entering services through an ingress load-balancing configuration. It is not required for ordinary internal service-to-service discovery inside one cluster. Using an ingress endpoint for every internal microservice would add routing and load-balancing complexity when a ClusterIP Service and its DNS name already provide stable internal discovery.</p><p>C is incorrect because the Service would provide stable discovery, but a single unmanaged Pod does not give the company a declarative replica count, rolling updates, or automatic replacement behavior equivalent to a Deployment. The requirement specifically includes configuring the number of replicas for each microservice, which is a workload-controller responsibility.</p><p>D is incorrect because it combines both limitations: an individual Pod is not the correct controller for maintaining a replica set, and Ingress is unnecessary for internal cluster-only communication. Deployments plus internal Services are the Kubernetes-native combination designed for this pattern.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/concepts/service">cloud.google.com/kubernetes-engine/docs/concepts/service</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/concepts/about-service-discovery">cloud.google.com/kubernetes-engine/docs/concepts/about-service-discovery</a></p><p><a href="http://kubernetes.io/docs/concepts/workloads/controllers/deployment/">kubernetes.io/docs/concepts/workloads/controllers/deployment/</a></p><p><strong>Q8.</strong></p><p><strong>At Sentrava Security, a fictional company, all Google Cloud logs are collected in Cloud Logging. The security team wants near-real-time automated reaction when selected log entries indicate anomalies such as an unauthorized firewall change or a server compromise. You want to use managed Google Cloud services and follow recommended practices.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Create a Cloud Scheduler job that queries Cloud Logging once every minute for the security events.</strong></p><p><strong>B. Route all logs to BigQuery and repeatedly run scheduled BigQuery queries to search the exported log data for the events.</strong></p><p><strong>C. Create a Cloud Logging sink that routes the relevant log entries to Pub/Sub, and trigger a Cloud Run function from the Pub/Sub messages to perform the required security-response logic.</strong></p><p><strong>D. Route the relevant logs to a Cloud Storage bucket and trigger a Cloud Run service only after exported log objects are written to the bucket.</strong></p><p>A is incorrect because polling Cloud Logging every minute adds custom query logic, scheduled operations, duplicate-detection concerns, and up to a polling interval of avoidable latency. It also repeatedly queries when no relevant event has occurred. Cloud Logging can route matching entries as they arrive, so an event-driven workflow is more efficient and responds faster.</p><p>B is incorrect because BigQuery is valuable for long-term analysis, correlation, reporting, and large-scale searches over exported logs, but repeatedly querying a warehouse is not the most direct mechanism for immediate reaction to individual security events. It creates additional ingestion and scheduled-query latency and consumes analytical resources even when there are no matching incidents.</p><blockquote><strong><em>C is correct because Cloud Logging sinks can filter log entries and route only the relevant events to a Pub/Sub topic. Pub/Sub provides an asynchronous event-delivery boundary, and a Cloud Run function can be triggered from messages on that topic through the supported event-triggering integration. The function can then notify responders, enrich the finding, isolate a resource, create a ticket, or invoke another approved response workflow. This event-driven design reacts soon after the relevant log entry is produced and avoids repeatedly polling the log store.</em></strong></blockquote><p>D is incorrect because Cloud Storage log sinks are primarily useful for archival and batch-oriented log retention. Log files are written as objects rather than delivered as an immediate per-event stream, so this path introduces more latency than routing the matching entries to Pub/Sub. For rapid automated reaction, Pub/Sub is the appropriate Cloud Logging sink destination among the listed options.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/logging/docs/routing/overview">cloud.google.com/logging/docs/routing/overview</a></p><p><a href="http://cloud.google.com/run/docs/triggering/trigger-functions-from-log-entries">cloud.google.com/run/docs/triggering/trigger-functions-from-log-entries</a></p><p><a href="http://cloud.google.com/run/docs/triggering/pubsub-triggers">cloud.google.com/run/docs/triggering/pubsub-triggers</a></p><p><strong>Q9.</strong></p><p><strong>At Arqelion Labs, a fictional company, you are testing a microservices application on Google Kubernetes Engine (GKE) with Cloud Service Mesh. You want to validate how the rest of the application behaves if one specific microservice suddenly fails, without actually destroying cluster infrastructure.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Add a taint to one GKE node and configure Pod anti-affinity for the target microservice by using that node’s identity.</strong></p><p><strong>B. Configure Cloud Service Mesh fault injection for the target microservice so that selected requests are aborted, simulating service failure.</strong></p><p><strong>C. Delete one of the GKE worker nodes and observe the application.</strong></p><p><strong>D. Configure Cloud Service Mesh traffic management to route traffic away from the target microservice so that clients do not experience its failure.</strong></p><p>A is incorrect because taints influence which Pods can be scheduled on a node, while Pod affinity and anti-affinity influence placement according to labels and topology. These scheduling controls do not simulate a microservice returning failures to its callers. They can change where the workload runs, but they do not provide a controlled application-level fault that exercises retry, timeout, fallback, or error-handling behavior.</p><blockquote><strong><em>B is correct because Cloud Service Mesh supports traffic fault injection specifically for resilience testing. A fault-injection policy can intentionally abort a configurable percentage of requests or add delays, allowing the team to simulate service errors, overload, latency, or other failure conditions without terminating the actual microservice or damaging the GKE cluster. By targeting the route for the selected service, the team can observe how upstream and downstream microservices react to the simulated crash-like behavior and validate retries, circuit breaking, timeouts, fallback behavior, and user-facing error handling in a controlled way.</em></strong></blockquote><p>C is incorrect because deleting a worker node tests node and workload scheduling resilience rather than the behavior of the application when one particular microservice fails. Kubernetes may simply reschedule the affected Pods to another node, and unrelated workloads on the deleted node can also be disrupted. The blast radius is therefore broader and the simulated failure is less precise than a service-mesh fault.</p><p>D is incorrect because steering traffic away from the target microservice avoids exposing callers to the failure rather than testing how they behave when the service fails. Traffic management is useful for failover and controlled routing, but the objective is to deliberately inject a failure and observe resilience. Fault injection is the Cloud Service Mesh capability intended for that test.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/service-mesh/docs/service-routing/advanced-traffic-management">cloud.google.com/service-mesh/docs/service-routing/advanced-traffic-management</a></p><p><a href="http://cloud.google.com/service-mesh/docs/overview">cloud.google.com/service-mesh/docs/overview</a></p><p><a href="http://cloud.google.com/products/service-mesh">cloud.google.com/products/service-mesh</a></p><p><strong>Q10.</strong></p><p><strong>At Trivora Platforms, a fictional company, an application runs as a Deployment on Google Kubernetes Engine (GKE). The company has separate development, staging, and production clusters. The team can currently deploy a container image directly to production without first proving that the image passed the required development and staging checks. You want a Google Cloud control that can be implemented quickly, allows the team to retain normal development autonomy, and prevents unapproved images from being deployed to later environments.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Configure a Kubernetes container lifecycle hook that prevents a container from starting when the image has not been approved for that environment.</strong></p><p><strong>B. Create a written corporate policy stating that teams must not deploy a container image to an environment unless the image was tested in an earlier environment.</strong></p><p><strong>C. Configure Binary Authorization policies for the development, staging, and production clusters, and create the required attestations as part of the CI/CD promotion process.</strong></p><p><strong>D. Develop and operate a custom Kubernetes admission controller that prevents a container from starting when the image has not been approved for that environment.</strong></p><p>A is incorrect because lifecycle hooks execute as part of a container’s lifecycle after Kubernetes has already admitted and scheduled the workload. They are application or container lifecycle mechanisms, not a centralized software-supply-chain admission policy. A malicious or incorrectly configured workload could also omit the expected hook, so this is not a reliable enforcement boundary for production deployment authorization.</p><p>B is incorrect because a written process is useful governance, but it does not technically prevent a user who has deployment permission from bypassing the required sequence. The problem in the scenario is that direct production deployment is already possible, so the solution needs an enforceable admission control rather than another procedural instruction.</p><blockquote><strong><em>C is correct because Binary Authorization is a managed Google Cloud deployment-control service for GKE that can require verifiable attestations before an image is admitted. The CI/CD workflow can create an attestation after an image successfully completes a required stage, and the policy on the next cluster can require that attestation before allowing deployment. Production can therefore require evidence that the exact image passed the designated earlier controls. This provides centralized policy enforcement with less custom engineering than writing and maintaining a bespoke admission controller.</em></strong></blockquote><p>D is incorrect because a custom admission controller could be written to enforce similar logic, but it requires the team to design, secure, deploy, monitor, test, and maintain custom admission code. Binary Authorization already provides Google-managed policy and attestation capabilities for this software-supply-chain use case, so a custom controller does not meet the goal of quick implementation with minimal effort.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/binary-authorization/docs/overview">cloud.google.com/binary-authorization/docs/overview</a></p><p><a href="http://cloud.google.com/binary-authorization/docs/making-attestations">cloud.google.com/binary-authorization/docs/making-attestations</a></p><p><a href="http://cloud.google.com/binary-authorization/docs/configure-policy-gke">cloud.google.com/binary-authorization/docs/configure-policy-gke</a></p><p><strong>Q11.</strong></p><p><strong>At Qorvexa Research, a fictional company, sensitive data is stored in a Cloud Storage bucket and processed by Managed Service for Apache Spark. Regulatory policy requires the company to control and rotate the encryption key that protects the stored objects. You want to follow Google-recommended security practices.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Create a key in Cloud Key Management Service (Cloud KMS) and call the Cloud KMS encrypt API yourself to application-encrypt every object before uploading it to Cloud Storage.</strong></p><p><strong>B. Create and manage a key in Cloud KMS, grant the Cloud Storage service agent the required key permission, and configure the bucket to use that Cloud KMS key as its default customer-managed encryption key.</strong></p><p><strong>C. Generate a GPG key pair, manually encrypt each file with GPG, and upload only the encrypted files to Cloud Storage.</strong></p><p><strong>D. Generate an AES-256 key outside Google Cloud and use Cloud Storage customer-supplied encryption keys for the objects.</strong></p><p>A is incorrect because application-layer encryption with direct Cloud KMS encrypt calls shifts encryption, metadata, key-version tracking, and decryption behavior into custom application code. Cloud KMS also has payload-size constraints on direct symmetric encryption operations and is intended to protect key material or relatively small payloads rather than replace the storage service’s integrated envelope-encryption workflow for arbitrary large objects. Cloud Storage provides native CMEK integration, which is simpler and easier to audit and operate.</p><blockquote><strong><em>B is correct because Cloud Storage integrates directly with Cloud KMS customer-managed encryption keys. Configuring a bucket default CMEK causes new object writes that do not explicitly select another encryption method to use the configured Cloud KMS key, while the company retains control over key IAM, protection level, rotation, disablement, and audit logs. Cloud KMS can rotate the CryptoKey to a new primary key version, after which new encryption operations use the new version while earlier versions remain available to decrypt existing objects. If policy requires old objects themselves to be re-encrypted under the new version or a replacement key, the company can rewrite those objects or use supported batch operations. This is the Google-managed integration intended for customer-controlled storage encryption and is compatible with downstream services when their service identities have the required KMS permissions.</em></strong></blockquote><p>C is incorrect because GPG encryption can protect data before it reaches Google Cloud, but the organization would have to design key distribution, rotation, decryption, access control, recovery, and integration with its data-processing jobs itself. Managed Service for Apache Spark would need custom logic to decrypt the content before processing it. Native Cloud Storage CMEK integration provides centralized Cloud KMS policy and auditability with less custom operational burden.</p><p>D is incorrect because Cloud Storage supports customer-supplied encryption keys, but the customer must provide the key material for every operation that needs to decrypt the object and must independently manage secure storage, distribution, recovery, and rotation. Google recommends Cloud KMS-backed CMEK when an organization wants customer control with integrated key lifecycle management. CMEK also provides centralized IAM and key-usage audit capabilities that are easier to operate for a regulated analytics environment.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/storage/docs/encryption/customer-managed-keys">cloud.google.com/storage/docs/encryption/customer-managed-keys</a></p><p><a href="http://cloud.google.com/storage/docs/encryption/using-customer-managed-keys">cloud.google.com/storage/docs/encryption/using-customer-managed-keys</a></p><p><a href="http://cloud.google.com/managed-spark/docs/concepts/security-best-practices">cloud.google.com/managed-spark/docs/concepts/security-best-practices</a></p><p><strong>Q12.</strong></p><p><strong>At Cendrixa Analytics, a fictional company, a Google Cloud project uses BigQuery for data warehousing. The on-premises environment connects to Google Cloud through Cloud VPN. The security team wants to reduce the risk of data exfiltration caused by malicious insiders, compromised code, or accidental data sharing.</strong></p><p><strong>What should the security team do?</strong></p><p><strong>A. Configure Private Google Access for on-premises hosts without creating a VPC Service Controls perimeter.</strong></p><p><strong>B. Create one service account, grant it BigQuery Job User and Storage Object Viewer permissions, and remove all other IAM access from the project.</strong></p><p><strong>C. Protect the relevant projects and BigQuery service with VPC Service Controls, and configure Private Google Access for on-premises hosts through the authorized VPC network, preferably using restricted.googleapis.com for supported protected APIs.</strong></p><p><strong>D. Enable Private Google Access on the VPC network without configuring VPC Service Controls.</strong></p><p>A is incorrect because Private Google Access for on-premises hosts provides a private network path from an on-premises environment to supported Google APIs and services through Cloud VPN or Cloud Interconnect, but it does not by itself define a data-exfiltration boundary. A properly authorized identity could still attempt to move protected data to an unauthorized Google Cloud resource if no service perimeter restricts that exchange. Private connectivity and exfiltration controls solve different problems.</p><p>B is incorrect because consolidating access into one service account is neither a complete exfiltration control nor a practical least-privilege design for all users and applications. Compromised code running as that service account could still misuse its permissions, and removing all other project access can break legitimate administrative and analytical workflows. IAM should remain carefully scoped, but the scenario specifically calls for an additional boundary that limits where protected data can move.</p><blockquote><strong><em>C is correct because VPC Service Controls is designed to mitigate data exfiltration from supported Google-managed services such as BigQuery by creating service perimeters and controlling ingress and egress across those boundaries. For on-premises clients, the organization can extend trusted access through the VPC network that terminates the Cloud VPN connection and use Private Google Access for on-premises hosts. Google recommends restricted.googleapis.com when only VPC Service Controls-supported APIs are required because the restricted VIP blocks access to unsupported Google APIs and provides an additional exfiltration safeguard. IAM should still be used for identity authorization, making this a defense-in-depth architecture.</em></strong></blockquote><p>D is incorrect because Private Google Access changes the network path used to reach Google APIs; it does not prevent an authorized or compromised client from moving BigQuery data to an unauthorized resource. Without a service perimeter and appropriate ingress and egress policy, the private path alone does not meet the stated requirement to mitigate malicious or accidental data exfiltration.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/vpc-service-controls/docs/overview">cloud.google.com/vpc-service-controls/docs/overview</a></p><p><a href="http://cloud.google.com/vpc-service-controls/docs/private-connectivity">cloud.google.com/vpc-service-controls/docs/private-connectivity</a></p><p><a href="http://cloud.google.com/bigquery/docs/vpc-sc">cloud.google.com/bigquery/docs/vpc-sc</a></p><p><strong>Q13.</strong></p><p><strong>At Prymara Engineering, a fictional company, a custom C++ test suite runs throughout the day on Linux VMs. A complete test run takes several hours because only a small number of on-premises test servers are available. The company wants to move the test infrastructure to Google Cloud, substantially shorten the full test cycle by running more test capacity when needed, and change the existing tests as little as possible.</strong></p><p><strong>Which infrastructure should you recommend?</strong></p><p><strong>A. Compute Engine unmanaged instance groups with an external passthrough Network Load Balancer.</strong></p><p><strong>B. Compute Engine managed instance groups with autoscaling.</strong></p><p><strong>C. Managed Service for Apache Spark clusters that run each C++ test as an Apache Hadoop job.</strong></p><p><strong>D. App Engine with Cloud Logging and Cloud Monitoring.</strong></p><p>A is incorrect because an unmanaged instance group can group independently managed VMs and can be used as a load-balancing backend, but Compute Engine autoscaling is designed for managed instance groups rather than unmanaged groups. The company would therefore need to provision and remove test VMs itself. A Network Load Balancer also does not solve the principal problem, which is dynamically increasing parallel test-execution capacity.</p><blockquote><strong><em>B is correct because a Compute Engine managed instance group can create identical Linux VMs from an instance template and automatically increase or decrease the number of instances according to configured autoscaling signals. The existing C++ test binaries and supporting Linux environment can therefore be packaged into an image or startup configuration with relatively little application change. If the test suite can partition test cases across workers, increasing the number of test VMs allows more tests to run concurrently and reduces the wall-clock time required for the complete test cycle. The managed group also automatically recreates unhealthy instances and simplifies fleet management compared with individually managed VMs.</em></strong></blockquote><p>C is incorrect because Managed Service for Apache Spark is intended for Spark, Hadoop, and related distributed data-processing workloads. A custom C++ test suite that already runs on ordinary Linux VMs would need substantial adaptation to turn each test into a compatible distributed data-processing job. That directly conflicts with the requirement to change the test system as little as possible.</p><p>D is incorrect because App Engine is a managed web application platform and is not a general replacement for arbitrary Linux VM test workers running a custom C++ test harness. Cloud Logging and Cloud Monitoring provide observability but do not supply the additional Linux test execution capacity required to shorten the test suite. The workload’s existing VM-oriented execution model maps much more naturally to Compute Engine.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/compute/docs/instance-groups">cloud.google.com/compute/docs/instance-groups</a></p><p><a href="http://cloud.google.com/compute/docs/autoscaler">cloud.google.com/compute/docs/autoscaler</a></p><p><strong>Q14.</strong></p><p><strong>At Velquorin Software, a fictional company, the business places a high value on responding quickly to customer needs. Its primary engineering objectives are release speed and agility. You want to reduce the chance that security defects are accidentally introduced without creating slow, manual approval bottlenecks.</strong></p><p><strong>Which two actions should you take? (Choose two.)</strong></p><p><strong>A. Require every code check-in to be peer reviewed by a security subject-matter expert.</strong></p><p><strong>B. Use source-code security analyzers as part of the CI/CD pipeline.</strong></p><p><strong>C. Ensure that you have stubs to unit test all interfaces between components.</strong></p><p><strong>D. Enable code signing and use a trusted artifact repository integrated with the CI/CD pipeline.</strong></p><p><strong>E. Run vulnerability security scanning as part of the CI/CD pipeline.</strong></p><p>A is incorrect because expert security review can identify subtle design problems that automation might miss, but requiring a security specialist to review every individual check-in creates a serial manual dependency in the development process. That requirement can become a bottleneck as the number of commits and teams grows and therefore conflicts with the company’s primary goals of release speed and agility. Security experts are better used to define secure coding standards, review high-risk designs, improve automated checks, and investigate findings while normal code changes are subjected to scalable automated security controls.</p><blockquote><strong><em>B is correct because automated source-code security analysis can run on every change as part of continuous integration and provide rapid feedback before vulnerable code advances through the delivery pipeline. Static analysis, secret scanning, dependency analysis, policy checks, and similar automated techniques help detect insecure coding patterns and accidentally committed sensitive material early in the software lifecycle. Google Cloud software-supply-chain guidance recommends shifting security checks earlier and integrating automated security analysis into developer and CI/CD workflows. This reduces reliance on manual review while preserving fast iteration.</em></strong></blockquote><p>C is incorrect because interface stubs and unit tests are valuable software-quality practices, but they primarily validate functional behavior and component contracts. They do not systematically identify security problems such as vulnerable dependencies, injection risks, unsafe cryptographic use, embedded credentials, or insecure configurations. Unit testing should remain part of the pipeline, but it does not replace security-specific automated analysis.</p><p>D is incorrect because code signing, trusted build provenance, and a controlled artifact repository protect software-supply-chain integrity. They help demonstrate where an artifact came from and can prevent unauthorized or altered artifacts from being deployed. However, a correctly signed artifact can still contain a vulnerability that a developer accidentally introduced. This control therefore addresses artifact authenticity and integrity rather than directly reducing the probability that security defects enter the source code. It is a strong complementary control, but it is not one of the two best answers to the specific question.</p><blockquote><strong><em>E is correct because automated vulnerability scanning lets the organization detect known vulnerabilities and malicious or vulnerable packages without inserting a human approval step into every release. Artifact Analysis, for example, can scan artifacts stored in Artifact Registry and can be incorporated into a software-supply-chain policy so that vulnerable artifacts are detected before deployment. Combining vulnerability scanning with source-code analysis gives fast feedback at more than one stage of the pipeline and supports both agility and continuous security.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/software-supply-chain-security/docs/safeguard-source">cloud.google.com/software-supply-chain-security/docs/safeguard-source</a></p><p><a href="http://cloud.google.com/software-supply-chain-security/docs/safeguard-builds">cloud.google.com/software-supply-chain-security/docs/safeguard-builds</a></p><p><a href="http://cloud.google.com/artifact-analysis/docs/scanning-types">cloud.google.com/artifact-analysis/docs/scanning-types</a></p><p><strong>Q15.</strong></p><p><strong>At Lumeraxa Group, a fictional company, the organization uses Google Workspace, Cloud Identity, and a Google Cloud organization resource. Several legacy developer projects were created before centralized governance was established and still need to be brought under the organization. You want developers to be able to create development projects but prevent them from modifying production projects. Policies that apply everywhere should be managed centrally, while production must be able to inherit additional, more restrictive policies. You want a structure that remains easy to adapt as business requirements change.</strong></p><p><strong>How should you design the resource hierarchy?</strong></p><p><strong>A. 1. Create a second Google Workspace account and a second Google Cloud organization. 2. Grant all developers the Project Creator role on the new organization. 3. Move the developer projects into the new organization. 4. Configure common policies independently on both organizations. 5. Configure additional production policies on the original organization.</strong></p><p><strong>B. 1. Create a Production folder under the organization. 2. Grant all developers the Project Creator role on the organization. 3. Move the developer projects directly under the organization. 4. Configure common policies on the organization. 5. Configure additional production policies on the Production folder.</strong></p><p><strong>C. 1. Create Development and Production folders under the organization. 2. Grant developers the Project Creator role on the Development folder. 3. Move the legacy developer projects into the Development folder. 4. Configure common policies on the organization. 5. Configure additional production policies on the Production folder.</strong></p><p><strong>D. 1. Reserve the organization resource for production projects only. 2. Do not grant developers Project Creator on the organization. 3. Keep development projects outside the organization. 4. Configure policies on the organization. 5. Configure production policies individually on every production project.</strong></p><p>A is incorrect because creating a separate Google Workspace account and organization fragments central governance. Common IAM and organization policies would need to be maintained in multiple hierarchy roots, resource visibility would be split, and moving workloads between development and production governance domains would be more disruptive. Google Cloud folders exist specifically to create administration and policy boundaries inside one organization.</p><p>B is incorrect because the Production folder is useful, but granting Project Creator at the organization level lets the developers create projects under the organization outside the intended Development boundary when they have the required placement permissions. The design also does not create a dedicated Development folder to serve as a stable policy and delegated-administration boundary. Separating the environments at the folder layer provides clearer governance.</p><blockquote><strong><em>C is correct because the organization resource provides the common root for centrally inherited policies, while Development and Production folders create durable environment-specific policy boundaries. Project Creator can be delegated on the Development folder so that developers create projects in the part of the hierarchy they control without receiving equivalent rights over Production. Common organization policies can be enforced at the organization, and stricter production controls can be applied at the Production folder and inherited by all production projects. The legacy projects can also be moved into the Development folder once the required migration and IAM permissions are in place. This structure follows the Google Cloud resource hierarchy and adapts cleanly as projects are created, retired, or reorganized.</em></strong></blockquote><p>D is incorrect because keeping development projects outside the organization prevents central governance from consistently applying to those projects and undermines organization-wide visibility. It also duplicates production policy administration by setting restrictions independently on individual projects instead of using folder inheritance. Modern managed Google Workspace and Cloud Identity users normally create new projects within their organization resource, so the long-term design should bring governed workloads into the hierarchy rather than deliberately preserve shadow projects.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy">cloud.google.com/resource-manager/docs/cloud-platform-resource-hierarchy</a></p><p><a href="http://cloud.google.com/resource-manager/docs/access-control-folders">cloud.google.com/resource-manager/docs/access-control-folders</a></p><p><a href="http://cloud.google.com/resource-manager/docs/creating-managing-projects">cloud.google.com/resource-manager/docs/creating-managing-projects</a></p><p><strong>Q16.</strong></p><p><strong>At Pyralune Analytics, a fictional company, a data-warehousing team receives data from external partners for analysis. The incoming data contains personally identifiable information (PII), but the company must process the feed and persist only a version from which the PII has been removed or de-identified. The raw PII must not first be stored in the analytical destination.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Create a Dataflow pipeline that retrieves the data from the external sources, uses Sensitive Data Protection during the pipeline to identify and remove or transform PII, and writes only the sanitized result to BigQuery.</strong></p><p><strong>B. Create a Dataflow pipeline that retrieves the data from the external sources, writes non-PII data to BigQuery, and stores all PII fields in a Cloud Storage bucket with a retention policy.</strong></p><p><strong>C. Ask the external partners to upload the raw data to Cloud Storage, configure Bucket Lock on that bucket, run a Dataflow pipeline that uses Sensitive Data Protection to remove PII, and write the sanitized result to BigQuery.</strong></p><p><strong>D. Ask the external partners to load the raw data directly into BigQuery, and then run a Dataflow pipeline that copies the data to a new table while omitting columns that contain PII.</strong></p><blockquote><strong><em>A is correct because Sensitive Data Protection provides purpose-built inspection and de-identification capabilities for sensitive information such as names, identifiers, addresses, payment data, and other PII. Applying those transformations inside the ingestion or processing pipeline lets the organization avoid persisting the unredacted partner data in BigQuery and store only the sanitized analytical representation. Dataflow provides the scalable processing layer and BigQuery provides the downstream analytical warehouse. The exact design should select the appropriate de-identification transformation, such as redaction, masking, replacement, or tokenization, based on the business use case.</em></strong></blockquote><p>B is incorrect because the requirement says that PII must not be stored. Writing the PII to a Cloud Storage bucket still persists the sensitive data, even if the bucket has a retention policy. A retention policy protects data from premature deletion; it does not remove its sensitivity or satisfy a requirement that the company not retain the PII.</p><p>C is incorrect because this workflow first persists the complete raw dataset, including the PII, in Cloud Storage. Bucket Lock makes those objects harder to delete and can therefore increase rather than reduce the period during which the prohibited raw PII is retained. The data should be de-identified before it is written to the long-term analytical storage layer if the policy is to avoid storing PII.</p><p>D is incorrect because importing the source data directly into BigQuery means the PII has already been stored in the warehouse before the sanitizing pipeline runs. Deleting or excluding it from a second table does not change the fact that the original table contained the sensitive information. The processing path should transform the sensitive fields before persistence in the destination data store.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/sensitive-data-protection/docs/deidentify-sensitive-data">cloud.google.com/sensitive-data-protection/docs/deidentify-sensitive-data</a></p><p><a href="http://cloud.google.com/sensitive-data-protection/docs/transformations-reference">cloud.google.com/sensitive-data-protection/docs/transformations-reference</a></p><p><a href="http://cloud.google.com/dataflow/docs/overview">cloud.google.com/dataflow/docs/overview</a></p><p><strong>Q17.</strong></p><p><strong>At Torvexa Applications, a fictional company, an application runs as a Deployment on Google Kubernetes Engine (GKE). During rolling releases, production outages have occurred because new Pods contain configuration values used only in production and sometimes start in a state that cannot serve requests correctly. You want a preventive platform control that keeps bad replacement Pods from receiving traffic during the rollout.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Configure appropriate readiness and liveness probes in the Pod specification.</strong></p><p><strong>B. Configure health checks only on the Compute Engine managed instance group that provides GKE worker nodes.</strong></p><p><strong>C. Create a scheduled job that periodically checks whether the application is available.</strong></p><p><strong>D. Configure an uptime-check alert in Cloud Monitoring.</strong></p><blockquote><strong><em>A is correct because a Kubernetes readiness probe determines whether a Pod is currently ready to receive traffic. A newly deployed Pod with an invalid production configuration can remain NotReady, which keeps it out of normal Service endpoints and prevents Kubernetes from treating it as a healthy replacement during the rollout. With container-native GKE load balancing, readiness and load-balancer readiness gates also help ensure that newly created endpoints are healthy before old Pods are removed. A liveness probe complements this by restarting a container when it enters a state from which it cannot recover. Correctly designed probes therefore provide an automated preventive guard against routing production traffic to a broken revision.</em></strong></blockquote><p>B is incorrect because node-level health checks determine whether worker VMs are functioning; they do not validate that a particular application container has the correct production-only parameters or is ready to serve requests. A healthy GKE node can host an unhealthy or misconfigured Pod. Application health must therefore be represented at the Pod or workload level.</p><p>C is incorrect because a scheduled availability check is reactive and occurs only at its polling interval. It can tell the team that an outage has already occurred, but it does not participate in Kubernetes endpoint readiness or rollout decisions and therefore cannot stop traffic from being sent to a bad replacement Pod.</p><p>D is incorrect because a Cloud Monitoring uptime alert is an important detection mechanism, but an alert is generated after the monitored endpoint begins violating its availability condition. It does not prevent an unhealthy Pod from becoming a serving endpoint in the first place. Readiness probes provide the preventive behavior that the question requires.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/concepts/ingress-routing-security">cloud.google.com/kubernetes-engine/docs/concepts/ingress-routing-security</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/troubleshooting/ingress-health-checks">cloud.google.com/kubernetes-engine/docs/troubleshooting/ingress-health-checks</a></p><p><a href="http://cloud.google.com/kubernetes-engine/docs/how-to/stateless-apps">cloud.google.com/kubernetes-engine/docs/how-to/stateless-apps</a></p><p><strong>Q18.</strong></p><p><strong>At Qorvexa Health, a fictional healthcare provider based in Qatar, clinicians submit sensitive patient notes that must be summarized by a powerful pre-trained generative model. A strict data-residency requirement states that the patient-note content must never be processed outside Qatar. You want to use current Google Cloud capabilities while meeting this requirement.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Write application code that sends the patient notes to a generic Gemini API endpoint and use VPC Service Controls to protect the API request.</strong></p><p><strong>B. Use Gemini Enterprise Agent Platform Model Garden to select an appropriate Gemma open model, deploy the model to a dedicated endpoint in me-central1 (Doha, Qatar), and send summarization requests only to that regional endpoint.</strong></p><p><strong>C. Use the Cloud Natural Language API to analyze the text and configure it to generate a generative summary of the patient notes.</strong></p><p><strong>D. Gather a large anonymized dataset of medical notes, train a new summarization model from scratch with Gemini Enterprise Agent Platform custom training in me-central1, and deploy the custom model to a Doha endpoint.</strong></p><p>A is incorrect because VPC Service Controls can help restrict data movement across a service perimeter, but it does not create a Qatar-resident model endpoint or force a generic or global Gemini endpoint to process data in Qatar. A global model endpoint can process requests outside the requested geographic boundary, and the data-residency requirement must be satisfied by the location in which model inference actually runs. The application therefore needs a model deployment whose serving resources and endpoint are explicitly located in the Qatar region rather than relying on a security perimeter to change where an otherwise non-regional model executes.</p><blockquote><strong><em>B is correct because me-central1 is the Google Cloud region in Doha, Qatar, and Gemini Enterprise Agent Platform supports custom model online inference in that region. Model Garden provides deployable open models such as Gemma, and the organization can deploy a suitable Gemma model to a regional Agent Platform endpoint so inference runs on resources in the selected region. Google Cloud’s locational endpoint guidance states that regional endpoints provide geographic processing controls, while global endpoints do not provide regional data-residency guarantees. By using the Doha endpoint for every summarization request and avoiding global serving endpoints, the provider can use a pre-trained generative model while keeping the processing boundary aligned with the Qatar requirement. Additional controls such as resource-location organization policies, VPC Service Controls, CMEK where applicable, and strict IAM can strengthen governance but do not replace the regional deployment itself.</em></strong></blockquote><p>C is incorrect because the Cloud Natural Language API is intended for natural-language analysis tasks such as entities, sentiment, syntax, and content classification. It is not a configurable general-purpose generative summarization platform that can be instructed to produce free-form summaries of clinical notes in the way required by the scenario. It also does not solve the requirement to deploy the model’s inference workload specifically within the Doha region.</p><p>D is incorrect because custom model training in me-central1 could satisfy the geographic-processing constraint when all supporting resources and operations are correctly regionalized, but training a large summarization model from scratch is unnecessary when an appropriate pre-trained Gemma model is already available through Model Garden. Training from scratch requires a suitable labeled dataset, substantial accelerator capacity, model-development expertise, evaluation, tuning, and longer time to production. The requirement is to use a powerful pre-trained model, so deploying the existing model regionally is more direct and operationally efficient.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/locations">cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/locations</a></p><p><a href="http://cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/deploy-and-inference-tutorial">cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/deploy-and-inference-tutorial</a></p><p><a href="http://cloud.google.com/gemini-enterprise-agent-platform/resources/data-residency">cloud.google.com/gemini-enterprise-agent-platform/resources/data-residency</a></p><p><strong>Q19.</strong></p><p><strong>At Quoralis Digital, a fictional company, a product team is building a critical customer-facing application. The developers prefer Spanner because the application needs strong consistency, horizontal scalability, and low database-administration overhead. The FinOps team is concerned about Spanner’s direct monthly service cost and proposes running a self-managed PostgreSQL cluster on Compute Engine instead. You need to help the teams make an architecturally and financially sound decision.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Give the development team a reference architecture for a highly available PostgreSQL cluster on a regional managed instance group and proceed with that design.</strong></p><p><strong>B. Select Cloud SQL for PostgreSQL as an immediate compromise solely because its direct monthly price is expected to be lower than Spanner.</strong></p><p><strong>C. Develop a total-cost-of-ownership analysis that includes infrastructure charges, database administration, patching, scaling, backups, high availability, disaster recovery, security operations, staffing, downtime risk, and business value, and review the analysis jointly with the development and FinOps teams.</strong></p><p><strong>D. Cite the reliability and performance optimization pillars of the Google Cloud Well-Architected Framework and approve Spanner without performing a cost analysis.</strong></p><p>A is incorrect because a self-managed PostgreSQL architecture should not be selected before comparing whether it can satisfy the application’s technical requirements and whether its apparent infrastructure savings survive a complete cost analysis. Operating a production-grade PostgreSQL cluster on VMs requires database installation, patching, upgrades, replication, backups, failover, monitoring, security hardening, capacity planning, incident response, and specialist staffing. A reference design is useful as one candidate for comparison, but choosing it first ignores both the product team’s scalability requirements and the operational costs raised by a self-managed database.</p><p>B is incorrect because Cloud SQL can be an excellent managed PostgreSQL service and might be a valid candidate if the workload fits its scale, availability, consistency, and feature requirements. However, choosing it solely as a compromise on direct service price is not an architecture decision. The company first needs to quantify the workload requirements and compare the complete cost and business value of the viable options. A less expensive service that cannot meet the required horizontal scale or reliability would create higher downstream cost and risk.</p><blockquote><strong><em>C is correct because the Well-Architected cost-optimization guidance explicitly recommends evaluating total cost of ownership rather than comparing only resource line-item prices. A self-managed database can look inexpensive when only VM and disk charges are counted, but its TCO also includes engineering labor for patching, scaling, backups, security, high availability, disaster recovery, on-call operations, and outage risk. Spanner has a higher visible service price for some workloads but provides managed replication, scaling, availability, and operational automation that can reduce these indirect costs. A joint workshop based on quantified technical requirements, TCO, and business value lets FinOps and engineering make a transparent decision rather than allowing either direct price or technical preference to dominate by itself.</em></strong></blockquote><p>D is incorrect because reliability and performance are important Well-Architected concerns, but the framework also includes cost optimization and explicitly recommends balancing cost with business value and operational requirements. Simply citing two pillars to justify a predetermined choice does not resolve the FinOps concern and does not prove that Spanner is the best economic solution for this particular workload. The correct process is to evaluate all material costs and requirements together.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/architecture/framework/cost-optimization/align-cloud-spending-business-value">cloud.google.com/architecture/framework/cost-optimization/align-cloud-spending-business-value</a></p><p><a href="http://cloud.google.com/architecture/framework/cost-optimization">cloud.google.com/architecture/framework/cost-optimization</a></p><p><a href="http://cloud.google.com/spanner/pricing">cloud.google.com/spanner/pricing</a></p><p><strong>Q20.</strong></p><p><strong>At Torvexa Financial, a fictional financial-services company, you recently acquired a fintech startup. Its core system is a monolithic Python application running on a Compute Engine managed instance group with one large PostgreSQL database. Slow deployment cycles make it difficult to introduce new capabilities, including machine-learning-powered fraud detection. You need a long-term strategy that improves developer agility and positions the company to use Google Cloud data and AI capabilities for future innovation.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Leave the monolith intact, deploy only the fraud-detection model to a Gemini Enterprise Agent Platform endpoint, and modify the monolithic Python application to call the model’s REST endpoint.</strong></p><p><strong>B. Containerize the entire monolith, deploy it unchanged to GKE, and migrate PostgreSQL to Cloud SQL for PostgreSQL.</strong></p><p><strong>C. Implement a phased migration from the monolith to an event-driven microservices architecture, use Pub/Sub for asynchronous domain events, and deploy fraud-detection models to Gemini Enterprise Agent Platform inference endpoints.</strong></p><p><strong>D. Keep the monolithic application architecture, migrate PostgreSQL to Cloud SQL for PostgreSQL, replicate database changes into BigQuery with Datastream, and train and deploy the fraud-detection logic only through BigQuery ML.</strong></p><p>A is incorrect because adding a managed inference endpoint is a good way to integrate a fraud model, but it leaves the monolithic architecture and its slow deployment cycle fundamentally unchanged. Every new domain capability still has to be integrated, tested, and released through the same tightly coupled Python application. This can provide a useful short-term step, but it is not the long-term modernization strategy requested by the question.</p><p>B is incorrect because containers and GKE can improve packaging, deployment consistency, and runtime orchestration, and Cloud SQL reduces database administration. However, moving one large monolith into a container does not turn it into independently deployable services. The development team would still face coupling among modules and coordinated releases, so the principal agility problem remains. Kubernetes should not be treated as an architectural decomposition mechanism.</p><blockquote><strong><em>C is correct because a phased microservices migration lets the organization incrementally extract bounded business capabilities from the monolith instead of attempting a high-risk one-time rewrite. Independently deployable services can evolve at different speeds and can scale according to their own workloads. Pub/Sub provides asynchronous event delivery that reduces temporal coupling among services and supports event-driven integration with analytics and ML workflows. Fraud models can be versioned and deployed to Agent Platform inference endpoints so services consume them through a managed prediction interface. This approach improves developer autonomy while establishing architectural seams through which additional cloud-native data and AI capabilities can be introduced over time.</em></strong></blockquote><p>D is incorrect because Cloud SQL, Datastream, BigQuery, and BigQuery ML can form an excellent managed analytics and ML architecture, but this option leaves the application itself as the same tightly coupled monolith. It therefore does not directly solve the slow deployment-cycle and developer-agility problem that motivates the long-term strategy. It also constrains future ML integration to database-replicated analytical data rather than creating a broader event-driven platform.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/learn/what-is-microservices-architecture">cloud.google.com/learn/what-is-microservices-architecture</a></p><p><a href="http://cloud.google.com/pubsub/docs/overview">cloud.google.com/pubsub/docs/overview</a></p><p><a href="http://cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/deployment">cloud.google.com/gemini-enterprise-agent-platform/machine-learning/general/deployment</a></p><p><strong>Q21.</strong></p><p><strong>At Pryvanta Infrastructure, a fictional company, you are designing a centralized Terraform deployment process that runs through Cloud Build. The security team prohibits long-lived service account keys in CI/CD. Developers may propose infrastructure changes and participate in peer review, but they must not have permission to directly apply changes to the production project. You need an automated workflow with appropriate governance.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Configure Cloud Build to use service account impersonation, and automatically execute terraform apply as soon as a pull request is merged without a separate deployment approval.</strong></p><p><strong>B. Use short-lived service account impersonation or a least-privilege Cloud Build service account for the pipeline. Run terraform plan for proposed changes, and require an authorized manual approval gate before the production terraform apply step is allowed to run.</strong></p><p><strong>C. Configure the pipeline to run only terraform plan. After the pull request is approved, have an authorized developer run terraform apply manually from a secured workstation.</strong></p><p><strong>D. Create a privileged service account, store its JSON private key in Secret Manager, and have Cloud Build retrieve that long-lived key whenever Terraform is applied.</strong></p><p>A is incorrect because service account impersonation avoids long-lived static keys, but automatically applying every merged change removes the independent production-deployment control described in the requirement. Peer review of source code is valuable, yet a merge alone might not be sufficient authorization for a high-impact infrastructure change. Cloud Build supports approval-gated builds, allowing the organization to separate change proposal and review from final production execution.</p><blockquote><strong><em>B is correct because the workflow keeps privileged production credentials out of developers’ hands and out of source repositories. Cloud Build can execute with a tightly scoped build identity or use service account impersonation to obtain short-lived authority for the production operation instead of storing a private key. Terraform plan gives reviewers a concrete preview of resource changes before execution, while a Cloud Build approval gate can restrict the production apply to users holding the appropriate Cloud Build approval permission. This creates separation of duties: developers author and review the infrastructure code, while the controlled automation identity performs the actual change only after an authorized approval.</em></strong></blockquote><p>C is incorrect because this option prevents developers without production credentials from applying changes, but it removes the benefit of a fully automated and consistently audited production deployment. An operator running terraform apply from a workstation introduces differences in local Terraform versions, credentials, environment variables, network conditions, state access, and command usage. Keeping the apply inside the approved pipeline is more repeatable and easier to audit.</p><p>D is incorrect because storing a service account JSON key in Secret Manager protects the secret better than placing it in source control, but the key is still a long-lived credential, which the security requirement explicitly prohibits. If the key is extracted or misused, it remains valid until it is revoked. Cloud Build supports managed service accounts, user-specified service accounts, approval gates, and short-lived impersonation mechanisms, so a static private key is unnecessary.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/build/docs/securing-builds/gate-builds-on-approval">cloud.google.com/build/docs/securing-builds/gate-builds-on-approval</a></p><p><a href="http://cloud.google.com/build/docs/securing-builds/configure-user-specified-service-accounts">cloud.google.com/build/docs/securing-builds/configure-user-specified-service-accounts</a></p><p><a href="http://cloud.google.com/iam/docs/service-account-impersonation">cloud.google.com/iam/docs/service-account-impersonation</a></p><p><strong>Q22.</strong></p><p><strong>At Velarix Global Operations, a fictional company, you manage a highly distributed hybrid and multi-cloud environment. Developers already rely heavily on Prometheus and PromQL. You need a cloud-based enterprise monitoring solution that scales globally, requires little infrastructure maintenance, preserves PromQL workflows, and supports fast investigation and alerting.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Deploy and operate a Prometheus Operator and self-managed Prometheus infrastructure in every Kubernetes and serverless environment across all clouds.</strong></p><p><strong>B. Use only standard Cloud Monitoring system metrics as a single pane of glass and require developers to replace their PromQL queries with a different query language.</strong></p><p><strong>C. Use Google Cloud Managed Service for Prometheus to collect Prometheus-compatible metrics across the environments and query and alert on them with PromQL through Cloud Monitoring or compatible tools.</strong></p><p><strong>D. Build and operate a custom SaaS monitoring platform that implements enough of the Prometheus API to display metrics from each cloud.</strong></p><p>A is incorrect because a Prometheus Operator can simplify Kubernetes deployment of Prometheus, but the organization would still operate Prometheus servers, storage, sharding, federation, upgrades, availability, retention, and capacity across a large multi-cloud footprint. That directly conflicts with the requirement for a low-maintenance enterprise service. Self-managed Prometheus can remain appropriate for special edge or disconnected environments, but it should not be the primary global backend when a managed service is requested.</p><p>B is incorrect because Cloud Monitoring is an important part of the solution, but forcing developers to abandon PromQL would unnecessarily disrupt the workflows they already use. Cloud Monitoring itself supports PromQL, and Google Cloud Managed Service for Prometheus integrates Prometheus-compatible ingestion with Cloud Monitoring’s globally scalable backend. The company can preserve rather than replace existing Prometheus query expertise.</p><blockquote><strong><em>C is correct because Google Cloud Managed Service for Prometheus is a fully managed, multi-cloud and hybrid-compatible Prometheus monitoring solution built on the globally scalable Cloud Monitoring backend. It supports Prometheus exporters and OpenTelemetry, stores metrics centrally, and lets users continue querying with PromQL. Developers can use Cloud Monitoring, Grafana, or other Prometheus API-compatible tools, and alerting can also be expressed with PromQL. This preserves the open Prometheus interface while removing the need to operate a large federation and long-term storage platform manually.</em></strong></blockquote><p>D is incorrect because developing a custom monitoring SaaS would require the organization to build and operate ingestion, authentication, tenant isolation, storage, retention, global querying, PromQL compatibility, alert evaluation, upgrades, and high availability. That is substantially more engineering and operational effort than adopting the managed service that already provides these capabilities.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/managed-prometheus">cloud.google.com/managed-prometheus</a></p><p><a href="http://cloud.google.com/monitoring/promql">cloud.google.com/monitoring/promql</a></p><p><a href="http://cloud.google.com/monitoring/alerts/using-promql">cloud.google.com/monitoring/alerts/using-promql</a></p><p><strong>Q23.</strong></p><p><strong>At Brinovexa Systems, a fictional company, you are migrating a large on-premises application to Google Cloud. The application consists of several interconnected virtual machines. You want to create a detailed migration plan with minimal manual discovery effort. You need to understand the existing environment, identify application and network dependencies, and estimate the total cost of ownership for the target Google Cloud environment.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Write a custom script that queries the VMware vSphere API for virtual-machine information and imports the collected data into BigQuery for analysis.</strong></p><p><strong>B. Use Migration Center to perform automated discovery and assessment of the on-premises environment.</strong></p><p><strong>C. Use Config Connector to declare the desired state of the future Google Cloud resources in Kubernetes-style manifests.</strong></p><p><strong>D. Use the Google Cloud cost calculator to manually enter the specifications of the on-premises servers and use that estimate as the complete migration assessment.</strong></p><p>A is incorrect because a custom vSphere inventory script can collect some source-environment facts, but the company would need to design the data model, dependency analysis, utilization interpretation, target sizing logic, reporting, and TCO calculations itself. A VM inventory alone also does not provide the complete application and network dependency view required for a migration plan. Migration Center already provides managed discovery and assessment workflows and therefore avoids unnecessary custom engineering.</p><blockquote><strong><em>B is correct because Migration Center is the Google Cloud service designed for migration discovery, assessment, planning, and estimation. It can collect or import information about on-premises infrastructure, build a detailed asset inventory, identify application and network dependencies, group related assets, recommend target Google Cloud products and sizes, and generate total cost of ownership estimates. This directly addresses the need to understand the current environment before deciding migration waves and target architecture. The resulting assessment can then be refined with business criticality, application ownership, performance requirements, and migration risk.</em></strong></blockquote><p>C is incorrect because Config Connector is an infrastructure-management capability that represents Google Cloud resources as Kubernetes custom resources. It is useful after the desired target infrastructure has been designed, but it does not discover the existing vSphere estate, map workload dependencies, or create a TCO assessment for the migration.</p><p>D is incorrect because the Google Cloud cost calculator can estimate the price of a known target configuration, but manually entering server specifications does not provide automated discovery, application dependency mapping, utilization analysis, or migration recommendations. It can supplement migration planning after the target design is understood, but it is not a substitute for Migration Center’s discovery and assessment process.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/migration-center/docs/discovery-and-assessment-overview">cloud.google.com/migration-center/docs/discovery-and-assessment-overview</a></p><p><a href="http://cloud.google.com/migration-center/docs/migration-planning-overview">cloud.google.com/migration-center/docs/migration-planning-overview</a></p><p><a href="http://cloud.google.com/migration-center/docs/estimate/start-estimation">cloud.google.com/migration-center/docs/estimate/start-estimation</a></p><p><strong>Q24.</strong></p><p><strong>For this question, refer to the Altostrat Media case study. Altostrat still uses an on-premises file system for part of its content-ingestion workflow. New media files arrive throughout the day and must be copied to Cloud Storage repeatedly until the legacy system is retired. The operations team wants resumable transfers, centralized monitoring, and the ability to scale transfer throughput without writing and maintaining its own file-copy framework.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Mount the Cloud Storage bucket on the on-premises servers with Cloud Storage FUSE and make the legacy application treat the bucket as its primary POSIX file system.</strong></p><p><strong>B. Use Storage Transfer Service for file-system transfers, install transfer agents with access to the required source directories, and schedule recurring transfers to Cloud Storage.</strong></p><p><strong>C. Create a cron job that runs gcloud storage cp from every content server and maintain custom retry and checkpoint logic in shell scripts.</strong></p><p><strong>D. Ship a Transfer Appliance to the data center every week and use each appliance as the recurring ingestion mechanism.</strong></p><p>A is incorrect because Cloud Storage FUSE presents Cloud Storage through a file-system interface, but it does not turn object storage into a drop-in replacement for every POSIX application. Replacing the legacy application’s working file system would change its runtime storage semantics and could introduce differences in metadata, locking, rename, and latency behavior. The requirement is to move files reliably from the existing file system to Cloud Storage while the legacy ingestion system remains in service, not to redesign the legacy application’s storage layer.</p><blockquote><strong><em>B is correct because Storage Transfer Service supports agent-based transfers from POSIX-compatible file systems to Cloud Storage. Transfer agents run on machines that can access the source data, while the managed service coordinates jobs, retries, scheduling, progress, and scaling. Altostrat can add multiple agents to increase throughput and can restrict agent access to only the required directories. This lets the company modernize the ingestion path incrementally without rewriting the legacy producer or maintaining a custom file-transfer control plane.</em></strong></blockquote><p>C is incorrect because gcloud storage can copy objects, but a fleet of scheduled shell scripts would require Altostrat to build and operate its own discovery, retry, checkpointing, monitoring, concurrency, and error-recovery framework. Scripts running independently on many servers are also harder to observe consistently and can duplicate or miss work if hosts fail. A managed transfer service is a better fit for a recurring, production-scale migration path.</p><p>D is incorrect because Transfer Appliance is designed for offline movement of large datasets when network transfer is impractical. Repeating a physical appliance workflow every week would introduce shipping delays, handling effort, and operational complexity and would not provide the continuous or frequently scheduled ingestion process described. Transfer Appliance can be useful for an initial bulk migration, but it is not the preferred recurring synchronization mechanism.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/storage-transfer/docs/create-transfers/agent-based/file-system-to-cloud-storage">docs.cloud.google.com/storage-transfer/docs/create-transfers/agent-based/file-system-to-cloud-storage</a></p><p><a href="http://docs.cloud.google.com/storage-transfer/docs/on-prem-agent-best-practices">docs.cloud.google.com/storage-transfer/docs/on-prem-agent-best-practices</a></p><p><a href="http://docs.cloud.google.com/storage-transfer/docs/source-filesystem">docs.cloud.google.com/storage-transfer/docs/source-filesystem</a></p><p><strong>Q25.</strong></p><p><strong>For this question, refer to the Altostrat Media case study. As Altostrat expands media processing, GKE, BigQuery, Cloud Storage, and generative AI usage, finance wants to understand cloud cost by product, environment, and cost center. The operations team also wants automated notification when actual or forecasted monthly spend approaches approved limits.</strong></p><p><strong>Which two actions should you take? (Choose two.)</strong></p><p><strong>A. Enable Cloud Billing data export to BigQuery and apply a consistent resource labeling or tagging strategy so costs can be analyzed by product, environment, and cost center.</strong></p><p><strong>B. Create a separate Cloud Billing account for every GKE namespace and every Cloud Run function.</strong></p><p><strong>C. Use Cloud Logging ingestion volume as the authoritative source for calculating the cost of all Google Cloud services.</strong></p><p><strong>D. Create Cloud Billing budgets with actual and forecast threshold rules, and connect programmatic budget notifications to Pub/Sub when automated responses are required.</strong></p><p><strong>E. Purchase long-term commitments for all major services immediately, based only on the first week of usage.</strong></p><blockquote><strong><em>A is correct because Cloud Billing export provides detailed usage and cost data in BigQuery, which gives Altostrat a durable analytical source for FinOps reporting and trend analysis. Consistent labels and supported tags add organizational context to billing records, allowing finance to group or allocate costs by dimensions such as product, environment, and cost center. This approach scales across many projects and services and provides far better cost transparency than manual reconciliation.</em></strong></blockquote><p>B is incorrect because Cloud Billing accounts are not intended to be created at the granularity of Kubernetes namespaces or individual serverless functions. That design would create unnecessary billing administration and would not map cleanly to many Google Cloud resource types. Projects, folders, labels, tags, and exported billing data provide more appropriate mechanisms for cost allocation and governance.</p><p>C is incorrect because log ingestion volume represents only one observability cost dimension and has no direct one-to-one relationship with total charges for compute, storage, networking, analytics, AI, or other services. Cloud Billing is the authoritative source for billed usage and costs. Using logs as a substitute would produce incomplete and inaccurate financial reporting.</p><blockquote><strong><em>D is correct because Cloud Billing budgets can monitor actual and forecasted spend against thresholds and notify stakeholders before costs exceed planned limits. Programmatic notifications can be published to Pub/Sub, allowing Altostrat to integrate budgets with chat systems, ticketing, or carefully designed automated cost-control workflows. Budgets provide proactive cost visibility, while billing export provides the detailed data needed for analysis and allocation.</em></strong></blockquote><p>E is incorrect because commitments can reduce cost for predictable, sustained workloads, but Altostrat’s expanding media and AI workloads can have variable and evolving demand. Purchasing broad long-term commitments after only one week of observations risks overcommitting to resources that later become idle or change shape. Altostrat should first establish a stable usage baseline and evaluate service-specific commitment economics before making long-term commitments.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/billing/docs/how-to/export-data-bigquery">cloud.google.com/billing/docs/how-to/export-data-bigquery</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/budgets">cloud.google.com/billing/docs/how-to/budgets</a></p><p><a href="http://cloud.google.com/billing/docs/how-to/budgets-programmatic-notifications">cloud.google.com/billing/docs/how-to/budgets-programmatic-notifications</a></p><p><strong>Q26.</strong></p><p><strong>For this question, refer to the Altostrat Media case study. Altostrat currently uses event-driven functions for some video transcoding work. Media volume is growing, and the company wants to reduce custom transcoding infrastructure and reliably produce standardized renditions for web, mobile, and connected-TV playback, including HLS and MPEG-DASH outputs. Source and output media are stored in Cloud Storage.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Package FFmpeg in containers and run one GKE Job per asset, using cluster autoscaling and custom job-management code to control retries and rendition settings.</strong></p><p><strong>B. Submit asynchronous jobs to the Transcoder API using Cloud Storage input and output URIs, and use job templates for standardized rendition ladders.</strong></p><p><strong>C. Package FFmpeg in Cloud Run jobs and build a custom orchestration layer that starts a separate job for every rendition of every media asset.</strong></p><p><strong>D. Implement transcoding inside a Dataflow pipeline by invoking custom codec binaries from Dataflow workers.</strong></p><p>A is incorrect because GKE Jobs can run FFmpeg and GKE can scale worker capacity, so this is technically viable. However, Altostrat would still own the container images, codec patching, job orchestration, retry semantics, node sizing, accelerator decisions where applicable, and standardized rendition configuration. The requirement is to reduce custom transcoding infrastructure, which makes a purpose-built managed transcoding API the better choice.</p><blockquote><strong><em>B is correct because the Transcoder API is a managed media-processing service designed specifically to convert video assets and package them for delivery to web, mobile, and connected-TV clients. It reads from Cloud Storage, writes outputs back to Cloud Storage, supports formats including HLS and MPEG-DASH, and provides job templates so Altostrat can standardize codec and rendition settings. Compared with building its own transcoder scheduler on general-purpose compute, this reduces infrastructure, retry, worker-scaling, and media-pipeline management.</em></strong></blockquote><p>C is incorrect because Cloud Run jobs can execute batch containers without a permanent cluster and therefore reduce some infrastructure administration. However, Altostrat would still need to build the complete video-transcoding control plane: queueing, workflow state, codec packaging, rendition fan-out, retries, manifests, and output conventions. A managed media service already provides these capabilities directly.</p><p>D is incorrect because Dataflow is optimized for distributed batch and stream data-processing pipelines using the Apache Beam model. It is not a purpose-built media codec and packaging service. Embedding video transcoding binaries in Dataflow workers would create an unusual and operationally complex design and would not provide the standard media job and template abstractions available from the Transcoder API.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/transcoder/docs/concepts/overview">cloud.google.com/transcoder/docs/concepts/overview</a></p><p><a href="http://cloud.google.com/transcoder/docs/transcode-video">cloud.google.com/transcoder/docs/transcode-video</a></p><p><a href="http://cloud.google.com/transcoder/docs/concepts/supported-input-and-output-formats">cloud.google.com/transcoder/docs/concepts/supported-input-and-output-formats</a></p><p><strong>Q27.</strong></p><p><strong>For this question, refer to the Altostrat Media case study. Altostrat exposes customer-facing HTTP(S) APIs through a global external Application Load Balancer. Security teams want a centrally managed way to block common web-application attacks and reduce the impact of Layer 7 denial-of-service traffic before requests reach the GKE and serverless backends. Application teams should not have to implement the same filtering logic in every service.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Attach a Google Cloud Armor security policy to the Application Load Balancer and use preconfigured WAF rules and appropriate DDoS-protection capabilities.</strong></p><p><strong>B. Add host-based iptables rules to every GKE node and reproduce the same rules in each serverless application.</strong></p><p><strong>C. Create Cloud DNS response policies that return NXDOMAIN for clients that send malformed HTTP requests.</strong></p><p><strong>D. Export load-balancer logs to BigQuery and block attacking clients manually after analysts review the data.</strong></p><blockquote><strong><em>A is correct because Google Cloud Armor integrates with supported external Application Load Balancers and provides centrally managed security policies at Google’s edge. Altostrat can use preconfigured WAF rules based on industry rule sets to detect common web attacks and can apply additional rules, rate controls, and DDoS-related protections before traffic reaches the application backends. One policy can protect multiple services behind the load-balancing architecture, reducing duplicated security logic and lowering the amount of malicious traffic that consumes backend resources.</em></strong></blockquote><p>B is incorrect because host-level packet filters operate on the individual compute hosts and do not provide one consistent Layer 7 policy across GKE and serverless backends. They also cannot inspect HTTP requests with the same managed web-application firewall capabilities as a load-balancer security policy. Repeating local rules across workloads increases configuration drift and allows malicious traffic to consume load-balancer and backend resources before it is rejected.</p><p>C is incorrect because Cloud DNS response policies affect DNS resolution, not the content of HTTP requests after a client has resolved the application hostname. DNS cannot inspect SQL injection patterns, cross-site scripting payloads, request paths, or other Layer 7 characteristics of an established connection. It is therefore not the control plane for a web-application firewall.</p><p>D is incorrect because BigQuery is useful for historical security analysis, but a log-analysis workflow is reactive. By the time an analyst identifies a source and changes an application or firewall rule, the attack traffic has already reached the platform. The requirement is a preventive, centrally enforced control in the request path, which is what Cloud Armor supplies.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/armor/docs/overview">docs.cloud.google.com/armor/docs/overview</a></p><p><a href="http://docs.cloud.google.com/armor/docs/waf-rules">docs.cloud.google.com/armor/docs/waf-rules</a></p><p><a href="http://docs.cloud.google.com/load-balancing/docs/application-load-balancer">docs.cloud.google.com/load-balancing/docs/application-load-balancer</a></p><p><strong>Q28.</strong></p><p><strong>For this question, refer to the Cymbal Retail case study. Cymbal’s customer and catalog data is spread across existing MySQL and Microsoft SQL Server systems. The analytics team wants a near-real-time BigQuery view of changes so it can build a unified customer journey without running frequent full-table extracts against the operational databases. The source applications must continue using their existing databases during modernization.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Schedule a full export of every operational table to CSV once per hour and reload the complete files into BigQuery.</strong></p><p><strong>B. Give analysts direct read access to the production databases and have every dashboard query the operational systems in real time.</strong></p><p><strong>C. Migrate all source applications to BigQuery before the analytics team begins building the unified view.</strong></p><p><strong>D. Use Datastream change data capture from the supported MySQL and SQL Server sources and configure BigQuery as the destination for continuously applied changes.</strong></p><p>A is incorrect because repeated full-table exports perform substantially more source I/O and data transfer than capturing only inserts, updates, and deletes. As the catalog and customer datasets grow, hourly snapshots become slower, more expensive, and harder to reconcile. They also introduce up to an hour of planned freshness delay. Change data capture is a better fit when Cymbal wants to preserve the operational databases while making incremental changes available to analytics.</p><p>B is incorrect because running analytical dashboards directly against production databases couples reporting demand to transaction-processing capacity. Large scans and joins can contend with customer-facing workloads and create multiple inconsistent access patterns across MySQL and SQL Server. It also does not create the unified analytical layer Cymbal wants. BigQuery should serve analytical queries while the operational databases remain focused on application transactions.</p><p>C is incorrect because BigQuery is an analytical data warehouse and should not be treated as a universal replacement for transactional MySQL and SQL Server applications merely to create a reporting view. Rewriting all applications first would create a large, unnecessary modernization dependency and delay the analytics benefit. Cymbal explicitly needs an incremental approach that works while the existing systems continue operating.</p><blockquote><strong><em>D is correct because Datastream provides serverless change data capture for supported relational sources, including MySQL and SQL Server, and can write the streamed changes to BigQuery. BigQuery applies the CDC records so the analytical tables can reflect source changes with controlled staleness without repeated full extracts. This lets Cymbal create a unified analytical view while minimizing load on the operational systems and avoiding a prerequisite application rewrite.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/datastream/docs/configure-bigquery-destination">docs.cloud.google.com/datastream/docs/configure-bigquery-destination</a></p><p><a href="http://docs.cloud.google.com/datastream/docs/configure-self-managed-mysql">docs.cloud.google.com/datastream/docs/configure-self-managed-mysql</a></p><p><a href="http://docs.cloud.google.com/datastream/docs/configure-self-managed-sqlserver">docs.cloud.google.com/datastream/docs/configure-self-managed-sqlserver</a></p><p><strong>Q29.</strong></p><p><strong>For this question, refer to the EHR Healthcare case study. EHR Healthcare stores provider-level analytical data in shared BigQuery tables. Analysts who work for Provider A must see only rows that belong to Provider A, analysts for Provider B must see only Provider B rows, and EHR’s central analytics team must retain access to all providers. EHR wants to keep one physical table rather than create a copy for every provider.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Create BigQuery row access policies that filter rows by provider identity or group, and create an appropriate full-access policy for EHR’s central analytics principals.</strong></p><p><strong>B. Create one BigQuery dataset per analyst and copy the shared table into every dataset each night.</strong></p><p><strong>C. Grant all provider analysts BigQuery Data Viewer on the table and rely on each query to include the correct provider_id filter.</strong></p><p><strong>D. Encrypt each provider’s rows with a different Cloud KMS key inside the same BigQuery STRING column.</strong></p><blockquote><strong><em>A is correct because BigQuery row-level security applies row access policies that work as enforced filters for specified users or groups. EHR can create policies whose filter expressions select the rows associated with each provider and grant those policies to the appropriate provider identities. Principals that need complete access, such as EHR’s central analytics or data-maintenance identities, can receive a policy that exposes all rows. This provides fine-grained least-privilege access without physically duplicating the data.</em></strong></blockquote><p>B is incorrect because duplicating the table for every analyst or provider creates additional storage, synchronization, schema-management, and lifecycle work. New analysts and providers would require more copies, and updates would not be visible until the next replication run. BigQuery provides a native row-level authorization mechanism that can preserve a single source table.</p><p>C is incorrect because Data Viewer grants read access to the table data but does not force an analyst’s SQL to include the correct provider filter. A user could omit or change the WHERE clause and read another provider’s rows. The provider boundary must be enforced by the platform rather than by a query-writing convention.</p><p>D is incorrect because application-level encryption of individual rows would require custom encryption, key lookup, decryption, and query logic and would make normal analytical operations difficult. Cloud KMS is valuable for controlling encryption keys, but it is not a substitute for BigQuery’s native row authorization when users need ordinary SQL access to different subsets of the same table.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/row-level-security-intro">docs.cloud.google.com/bigquery/docs/row-level-security-intro</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/managing-row-level-security">docs.cloud.google.com/bigquery/docs/managing-row-level-security</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/best-practices-row-level-security">docs.cloud.google.com/bigquery/docs/best-practices-row-level-security</a></p><p><strong>Q30.</strong></p><p><strong>For this question, refer to the KnightMotives Automotive case study. Autonomous-driving engineers run large simulation campaigns before regulatory pilots. Each campaign contains thousands of independent tasks, tasks can safely restart after interruption, and no interactive service must remain online between campaigns. The team wants to minimize infrastructure administration and use lower-cost interruptible capacity when suitable.</strong></p><p><strong>What should you do?</strong></p><p><strong>A. Run the campaigns as Kubernetes Jobs on a permanently maintained GKE Standard cluster and operate custom controllers that provision and drain large Spot node pools for every campaign.</strong></p><p><strong>B. Use Batch to execute the simulation tasks and configure fault-tolerant jobs to use Spot VMs where appropriate.</strong></p><p><strong>C. Create a Compute Engine managed instance group, write a custom queue consumer that scales the group from simulation backlog, and maintain the worker lifecycle code.</strong></p><p><strong>D. Package each simulation as a Cloud Run job and split every campaign manually into independent job executions regardless of its compute, networking, and runtime requirements.</strong></p><p>A is incorrect because GKE Jobs and Spot node pools can run fault-tolerant simulations and may be appropriate when KnightMotives already needs Kubernetes-specific scheduling or platform features. However, the option requires maintaining a GKE Standard cluster and custom campaign-level capacity controls for a workload whose primary characteristic is independent batch tasks. This adds cluster administration compared with a managed batch scheduler.</p><blockquote><strong><em>B is correct because Batch is designed to queue, schedule, and execute batch and high-performance computing tasks while provisioning Compute Engine resources for the job lifecycle. Thousands of independent simulation tasks can run in parallel without a permanently running application tier. Because the tasks are explicitly restartable, Spot VMs can be used for eligible capacity to reduce cost, with retry and job design accounting for possible preemption. This minimizes custom scheduler and worker-fleet operations.</em></strong></blockquote><p>C is incorrect because a managed instance group can autoscale compute, but the team would still need to create and operate the queue consumer, map tasks to instances, manage retries and checkpoints, drain workers, and coordinate campaign completion. Batch provides those batch-job abstractions directly.</p><p>D is incorrect because Cloud Run jobs provide managed batch container execution and can be a good fit for many stateless jobs, but the option assumes that every autonomous simulation fits Cloud Run job runtime and resource characteristics and still requires the team to manually fragment large campaigns. Batch is purpose-built for broad compute-intensive batch and HPC scheduling with Compute Engine resource selection.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/batch/docs/get-started">docs.cloud.google.com/batch/docs/get-started</a></p><p><a href="http://docs.cloud.google.com/batch/docs/create-run-basic-job">docs.cloud.google.com/batch/docs/create-run-basic-job</a></p><p><a href="http://docs.cloud.google.com/batch/docs/improve-obtainability-overview">docs.cloud.google.com/batch/docs/improve-obtainability-overview</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e8795364ab5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Professional Cloud DevOps Engineer Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-professional-cloud-devops-engineer-practice-questions-03948d72d98c?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/03948d72d98c</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[gcp-certification]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 12:33:41 GMT</pubDate>
            <atom:updated>2026-10-05T12:33:41.545Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*se_PE6TIWWOs-5gsnMHaYA.jpeg" /><figcaption>KloudExams GCP Google Cloud Professional Cloud DevOps Engineer Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Professional Cloud DevOps Engineer certification</strong>? Knowing what automation, delivery, and monitoring tools do is a starting point — understanding how to use them to improve software delivery and maintain reliable production systems is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing delivery, observability, and reliability approaches that fit specific operational requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on DevOps practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>At Ardentiva Systems, you manage production deployments to a set of Google Kubernetes Engine (GKE) clusters. You want to ensure that only container images that were successfully built by your trusted CI/CD pipeline can be deployed to production. What should you do?</strong></p><p><strong>A. Enable Web Security Scanner for the applications running on the clusters.</strong></p><p><strong>B. Enable Artifact Analysis vulnerability scanning for the images stored in Artifact Registry.</strong></p><p><strong>C. Configure the GKE clusters as private clusters.</strong></p><p><strong>D. Configure the GKE clusters to enforce Binary Authorization policies.</strong></p><p>A is incorrect because Web Security Scanner is intended to identify certain classes of web-application vulnerabilities by crawling publicly accessible or otherwise reachable web applications. It does not establish software-supply-chain provenance for container images and does not provide a deploy-time admission decision proving that an image came from an approved CI/CD pipeline. Even if an application image passed a web vulnerability scan, that would not prove that the image was produced by the trusted build process.</p><p>B is incorrect because Artifact Analysis vulnerability scanning helps identify known vulnerabilities in container images and other supported artifacts, but vulnerability status and build provenance are different controls. An image could have no known vulnerabilities and still have been built outside the trusted CI/CD process. Vulnerability scanning therefore cannot by itself enforce the requirement that only images successfully produced by an approved pipeline are admitted to the production clusters.</p><p>C is incorrect because a private GKE cluster changes the network exposure of cluster nodes and, depending on configuration, the control plane. Private clusters are useful for reducing network exposure, but they do not validate where a container image was built or whether it passed the trusted CI/CD pipeline. A private cluster could still deploy an untrusted image unless an admission control mechanism enforces an image policy.</p><blockquote><strong><em>D is correct because Binary Authorization provides deploy-time software-supply-chain enforcement for GKE. You can configure a Binary Authorization policy to require an attestation proving that an image completed an approved build or verification stage before the image is admitted. A trusted CI/CD pipeline can create the required attestation after a successful build, and Binary Authorization can reject images that do not have the required attestation. When Cloud Build is the trusted builder, Binary Authorization can also use Cloud Build provenance-related controls such as the built-by-cloud-build attestor where applicable.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/overview">docs.cloud.google.com/binary-authorization/docs/overview</a></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/cloud-build">docs.cloud.google.com/binary-authorization/docs/cloud-build</a></p><p><a href="http://docs.cloud.google.com/binary-authorization/docs/attestations">docs.cloud.google.com/binary-authorization/docs/attestations</a></p><p><strong>Q2.</strong></p><p><strong>At Olyndra Systems, your company follows Site Reliability Engineering practices. You are the Incident Commander for a new customer-impacting incident. You need to immediately assign two incident-management roles to support an effective response. Which roles should you assign? (Choose two.)</strong></p><p><strong>A. Operations Lead.</strong></p><p><strong>B. Engineering Lead.</strong></p><p><strong>C. Communications Lead.</strong></p><p><strong>D. Customer Impact Assessor.</strong></p><p><strong>E. External Customer Communications Lead.</strong></p><blockquote><strong><em>A is correct because the Operations Lead is one of the primary incident-management roles in the Google SRE incident command model. The Operations Lead directs the hands-on technical response, coordinates responders who are investigating and mitigating the incident, tracks operational work, and reports status to the Incident Commander. Assigning this role lets the Incident Commander remain focused on command and coordination rather than becoming absorbed in individual debugging tasks.</em></strong></blockquote><p>B is incorrect because Engineering Lead is not one of the standard primary roles in the incident-management structure described by Google SRE for this situation. Engineers can be assigned technical tasks under the Operations Lead, and specialized subteams can be created as the incident grows, but the immediate core role is Operations Lead rather than a separate generic Engineering Lead.</p><blockquote><strong><em>C is correct because the Communications Lead is the other primary role needed to support the Incident Commander. This person maintains clear incident status, communicates with stakeholders, manages incoming requests for information, and ensures updates are consistent and timely. Separating communications from technical mitigation prevents the Incident Commander and Operations Lead from being overwhelmed by stakeholder communication during a major outage.</em></strong></blockquote><p>D is incorrect because assessing customer impact is an important activity, but Customer Impact Assessor is not one of the standard core incident command roles in the Google SRE model. Impact assessment can be delegated to responders or incorporated into the work of the command, operations, planning, or communications functions depending on the incident.</p><p>E is incorrect because external customer communication can be part of the Communications Lead’s responsibilities or can later be delegated to a specialized communications function if the incident is sufficiently large. However, External Customer Communications Lead is not one of the two immediate core roles specified by the standard SRE incident-management model. The Incident Commander should first establish the Operations Lead and Communications Lead functions.</p><p><strong>Reference links:</strong></p><p><a href="http://sre.google/sre-book/managing-incidents/">sre.google/sre-book/managing-incidents/</a></p><p><a href="http://sre.google/workbook/incident-response/">sre.google/workbook/incident-response/</a></p><p><a href="http://sre.google/sre-book/incident-document/">sre.google/sre-book/incident-document/</a></p><p><strong>Q3.</strong></p><p><strong>At Corvanta Labs, your team is designing a new application for deployment both inside and outside Google Cloud. You need to collect detailed application performance and resource-consumption information, such as CPU usage and memory allocation, and analyze the data centrally in Google Cloud while minimizing setup effort. What should you do?</strong></p><p><strong>A. Add the Cloud Profiler agent or library for the application’s supported runtime and configure it to send profiling data to Cloud Profiler for centralized analysis.</strong></p><p><strong>B. Add a debugging package to the application and configure the application to emit debug messages containing timing information.</strong></p><p><strong>C. Instrument the code with a timing library and publish the measurements through a health-check endpoint that is scraped by Cloud Monitoring.</strong></p><p><strong>D. Install a separate third-party Application Performance Monitoring (APM) platform in both locations and configure both installations to export data to a central data store for analysis.</strong></p><blockquote><strong><em>A is correct because Cloud Profiler is designed to continuously collect low-overhead production profiling information such as CPU time, wall time, heap usage, and memory allocation for supported runtimes. It supports applications running in Google Cloud as well as applications running outside Google Cloud. For external environments, the profiling agent is associated with a Google Cloud project and authenticated so it can upload profiles to the Cloud Profiler service, giving the team a centralized Google Cloud interface for analysis. This directly addresses the cross-environment requirement without requiring a separate APM stack and data store.</em></strong></blockquote><p>B is incorrect because debugging telemetry and profiling telemetry serve different purposes. Debuggers are primarily used to inspect application state and behavior while diagnosing defects; emitting debug messages with timing information would require custom instrumentation and would not automatically provide sampled CPU and memory-allocation profiles correlated with application code. It would also create additional logging volume and analysis work.</p><p>C is incorrect because a health-check endpoint is intended to report whether an application is healthy enough to receive traffic, not to act as a general-purpose metrics transport. Building custom timing instrumentation, designing an endpoint format, arranging for scraping, and maintaining the collection path requires more development effort than using a purpose-built profiler. It also would not automatically provide the detailed CPU and heap attribution that Cloud Profiler supplies.</p><p>D is incorrect because a third-party APM product could potentially collect the required data, but deploying and operating it in both environments and then configuring a separate centralized data destination adds more components and administrative work. Cloud Profiler already supports external environments and provides centralized analysis in a Google Cloud project, which better satisfies the requirement to minimize setup effort.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/profiler/docs/about-profiler">docs.cloud.google.com/profiler/docs/about-profiler</a></p><p><a href="http://docs.cloud.google.com/profiler/docs/profiling-external">docs.cloud.google.com/profiler/docs/profiling-external</a></p><p><a href="http://docs.cloud.google.com/profiler/docs/concepts-profiling">docs.cloud.google.com/profiler/docs/concepts-profiling</a></p><p><strong>Q4.</strong></p><p><strong>At Kelvoria Systems, you use Cloud Build to build your application. You want to reduce repeated build time while minimizing additional cost and development effort. What should you do?</strong></p><p><strong>A. Cache reusable intermediate build artifacts or dependency directories in Cloud Storage and restore them in subsequent builds.</strong></p><p><strong>B. Deploy and operate multiple Jenkins agents so builds can be parallelized outside Cloud Build.</strong></p><p><strong>C. Split the build into many smaller Cloud Build steps even when the steps still execute sequentially.</strong></p><p><strong>D. Use a larger Cloud Build machine type for every build.</strong></p><blockquote><strong><em>A is correct because Cloud Build documentation explicitly supports caching directories in Cloud Storage so the output of an earlier build can be reused by later builds. This is useful when expensive dependency or intermediate-artifact computation can be reused and the cached files are relatively inexpensive to transfer. It can shorten builds without requiring a continuously running build server or permanently paying for larger build machines. For supported buildpack workflows, remote cache images are another current caching mechanism, but Cloud Storage caching remains a valid general Cloud Build optimization.</em></strong></blockquote><p>B is incorrect because introducing Jenkins creates an additional CI platform to provision, secure, update, monitor, and operate. Parallelism can reduce some build durations, but adding a separate Jenkins fleet conflicts with the requirement to minimize cost and development effort when Cloud Build already provides managed execution and caching options.</p><p>C is incorrect because creating more build steps does not inherently reduce total execution time. Unless independent steps are explicitly arranged to run in parallel, dividing the same work into smaller sequential steps can add orchestration overhead while performing the same amount of computation. The repeated work that dominates build time should instead be cached or otherwise optimized.</p><p>D is incorrect because a larger Cloud Build machine can reduce the runtime of CPU- or memory-constrained builds, but it increases build resource cost and might provide little benefit if the dominant time is spent repeatedly downloading dependencies or recomputing unchanged artifacts. Since the requirement includes minimizing cost, reusing previous build results is generally the better first optimization among the options.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/build/docs/optimize-builds/speeding-up-builds">docs.cloud.google.com/build/docs/optimize-builds/speeding-up-builds</a></p><p><a href="http://docs.cloud.google.com/docs/buildpacks/cache-images">docs.cloud.google.com/docs/buildpacks/cache-images</a></p><p><strong>Q5.</strong></p><p><strong>At Norvella Group, you are designing the VPC architecture for a new Google Cloud organization. Initially only a few departments will use the organization, but it is expected to grow to many departments. Each department owns many applications of different sizes. You need an architecture that keeps departmental networking manageable while allowing departments to evolve their applications independently. Which option is best among the following choices?</strong></p><p><strong>A. Create a separate VPC network for each department and connect the VPC networks that require general private network connectivity by using VPC Network Peering.</strong></p><p><strong>B. Create a separate VPC network for each department and use Private Service Connect as though it were general full-mesh VPC-to-VPC routing.</strong></p><p><strong>C. Create a separate VPC network for every application and use Private Service Connect to connect all of the application networks.</strong></p><p><strong>D. Create a separate VPC network for each department and create Cloud VPN tunnels between the Google Cloud VPC networks for routine internal connectivity.</strong></p><blockquote><strong><em>A is correct because a departmental VPC provides a reasonably coarse administrative and network boundary without creating a separate network for every application. VPC Network Peering provides private IP connectivity between Google Cloud VPC networks without requiring VPN gateways, tunnels, or traffic to traverse the public internet. Among the listed choices, it therefore introduces the least unnecessary connectivity infrastructure while preserving departmental autonomy. For a very large modern organization, architects should also evaluate Shared VPC and Network Connectivity Center because a large peering mesh has scale and transitivity limitations, but those alternatives are not offered in this question.</em></strong></blockquote><p>B is incorrect because Private Service Connect is primarily a service-oriented connectivity mechanism. It is excellent when a producer wants to publish a specific service privately to consumers without exposing the producer’s entire network. Treating it as a drop-in replacement for general bidirectional VPC routing is not appropriate for a requirement that assumes broad connectivity among departmental networks.</p><p>C is incorrect because one VPC per application creates a very large number of networks, policies, address plans, routes, and connectivity relationships. That substantially increases management overhead as the organization grows. Private Service Connect can be useful for selectively publishing individual services across network boundaries, but creating a separate VPC for every application is unnecessarily granular for the requirements given.</p><p>D is incorrect because Cloud VPN is principally used to create encrypted IPsec connectivity between separate networks, including hybrid or cross-cloud environments. Using VPN tunnels merely to interconnect VPC networks that are already on Google Cloud creates gateways, tunnels, availability design, routing configuration, and additional management that VPC Network Peering avoids for this use case.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/vpc/docs/vpc-peering">docs.cloud.google.com/vpc/docs/vpc-peering</a></p><p><a href="http://docs.cloud.google.com/vpc/docs/private-service-connect">docs.cloud.google.com/vpc/docs/private-service-connect</a></p><p><a href="http://docs.cloud.google.com/network-connectivity/docs/vpn/concepts/overview">docs.cloud.google.com/network-connectivity/docs/vpn/concepts/overview</a></p><p><strong>Q6.</strong></p><p><strong>At Orlivena Commerce, you have migrated an e-commerce application to Google Cloud and need to prepare for an upcoming high-traffic shopping season. You want to make evidence-based scaling and capacity decisions. What should you do first?</strong></p><p><strong>A. Load test the application at representative and expected peak traffic levels to identify bottlenecks, validate performance, and establish the capacity and scaling behavior you need.</strong></p><p><strong>B. Enable autoscaling on the production compute platforms without first validating how the application behaves under peak load.</strong></p><p><strong>C. Pre-provision exactly twice the compute capacity used during the previous season without testing the current application.</strong></p><p><strong>D. Write a runbook for manually expanding the disaster-recovery environment if demand increases, and use that as the primary capacity strategy.</strong></p><blockquote><strong><em>A is correct because the first step in peak-event preparation is to understand how the current application behaves under realistic load. A load test can expose bottlenecks in compute, databases, caches, networking, external dependencies, quotas, connection pools, and application code, and it can verify whether configured autoscaling reacts quickly enough. The results provide evidence for setting autoscaling limits, requesting quota increases, pre-warming resources, and deciding whether extra fixed capacity is necessary. Historical traffic should inform the test scenario, but the current system must be validated rather than assumed to behave like last season’s system.</em></strong></blockquote><p>B is incorrect because autoscaling is an important part of peak-event readiness, but enabling it without load testing does not prove that the application will scale correctly. Autoscaling can be constrained by quotas, maximum replica or instance settings, startup time, database bottlenecks, or a scaling metric that responds too slowly. Google Cloud peak-event guidance also notes that rapid event traffic can sometimes rise faster than autoscaling can react, so testing and, where appropriate, pre-warming are necessary.</p><p>C is incorrect because doubling last season’s compute capacity is an arbitrary forecast and assumes compute is the only bottleneck. The application, traffic mix, database behavior, quotas, and architecture may all have changed. The system could require much less than twice the capacity, wasting money, or more than twice the capacity, still failing under load. Load-test results provide a more defensible basis for provisioning.</p><p>D is incorrect because a disaster-recovery runbook is useful for resilience, but manually expanding a DR environment is not a substitute for understanding primary-system capacity and scaling behavior before a predictable peak event. Capacity preparation should be proactive and validated in advance rather than depending primarily on manual emergency expansion after traffic has already increased.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/support/docs/peak-events/optimize-resources">docs.cloud.google.com/support/docs/peak-events/optimize-resources</a></p><p><a href="http://docs.cloud.google.com/architecture/framework/operational-excellence/operational-readiness-and-performance-using-cloudops">docs.cloud.google.com/architecture/framework/operational-excellence/operational-readiness-and-performance-using-cloudops</a></p><p><a href="http://docs.cloud.google.com/compute/docs/tutorials/globally-autoscaling-a-web-service-on-compute-engine">docs.cloud.google.com/compute/docs/tutorials/globally-autoscaling-a-web-service-on-compute-engine</a></p><p><strong>Q7.</strong></p><p><strong>At Cendrixa Government Services, a regulatory requirement requires you to archive application logs for seven years. You need to route logs from Cloud Logging to long-term storage while minimizing storage cost. What should you do?</strong></p><p><strong>A. Create a Cloud Storage bucket and modify each application so that it sends its logs directly to the bucket.</strong></p><p><strong>B. Develop an App Engine application that continuously reads logs from Cloud Logging and writes them to BigQuery.</strong></p><p><strong>C. Create a Cloud Logging sink that routes logs to Pub/Sub and use Pub/Sub itself as the permanent seven-year archive.</strong></p><p><strong>D. Create a Cloud Logging sink that routes the required logs to a Cloud Storage bucket configured with an appropriate seven-year retention policy and long-term storage strategy.</strong></p><p>A is incorrect because modifying each application to write its own archival copy bypasses the centralized logging pipeline and creates application-specific code, authentication, retry, batching, and failure-handling logic. The applications are already producing logs that reach Cloud Logging, whose Log Router can route matching entries to supported destinations without changing the application. Duplicating the export responsibility inside every workload increases development and operational effort.</p><p>B is incorrect because BigQuery is designed for interactive analytics and large-scale querying, not primarily for the lowest-cost archival of log objects that only need to be retained for seven years. A continuously running App Engine exporter would also duplicate functionality already provided by Cloud Logging sinks. BigQuery is appropriate if the organization needs frequent SQL analysis of the retained logs, but it is unnecessary for simple long-term archival.</p><p>C is incorrect because Pub/Sub is a messaging service and is not intended to be a seven-year permanent archive. Messages are retained for a bounded period according to Pub/Sub retention configuration and are expected to be consumed by subscribers. Pub/Sub can be an intermediate transport in a log-processing architecture, but another durable system would still be required for a seven-year compliance archive.</p><blockquote><strong><em>D is correct because Cloud Logging sinks can route matching log entries directly to Cloud Storage. Cloud Storage is suitable for long-term archival and supports bucket retention policies that can prevent objects from being deleted before the required retention interval has elapsed. If compliance requires immutability, the organization can carefully evaluate locking the retention policy because that action is irreversible. Storage-class and lifecycle configuration can also be used to reduce long-term storage cost for logs that are accessed infrequently. This centralized design avoids changing each application and provides a durable retention control for the required seven-year period.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/export/storage">docs.cloud.google.com/logging/docs/export/storage</a></p><p><a href="http://docs.cloud.google.com/logging/docs/routing/overview">docs.cloud.google.com/logging/docs/routing/overview</a></p><p><a href="http://docs.cloud.google.com/storage/docs/bucket-lock">docs.cloud.google.com/storage/docs/bucket-lock</a></p><p><strong>Q8.</strong></p><p><strong>At Heliovex Media, some production services run in a GKE cluster in the europe-west1 region. Your build system runs in us-west1. You want to push container images from the build system to a scalable managed registry and optimize image-transfer latency and bandwidth for the production GKE workloads that pull those images. What should you do?</strong></p><p><strong>A. Push the images to an Artifact Registry repository in the us multi-region.</strong></p><p><strong>B. Push the images to an Artifact Registry repository in us-west1.</strong></p><p><strong>C. Push the images to an Artifact Registry repository in europe-west1.</strong></p><p><strong>D. Push the images to a private container registry running on a single Compute Engine instance in europe-west1.</strong></p><p>A is incorrect because an Artifact Registry repository in the us multi-region places the artifacts near the build system but far from the production GKE cluster in Europe. The build system pushes each image version relatively infrequently compared with the potentially many image pulls performed by production nodes during deployments, scaling, repairs, and node replacement. Locating the repository near the consumers is therefore the better optimization for production image delivery.</p><p>B is incorrect because placing the repository in us-west1 co-locates it with the build system but causes the GKE workloads in europe-west1 to pull images across geographic regions. Google Cloud recommends choosing an Artifact Registry location close to the services that consume the artifacts to reduce latency and, where applicable, data-transfer cost. The production cluster is the important recurring consumer in this scenario.</p><blockquote><strong><em>C is correct because Artifact Registry is the current managed Google Cloud container registry, and creating the repository in europe-west1 places the image repository in the same region as the GKE nodes that repeatedly pull the images. Google Cloud documentation recommends co-locating Artifact Registry repositories with GKE and other consumers where practical because this reduces latency and can avoid same-region data-transfer charges. It also supports GKE image-streaming locality requirements when that capability is used.</em></strong></blockquote><p>D is incorrect because a self-managed registry on one Compute Engine VM is not a scalable or highly available registry architecture. The team would need to manage storage, availability, backups, upgrades, TLS, authentication, capacity, and recovery. Artifact Registry already provides a managed and scalable repository service in europe-west1 and therefore better satisfies the requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/repositories/repo-locations">docs.cloud.google.com/artifact-registry/docs/repositories/repo-locations</a></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/repositories">docs.cloud.google.com/artifact-registry/docs/repositories</a></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/integrate-gke">docs.cloud.google.com/artifact-registry/docs/integrate-gke</a></p><p><strong>Q9.</strong></p><p><strong>At Oryvanta Platforms, you are building and deploying a microservice on Cloud Run. Many internal applications depend on the service. You are releasing a new version and need to test it extensively in both staging and production while minimizing user and developer impact. What should you do?</strong></p><p><strong>A. Deploy the new revision to staging and immediately send 1% of ordinary staging traffic to it before testing. If the test passes, gradually roll it out in staging and production.</strong></p><p><strong>B. Deploy the new revision to staging and immediately send 50% of ordinary traffic to it. If the test passes, move all staging traffic to the revision and repeat the same 50% procedure in production.</strong></p><p><strong>C. Deploy the new revision in staging with a revision tag and zero normal traffic. Test the tagged revision directly. If the tests pass, gradually migrate staging traffic to it. Repeat the zero-traffic tagged test and gradual rollout in production.</strong></p><p><strong>D. Create a completely new Cloud Run environment with a green tag, direct all staging traffic to it immediately after testing, delete the existing staging environment, and repeat the process for production.</strong></p><p>A is incorrect because sending even 1% of ordinary user or developer traffic to a new revision before completing the desired extensive testing creates avoidable impact. Cloud Run provides revision tags specifically so a zero-traffic revision can be reached through its tag URL for validation before it receives normal service traffic. The team should use that capability first.</p><p>B is incorrect because immediately exposing half of the environment’s normal traffic to an unvalidated revision creates a much larger blast radius than necessary. Cloud Run supports both zero-traffic tagged revisions and arbitrary gradual traffic migration, so there is no reason to start testing by putting 50% of dependent applications onto the new release.</p><blockquote><strong><em>C is correct because Cloud Run supports deploying a revision with no normal traffic while assigning it a tag. The tag gives testers a dedicated URL for the new revision even though the revision has a zero-percent traffic allocation. Google Cloud documentation describes this exact sequence for testing: deploy and test a tagged zero-traffic revision in staging, deploy and test the revision with zero traffic in production, and then gradually migrate traffic after validation. This minimizes impact while still testing the exact production revision before general users reach it.</em></strong></blockquote><p>D is incorrect because Cloud Run revisions already provide the isolation and traffic-management primitives needed for this release process. Creating and deleting whole environments for every release adds service configuration, IAM, networking, DNS, and lifecycle complexity. In addition, moving 100% of traffic at once after testing eliminates the risk reduction of a gradual rollout.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/rollouts-rollbacks-traffic-migration">docs.cloud.google.com/run/docs/rollouts-rollbacks-traffic-migration</a></p><p><a href="http://docs.cloud.google.com/run/docs/managing/revisions">docs.cloud.google.com/run/docs/managing/revisions</a></p><p><a href="http://docs.cloud.google.com/run/docs/samples/cloudrun-service-deploy-tag">docs.cloud.google.com/run/docs/samples/cloudrun-service-deploy-tag</a></p><p><strong>Q10.</strong></p><p><strong>At Eryndex Platforms, your teams deploy applications to three Google Kubernetes Engine (GKE) environments: development, staging, and production. GitHub repositories are the source of truth. You need to ensure that all three environments remain consistent and want to follow Google-recommended practices to install and continuously enforce network policies and a logging DaemonSet on every GKE cluster. What should you do?</strong></p><p><strong>A. Use Cloud Deploy to deploy the network policies and the DaemonSet. Use Cloud Monitoring to alert when those resources drift from the repository.</strong></p><p><strong>B. Use Cloud Deploy to deploy the DaemonSet and use Policy Controller to configure the network policies. Use Cloud Monitoring to detect drift and Cloud Run functions to correct the drift.</strong></p><p><strong>C. Use Cloud Build to render and deploy the network policies and the DaemonSet. Set up Config Sync to sync the configurations for the three environments.</strong></p><p><strong>D. Use Cloud Build to render and deploy the network policies and the DaemonSet. Use Policy Controller to enforce those exact configurations across all three environments.</strong></p><p>A is incorrect because Cloud Deploy is designed for application delivery and promotion through deployment targets, while Cloud Monitoring can detect many operational conditions but does not itself reconcile Kubernetes objects to match a Git repository. Alerting after configuration drift occurs still requires another process to restore the desired state. A GitOps reconciler is a better fit when Git is explicitly the source of truth and clusters must remain continuously synchronized.</p><p>B is incorrect because Policy Controller is designed to audit and enforce programmable Kubernetes policy constraints, not to act as the primary Git synchronization mechanism for arbitrary resources such as a logging DaemonSet. Adding Cloud Monitoring and Cloud Run functions to detect and repair drift creates custom remediation logic that Config Sync already provides as a managed GitOps capability.</p><blockquote><strong><em>C is correct because Config Sync is the GKE GitOps service for continuously reconciling cluster configuration with a central source of truth such as Git. It can manage Kubernetes resources including network-related configuration and DaemonSets across multiple clusters and continuously correct configuration drift. Cloud Build can be used where configuration rendering or validation is needed before artifacts are placed in the source that Config Sync consumes. This combination keeps Git authoritative and provides continuous reconciliation across development, staging, and production.</em></strong></blockquote><p>D is incorrect because Policy Controller can enforce rules and compliance constraints, but it does not by itself synchronize arbitrary desired Kubernetes manifests from Git or guarantee that a particular DaemonSet matches a repository version. Policy Controller and Config Sync can complement each other, but Config Sync is the component designed specifically for source-of-truth synchronization and drift reconciliation.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/config-sync/docs/overview">docs.cloud.google.com/kubernetes-engine/config-sync/docs/overview</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/config-sync/docs/tutorials/config-sync-multi-repo">docs.cloud.google.com/kubernetes-engine/config-sync/docs/tutorials/config-sync-multi-repo</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/policy-controller/docs/overview">docs.cloud.google.com/kubernetes-engine/policy-controller/docs/overview</a></p><p><strong>Q11.</strong></p><p><strong>At Galendra Security, your Google Cloud organization contains a Dev folder for development projects and a Prod folder for production projects. You need to route Cloud Logging entries from all current and future Dev projects to a BigQuery dataset named dev_dataset and entries from all current and future Prod projects to a BigQuery dataset named prod_dataset. You want to minimize the number of log sinks. What should you do?</strong></p><p><strong>A. Create one aggregated log sink at the organization level.</strong></p><p><strong>B. Create one log sink in every individual project.</strong></p><p><strong>C. Create two aggregated log sinks at the organization level and filter each sink by the IDs of the existing projects.</strong></p><p><strong>D. Create one aggregated log sink on the Dev folder and one aggregated log sink on the Prod folder, including child resources, and route each sink to its corresponding BigQuery dataset.</strong></p><p>A is incorrect because a single sink has one configured destination. The requirement has two independent destination datasets based on folder membership, so one organization-level sink cannot directly route Dev entries to dev_dataset and Prod entries to prod_dataset as two separate destinations. Additional routing logic or another sink would still be required.</p><p>B is incorrect because project-level sinks would need to be created and maintained separately for every project. That increases administration and, more importantly, would not automatically cover future projects created under the folders unless the provisioning process also creates a new sink every time. Aggregated folder sinks are intended to route logs from a folder and its descendants centrally.</p><p>C is incorrect because two organization-level sinks could technically separate traffic through filters, but filtering by the IDs of existing projects does not naturally include projects created later. The filters would have to be updated whenever the organization adds another development or production project. Using the resource hierarchy itself as the routing scope is simpler and automatically covers descendants.</p><blockquote><strong><em>D is correct because a folder-level aggregated sink can route log entries that originate in the folder and its descendant projects. Creating one aggregated sink on Dev and one on Prod naturally matches the organization hierarchy, minimizes the number of sinks to two, and includes future child projects without maintaining a list of project IDs. Each sink can route its selected log entries directly to the appropriate BigQuery dataset after the sink writer identity is granted permission on that destination.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/export/aggregated_sinks_overview">docs.cloud.google.com/logging/docs/export/aggregated_sinks_overview</a></p><p><a href="http://docs.cloud.google.com/logging/docs/export/bigquery">docs.cloud.google.com/logging/docs/export/bigquery</a></p><p><a href="http://docs.cloud.google.com/logging/docs/routing/overview">docs.cloud.google.com/logging/docs/routing/overview</a></p><p><strong>Q12.</strong></p><p><strong>At Orivexia Systems, your team is updating the postmortem policy for an external-facing application. You want to define objective and appropriate triggers that indicate when an incident requires a postmortem, following Site Reliability Engineering practices. Which two triggers should be included? (Choose two.)</strong></p><p><strong>A. An external stakeholder requests a postmortem.</strong></p><p><strong>B. Data is lost because of an incident.</strong></p><p><strong>C. An internal stakeholder requests a postmortem.</strong></p><p><strong>D. Monitoring detects the failure of one application instance, even when the service remains healthy and no significant incident occurs.</strong></p><p><strong>E. The continuous-delivery pipeline detects a bad release and automatically rolls it back before it creates a significant incident or requires on-call intervention.</strong></p><blockquote><strong><em>A is correct because Google SRE guidance explicitly states that, in addition to objective postmortem criteria, any stakeholder may request a postmortem for an event. An external stakeholder request is therefore a valid trigger. A stakeholder might have business, contractual, dependency, or user-impact context that makes formal analysis valuable even if another threshold was not crossed.</em></strong></blockquote><blockquote><strong><em>B is correct because data loss of any kind is explicitly listed as a common Google SRE postmortem trigger. Data loss can indicate serious weaknesses in durability, recovery, validation, or operational controls even if the affected service was restored quickly. A postmortem provides the structured analysis and corrective actions needed to understand the loss and reduce the probability or impact of recurrence.</em></strong></blockquote><p>C is incorrect because the wording makes this appear similar to option A, and Google SRE broadly says that any stakeholder can request a postmortem. However, the exam asks for two triggers and option A already represents the stakeholder-request criterion. The stronger independent objective criterion among the remaining choices is data loss. Selecting both A and C would duplicate the same stakeholder-request principle while omitting a specifically enumerated technical trigger.</p><p>D is incorrect because the failure of one instance is not automatically a significant incident. Well-designed distributed services routinely tolerate individual instance failures through redundancy, health checking, and rescheduling. A postmortem policy should define meaningful thresholds such as user-visible degradation, data loss, significant response time, on-call intervention, or monitoring failure, rather than requiring a full postmortem for every healthy failover event.</p><p>E is incorrect because an automated rollback that detects and contains a bad release before significant user impact or human intervention can be evidence that the delivery safeguards are working as intended. Google SRE does list on-call intervention such as a release rollback as a common postmortem trigger, but this option specifically describes an automatic pipeline rollback before a significant incident or on-call intervention. Teams can still review such events, but the scenario does not make every successfully contained automated rollback a mandatory postmortem trigger.</p><p><strong>Reference links:</strong></p><p><a href="http://sre.google/sre-book/postmortem-culture/">sre.google/sre-book/postmortem-culture/</a></p><p><a href="http://sre.google/workbook/postmortem-culture/">sre.google/workbook/postmortem-culture/</a></p><p><a href="http://docs.cloud.google.com/architecture/framework/reliability/conduct-postmortems">docs.cloud.google.com/architecture/framework/reliability/conduct-postmortems</a></p><p><strong>Q13.</strong></p><p><strong>At Torvessa Labs, your application’s performance on Google Cloud has degraded since the latest release. You suspect that downstream dependencies are causing some requests to take longer to complete. You need to determine which dependency or operation is contributing to the latency. What should you do?</strong></p><p><strong>A. Configure Error Reporting in the application.</strong></p><p><strong>B. Configure Google Cloud Managed Service for Prometheus in the application.</strong></p><p><strong>C. Configure Cloud Profiler in the application.</strong></p><p><strong>D. Configure Cloud Trace in the application and instrument the relevant service calls with distributed tracing.</strong></p><p>A is incorrect because Error Reporting groups and analyzes application errors and exceptions. It is useful for identifying recurring failures and stack traces, but a slow request can complete successfully without producing an exception. Error Reporting therefore does not reconstruct the timing of calls to downstream services or show where latency accumulated in an end-to-end request.</p><p>B is incorrect because Google Cloud Managed Service for Prometheus is a scalable metrics backend and can show service-level trends such as request latency, error rate, CPU usage, and saturation. Metrics can indicate that a dependency is generally unhealthy, but aggregated time-series data does not inherently reconstruct the individual sequence of downstream operations for a particular request. Distributed tracing is better suited to determining which dependency span delayed an end-to-end request.</p><p>C is incorrect because Cloud Profiler continuously samples application CPU usage, wall time, heap allocation, and similar runtime behavior to identify expensive code paths inside an application. It is valuable when slow performance is caused by code execution or resource usage, but it does not primarily correlate a request across several downstream services and display the latency contribution of each remote call.</p><blockquote><strong><em>D is correct because Cloud Trace is Google Cloud’s distributed tracing system. A trace represents an end-to-end request, and its spans represent the individual operations or downstream calls performed during that request. When trace context is propagated across service boundaries, Cloud Trace can display the duration and hierarchy of those spans, allowing the team to identify the dependency or operation responsible for the additional latency introduced after the release. OpenTelemetry is the current standards-based instrumentation approach for generating and exporting trace data.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/trace/docs">docs.cloud.google.com/trace/docs</a></p><p><a href="http://docs.cloud.google.com/trace/docs/traces-and-spans">docs.cloud.google.com/trace/docs/traces-and-spans</a></p><p><a href="http://docs.cloud.google.com/trace/docs/setup">docs.cloud.google.com/trace/docs/setup</a></p><p><strong>Q14.</strong></p><p><strong>At Halcyonix Finance, your company operates in a highly regulated industry and must retain all Google Cloud organization logs for seven years. You want to use managed services, minimize logging infrastructure complexity, automatically cover current and future projects, and protect retained logs from accidental deletion or retention-policy changes. What should you do?</strong></p><p><strong>A. Configure an organization-level aggregated Cloud Logging sink that routes all logs to a BigQuery dataset.</strong></p><p><strong>B. Configure an organization-level aggregated Cloud Logging sink that includes descendant resources and routes all required logs to a Cloud Storage bucket with a seven-year retention policy, and lock the retention policy with Bucket Lock.</strong></p><p><strong>C. Configure a separate Cloud Logging sink in every project and route all logs to a BigQuery dataset.</strong></p><p><strong>D. Configure a separate Cloud Logging sink in every project and route all logs to Cloud Storage buckets with seven-year retention policies and Bucket Lock.</strong></p><p>A is incorrect because an organization-level aggregated sink appropriately centralizes logs, but BigQuery is primarily an analytics platform and does not provide the Cloud Storage Bucket Lock control specified for immutable object retention. If the principal requirement is a long-term compliance archive protected from accidental deletion or reduction of its retention period, Cloud Storage with a locked retention policy is the stronger fit among the available options.</p><blockquote><strong><em>B is correct because an organization-level aggregated sink with descendant resources enabled can route matching log entries from the organization and its child folders and projects to one managed destination, including projects created later under the hierarchy. Cloud Storage is suitable for long-term log archival. A seven-year bucket retention policy prevents covered objects from being deleted or replaced before the retention period expires, and Bucket Lock permanently prevents that retention policy from being removed or shortened. This combination minimizes the number of sinks and provides strong protection against human error in the retained archive.</em></strong></blockquote><p>C is incorrect because project-level sinks must be created and maintained for every project, which increases configuration complexity and creates a risk that a new project is omitted from the compliance export. BigQuery also does not provide the same immutable bucket-retention mechanism as Cloud Storage Bucket Lock for this archival requirement.</p><p>D is incorrect because Cloud Storage with locked seven-year retention is appropriate for the archive itself, but creating an export sink separately in every project is unnecessary and operationally fragile. New projects could be created without a sink, and hundreds of independent configurations create more opportunities for misconfiguration. An organization-level aggregated sink provides centralized coverage with fewer moving parts.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/export/aggregated_sinks_overview">docs.cloud.google.com/logging/docs/export/aggregated_sinks_overview</a></p><p><a href="http://docs.cloud.google.com/logging/docs/export/storage">docs.cloud.google.com/logging/docs/export/storage</a></p><p><a href="http://docs.cloud.google.com/storage/docs/bucket-lock">docs.cloud.google.com/storage/docs/bucket-lock</a></p><p><strong>Q15.</strong></p><p><strong>At Klyvora Operations, your organization routes application logs from multiple Google Cloud projects into a central Cloud Logging project. The security team requires each project team to view only its own application’s logs, while the central operations team must be able to view all logs. You need to meet the requirement while minimizing additional data-storage cost. What should you do?</strong></p><p><strong>A. Grant every project team access to the _Default view in the central logging project. Grant Logs Viewer access to the operations team in the central logging project.</strong></p><p><strong>B. Create IAM roles for each project team and restrict access to the _Default log view in the team’s original Google Cloud project. Grant viewer access to the operations team in the central logging project.</strong></p><p><strong>C. Create a separate log view in the central log bucket for each project team’s application logs. Grant each team access only to its corresponding view, and grant the operations team access to the _AllLogs view.</strong></p><p><strong>D. Export the centralized logs into separate BigQuery tables for each project team. Grant each team access to its table, and grant the operations team access to all exported tables.</strong></p><p>A is incorrect because the _Default view is not a per-source-project isolation boundary. Granting every team access to the same central _Default view can expose log entries from other projects that are included in that view. The requirement calls for granular separation between teams, so each team needs a filtered view that contains only its own permitted entries.</p><p>B is incorrect because the logs have been centralized into the logging project, so controlling access only to log views in the original source projects does not govern access to the central copies. The security boundary needs to be implemented where the centralized logs are stored and queried.</p><blockquote><strong><em>C is correct because Cloud Logging log views are designed to provide granular access to subsets of the entries stored in a log bucket. A central organization can create a view filtered by source project or another reliable attribute for each team, then grant that team roles/logging.viewAccessor specifically on its view. Cloud Logging automatically provides an _AllLogs view for a log bucket, so the operations team can be granted access to that full view. This meets the isolation requirement without duplicating the data into another analytics service.</em></strong></blockquote><p>D is incorrect because BigQuery can implement row-, table-, or dataset-level access control, but exporting another copy of the logs increases storage and query-management overhead when Cloud Logging already provides log views for this purpose. The requirement asks to minimize cost, so using native views over the existing centralized log bucket is preferable.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/logs-views">docs.cloud.google.com/logging/docs/logs-views</a></p><p><a href="http://docs.cloud.google.com/logging/docs/central-log-storage">docs.cloud.google.com/logging/docs/central-log-storage</a></p><p><a href="http://docs.cloud.google.com/logging/docs/access-control">docs.cloud.google.com/logging/docs/access-control</a></p><p><strong>Q16.</strong></p><p><strong>At Fervanta Commerce, your company runs applications in Google Kubernetes Engine (GKE). Several applications rely on local ephemeral volumes, and some workloads have become unstable because worker nodes enter the DiskPressure condition. You need to identify which Pods are consuming the ephemeral storage, but you do not have execute access to the workloads or nodes. What should you do?</strong></p><p><strong>A. Check the node/ephemeral_storage/used_bytes metric in Cloud Monitoring Metrics Explorer.</strong></p><p><strong>B. Check the container/ephemeral_storage/used_bytes metric in Cloud Monitoring Metrics Explorer and group or filter the time series by Pod and container labels.</strong></p><p><strong>C. Locate all Pods with emptyDir volumes and run df -h inside them to measure disk usage.</strong></p><p><strong>D. Locate all Pods with emptyDir volumes and run du -sh against their filesystems to measure disk usage.</strong></p><p>A is incorrect because node-level ephemeral-storage usage can confirm that a node is consuming large amounts of local ephemeral storage, but it does not provide enough attribution to identify which Pod or container is responsible. The requirement is to isolate the workload causing DiskPressure, so a container-scoped metric with Kubernetes resource labels is more useful.</p><blockquote><strong><em>B is correct because GKE system metrics include container/ephemeral_storage/used_bytes on the k8s_container monitored resource. The metric records local ephemeral storage usage for individual containers and carries resource labels that let you identify the associated cluster, namespace, Pod, and container. Metrics Explorer can therefore compare or group the measurements without requiring shell access to the Pods or nodes, making it the appropriate observability-based diagnostic for the stated access restriction.</em></strong></blockquote><p>C is incorrect because this method requires execute access to the workload or node so that the df command can run in the relevant filesystem context. The question explicitly says that you do not have that access. In addition, df reports filesystem-level capacity and utilization and does not necessarily provide clean attribution across all container writable layers, logs, and emptyDir usage associated with a Pod.</p><p>D is incorrect because this option also depends on execute access that is unavailable. A du command can estimate directory-level file consumption when you can inspect a filesystem directly, but it is not a practical method when access to the workloads and nodes is prohibited. Cloud Monitoring already exposes the container-level ephemeral-storage metric needed for attribution.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/monitoring/api/metrics_kubernetes">docs.cloud.google.com/monitoring/api/metrics_kubernetes</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/view-observability-metrics">docs.cloud.google.com/kubernetes-engine/docs/how-to/view-observability-metrics</a></p><p><strong>Q17.</strong></p><p><strong>At Iverex Media, your Cloud Run application currently writes unstructured text logs to Cloud Logging. You want future application log entries to use JSON-based structured payloads so that fields can be queried independently. What should you do?</strong></p><p><strong>A. Modify the application to use a Cloud Logging client library and write structured log entries with a jsonPayload.</strong></p><p><strong>B. Add a Fluent Bit sidecar container and use a JSON parser to transform the application’s standard output.</strong></p><p><strong>C. Install the Google Cloud Ops Agent in the Cloud Run container image and use it to forward logs to Cloud Logging.</strong></p><p><strong>D. Configure the Google Cloud Ops Agent to convert the existing Cloud Run textPayload entries into jsonPayload entries after they are written.</strong></p><blockquote><strong><em>A is correct because Cloud Logging supports structured log entries whose application-defined fields are stored in jsonPayload. A Cloud Logging client library can write a structured LogEntry directly, which converts the application’s telemetry from an undifferentiated text string into individually queryable fields. For Cloud Run, an even lighter current implementation is to have the application write each log entry as a single serialized JSON object to standard output or standard error; Cloud Run detects and parses that JSON into structured Logging fields. Both approaches require the application to emit structured data instead of relying on a guest logging agent.</em></strong></blockquote><p>B is incorrect because Cloud Run can run multiple containers, but introducing a Fluent Bit sidecar solely to transform log formatting adds container resources, configuration, parsing rules, and lifecycle complexity. If the application can be changed, emitting JSON directly or using the Logging client library is much simpler and is natively understood by Cloud Run logging.</p><p>C is incorrect because Cloud Run is a fully managed serverless container platform and does not use the VM-oriented Ops Agent as the normal log collection mechanism. Cloud Run automatically captures logs written to standard output and standard error and also accepts entries written through the Cloud Logging API. Installing a guest telemetry agent inside the container is unnecessary.</p><p>D is incorrect because the Ops Agent is not the Cloud Run ingestion path and it does not retroactively transform already ingested textPayload records in Cloud Logging. Structured logging needs to be produced before or during ingestion, such as by writing JSON-formatted output or by sending a structured LogEntry through the Logging API.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/run/docs/logging">docs.cloud.google.com/run/docs/logging</a></p><p><a href="http://docs.cloud.google.com/logging/docs/structured-logging">docs.cloud.google.com/logging/docs/structured-logging</a></p><p><a href="http://docs.cloud.google.com/logging/docs/log-entry-data-model">docs.cloud.google.com/logging/docs/log-entry-data-model</a></p><p><strong>Q18.</strong></p><p><strong>At Klyvanta Enterprises, your company recently migrated to Google Cloud. You need a fast, reliable, and repeatable way to provision new Google Cloud projects and standard foundational resources by using established Google Cloud practices. What should you do?</strong></p><p><strong>A. Use the Google Cloud console to create every project manually.</strong></p><p><strong>B. Write a custom shell script that uses the gcloud CLI and passes parameters from each request. Store the script in a Git repository.</strong></p><p><strong>C. Write a new custom Terraform module and store it in source control. Manually copy and run terraform apply whenever a new project is requested.</strong></p><p><strong>D. Use maintained Terraform modules from the Cloud Foundation Toolkit, such as the Project Factory, and apply them with the appropriate parameters to create the project and related resources.</strong></p><p>A is incorrect because manual console provisioning is slow, difficult to reproduce exactly, and prone to configuration drift and human error. As the number of projects grows, operators would repeatedly perform the same IAM, API, network, billing, and policy configuration steps. Infrastructure as code provides a much stronger repeatability and audit trail.</p><p>B is incorrect because a gcloud-based script is more repeatable than manual console work, but the team becomes responsible for implementing idempotency, dependency ordering, error recovery, upgrades, and every project-governance feature it needs. Declarative Terraform modules provide a more maintainable model and can express the full desired state of the project and associated resources.</p><p>C is incorrect because a custom Terraform module can absolutely provide repeatable provisioning, but developing a project-factory implementation from scratch creates unnecessary engineering work when maintained Google Cloud Terraform modules already provide an opinionated, reusable project-provisioning baseline. The option also depends on manually copying and executing commands rather than integrating the module into a standardized automation workflow.</p><blockquote><strong><em>D is correct because the Cloud Foundation Toolkit contains reusable Terraform modules built for common Google Cloud foundation tasks. The Project Factory module creates and configures projects with capabilities such as API enablement, IAM, Shared VPC integration, service accounts, and other project-level settings. Using a maintained module reduces custom development, keeps provisioning declarative and version controlled, and provides a fast and repeatable foundation that can be integrated into an automated project-factory workflow.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/build/docs/terraform">docs.cloud.google.com/build/docs/terraform</a></p><p><a href="http://github.com/terraform-google-modules/terraform-google-project-factory">github.com/terraform-google-modules/terraform-google-project-factory</a></p><p><a href="http://github.com/terraform-google-modules/terraform-example-foundation">github.com/terraform-google-modules/terraform-example-foundation</a></p><p><strong>Q19.</strong></p><p><strong>At Orynthia Media, you operate Compute Engine n2-standard-2 instances that serve large files. Users report slow downloads, and Cloud Monitoring shows that the instances are reaching their maximum available network throughput. You need to improve per-instance network throughput. What should you do?</strong></p><p><strong>A. Add additional virtual network interface controllers to each VM.</strong></p><p><strong>B. Deploy a Cloud NAT gateway and attach it to the VMs’ subnet.</strong></p><p><strong>C. Change the VM machine type from n2-standard-2 to n2-standard-8.</strong></p><p><strong>D. Deploy the Google Cloud Ops Agent to export additional network metrics.</strong></p><p>A is incorrect because Compute Engine network bandwidth is accounted for per VM instance, not independently per virtual NIC or IP address. Adding another vNIC to an N2 VM does not multiply the machine type’s per-instance egress-bandwidth limit. Additional interfaces can be useful for network segmentation or specialized multi-NIC architectures, but they do not solve a machine-type bandwidth ceiling for this workload.</p><p>B is incorrect because Cloud NAT provides outbound internet connectivity for instances that do not have external IP addresses and can scale NAT capacity for appropriate egress scenarios. It does not raise the source VM’s own per-instance network-bandwidth limit. If the VM is already at its machine-type throughput ceiling, placing a NAT gateway in the path does not increase that ceiling.</p><blockquote><strong><em>C is correct because the VM’s machine type determines its maximum possible per-instance egress bandwidth. Current N2 machine-type specifications list n2-standard-2 with default egress bandwidth of up to 10 Gbps and n2-standard-8 with default egress bandwidth of up to 16 Gbps. Moving to the larger machine type therefore raises the VM’s potential bandwidth limit in addition to increasing CPU and memory resources. Actual achievable throughput still depends on the destination, traffic pattern, packet size, number of flows, and other documented factors, but this is the option that directly increases the relevant machine-level limit.</em></strong></blockquote><p>D is incorrect because the Ops Agent can provide additional telemetry for troubleshooting and capacity analysis, but monitoring does not increase available network throughput. The existing dashboard already shows that the VMs have reached their network limit, so the next action must change the network-capacity characteristics rather than collect more measurements.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/general-purpose-machines">docs.cloud.google.com/compute/docs/general-purpose-machines</a></p><p><a href="http://docs.cloud.google.com/compute/docs/network-bandwidth">docs.cloud.google.com/compute/docs/network-bandwidth</a></p><p><a href="http://docs.cloud.google.com/compute/docs/networking/network-overview">docs.cloud.google.com/compute/docs/networking/network-overview</a></p><p><strong>Q20.</strong></p><p><strong>At Jorynex Security, you are designing a new multi-tenant Google Kubernetes Engine (GKE) cluster. The customer wants to avoid long-lived credentials and requires every workload to receive only its minimum required Identity and Access Management (IAM) permissions. You specifically need an IAM service-account impersonation design using current Google-recommended GKE identity mechanisms. What should you do?</strong></p><p><strong>A. 1. Create an IAM service account.</strong></p><p><strong>2. Create a node pool and set the IAM service account as the node pool’s default identity.</strong></p><p><strong>3. Restrict the workload to that node pool by using node selectors, taints, and tolerations.</strong></p><p><strong>4. Repeat for every workload.</strong></p><p><strong>B. 1. Create an IAM service account.</strong></p><p><strong>2. Create a node pool without taints and set the IAM service account as the node pool’s default identity.</strong></p><p><strong>3. Grant the required IAM permissions to that service account.</strong></p><p><strong>C. 1. Create a dedicated IAM service account.</strong></p><p><strong>2. Create a dedicated Kubernetes ServiceAccount in a cluster with Workload Identity Federation for GKE enabled.</strong></p><p><strong>3. Grant that Kubernetes ServiceAccount roles/iam.workloadIdentityUser on the IAM service account and annotate the Kubernetes ServiceAccount with iam.gke.io/gcp-service-account.</strong></p><p><strong>4. Assign the Kubernetes ServiceAccount to the workload.</strong></p><p><strong>5. Repeat with separate least-privilege identities for each workload as needed.</strong></p><p><strong>D. 1. Create an IAM service account.</strong></p><p><strong>2. Create a user-managed service account JSON key.</strong></p><p><strong>3. Store the key in a Kubernetes Secret.</strong></p><p><strong>4. Mount the Secret in the workload and point GOOGLE_APPLICATION_CREDENTIALS to the key file.</strong></p><p><strong>5. Repeat for every workload.</strong></p><p>A is incorrect because a node service account is shared by workloads that run on that node pool and exists primarily as the identity used by the GKE nodes themselves. Creating a node pool for every application increases cluster-management complexity, wastes scheduling flexibility, and creates a coarser permission boundary than workload identity. Node scheduling controls are not an ideal substitute for per-workload IAM identities.</p><p>B is incorrect because every workload scheduled on the untainted node pool can potentially run in the context of the node’s service account for operations that use node credentials or metadata paths, depending on the cluster configuration. The design does not provide a distinct least-privilege IAM identity for each workload and therefore is especially unsuitable for a multi-tenant cluster.</p><blockquote><strong><em>C is correct because Workload Identity Federation for GKE is the current keyless mechanism for giving Kubernetes workloads Google Cloud identities. When IAM service account impersonation is specifically required, a Kubernetes ServiceAccount can be authorized with roles/iam.workloadIdentityUser on a dedicated IAM service account and linked through the iam.gke.io/gcp-service-account annotation. The workload then receives short-lived credentials rather than a private service account key. Giving each workload its own Kubernetes identity and narrowly scoped IAM service account supports least privilege and isolates tenant permissions. Where supported, direct IAM grants to GKE workload principals can be even simpler, but this option correctly implements the requested impersonation model.</em></strong></blockquote><p>D is incorrect because a user-managed service account key is a long-lived private credential that can be copied from the Kubernetes Secret and potentially used outside the cluster until it is disabled. It introduces key creation, distribution, rotation, audit, and revocation responsibilities that Workload Identity Federation for GKE was specifically designed to avoid.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity">docs.cloud.google.com/kubernetes-engine/docs/concepts/workload-identity</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/workload-identity">docs.cloud.google.com/kubernetes-engine/docs/how-to/workload-identity</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/service-accounts">docs.cloud.google.com/kubernetes-engine/docs/how-to/service-accounts</a></p><p><strong>Q21.</strong></p><p><strong>At Aurelvia API Labs, you manage a critical API running on Cloud Run that serves an average of 10,000 requests per minute. You need to define Service Level Objectives (SLOs) for availability and latency to ensure that the API meets user expectations, which include 99.9% availability and a maximum latency of 200 milliseconds for 95% of requests. You also need to ensure that these SLOs are actively monitored and measured. What should you do?</strong></p><p><strong>A. Configure Cloud Monitoring to send alerts when average API latency exceeds 150 ms or the error rate surpasses 0.1%.</strong></p><p><strong>B. Prioritize latency as the only SLO, targeting 100 ms for 99% of requests.</strong></p><p><strong>C. Set SLOs for 99% availability and 500 ms latency for 90% of requests. Use Cloud Monitoring to track the SLOs and alert on violations.</strong></p><p><strong>D. Set SLOs for the API by using availability and latency service level indicators that match the stated user expectations. Use Cloud Monitoring to track the SLOs and alert on SLO risk or violations.</strong></p><p>A is incorrect because fixed alert thresholds are not a substitute for explicitly defining and measuring the required SLOs. An error rate above 0.1% does correspond conceptually to availability below 99.9%, but the latency alert in this option is based on average latency rather than the stated requirement that 95% of requests complete within 200 ms. An average can remain low while a significant tail of requests is unacceptably slow. Cloud Monitoring SLO monitoring provides request-based availability and latency indicators that directly measure the proportion of good requests over a compliance period and can support error-budget or burn-rate alerting.</p><p>B is incorrect because the requirement contains two distinct user expectations: availability and latency. Defining only a latency objective leaves failures and unsuccessful requests outside the reliability model. In addition, the proposed latency objective of 100 ms for 99% of requests is substantially stricter than the stated 200 ms for 95% of requests and is not justified by the business requirement. SLOs should reflect meaningful user expectations rather than arbitrarily tightening targets.</p><p>C is incorrect because the proposed objectives are weaker than the requirements. A 99% availability objective permits approximately ten times the unavailable-request fraction of a 99.9% objective, and allowing 500 ms for 90% of requests does not satisfy a requirement that 95% complete within 200 ms. Monitoring incorrect SLOs accurately still does not protect the user expectations specified in the question.</p><blockquote><strong><em>D is correct because availability and latency are the two user-facing reliability dimensions explicitly required. The availability SLI should classify successful versus valid requests and use a 99.9% goal, while the latency SLI should classify requests completing within 200 ms as good and use a 95% goal over the chosen compliance period. Cloud Monitoring supports SLO resources based on availability and latency indicators and can display SLO compliance and error-budget status and create alerting policies when the SLO is being consumed too quickly or is at risk. This directly converts the stated user expectations into measurable operational objectives.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/monitoring/slo-monitoring">docs.cloud.google.com/monitoring/slo-monitoring</a></p><p><a href="http://docs.cloud.google.com/monitoring/api/ref_v3/rest/v3/services.serviceLevelObjectives">docs.cloud.google.com/monitoring/api/ref_v3/rest/v3/services.serviceLevelObjectives</a></p><p><a href="http://docs.cloud.google.com/monitoring/dashboards/slos">docs.cloud.google.com/monitoring/dashboards/slos</a></p><p><strong>Q22.</strong></p><p><strong>At Grevanta Health, regulatory requirements state that all resources containing company data must be created in United States locations. You attempt to create a secret in Secret Manager and receive this error:</strong></p><p><strong>Constraint constraints/gcp.resourceLocations violated for the project while attempting to create a secret in global.</strong></p><p><strong>You need to resolve the error while remaining compliant with the location restriction. What should you do?</strong></p><p><strong>A. Remove the organization policy referenced in the error message.</strong></p><p><strong>B. Create the secret with an automatic replication policy.</strong></p><p><strong>C. Create the secret with a user-managed replication policy that uses only permitted United States locations.</strong></p><p><strong>D. Add global to the organization policy’s allowed locations.</strong></p><p>A is incorrect because removing the resource-location organization policy would eliminate the control that enforces the regulatory requirement. Although the secret could then be created, the organization would no longer have a centralized mechanism preventing resources from being created in disallowed locations. The correct approach is to configure the secret so its storage locations comply with the existing policy.</p><p>B is incorrect because Secret Manager automatic replication is treated as global for purposes of the resource-location organization policy. Google Cloud manages where automatically replicated secret data is stored, and creation succeeds only when global is allowed by the policy. The error explicitly indicates that global is not permitted, so automatic replication does not satisfy the requirement.</p><blockquote><strong><em>C is correct because a user-managed Secret Manager replication policy lets the organization explicitly select the regions where secret payload data is replicated. If every chosen replica location is permitted by constraints/gcp.resourceLocations, the secret can be created while complying with the requirement that data remain in approved United States locations. This is the intended Secret Manager design when an organization needs explicit control over secret-data residency.</em></strong></blockquote><p>D is incorrect because allowing global would permit use of Secret Manager automatic replication, but it weakens the organization’s location restriction and no longer guarantees that the secret payload is stored only in the approved United States locations. The regulatory requirement should be satisfied through explicit user-managed replication rather than widening the organization policy simply to accommodate the default replication model.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/choosing-replication">docs.cloud.google.com/secret-manager/docs/choosing-replication</a></p><p><a href="http://docs.cloud.google.com/secret-manager/docs/locations">docs.cloud.google.com/secret-manager/docs/locations</a></p><p><a href="http://docs.cloud.google.com/organization-policy/reference/restrict-locations-supported-services">docs.cloud.google.com/organization-policy/reference/restrict-locations-supported-services</a></p><p><strong>Q23.</strong></p><p><strong>At Helyrix Development, you are responsible for creating development environments for your company’s development team. You want every developer to receive an identical managed IDE environment, and the workstations must not be exposed to public networks. You also want to minimize compute cost without interrupting developers who are actively working. What should you do?</strong></p><p><strong>A. Create multiple Compute Engine VM instances with public IP addresses and use Cloud NAT. Configure an instance schedule to shut down the VMs.</strong></p><p><strong>B. Create multiple Compute Engine VM instances without public IP addresses. Configure an instance schedule to shut down the VMs.</strong></p><p><strong>C. Create a Cloud Workstations private cluster, configure workstation VMs without public IP addresses, and set an idleTimeout on the workstation configuration.</strong></p><p><strong>D. Create a Cloud Workstations private cluster, configure workstation VMs without public IP addresses, and set a runningTimeout that stops each workstation after a fixed elapsed time regardless of activity.</strong></p><p>A is incorrect because assigning public IP addresses directly conflicts with the requirement that the development environments not be exposed to public networks. Cloud NAT is used to provide outbound internet access to resources without external IP addresses; it is not a reason to assign public addresses to the workstation VMs. A fixed VM schedule also does not account for whether a developer is actively working at the scheduled shutdown time.</p><p>B is incorrect because private Compute Engine VMs can satisfy the no-public-IP requirement, but the team must create and manage its own IDE images, VM lifecycle, access, persistent disks, identity integration, updates, and developer connectivity. A fixed instance schedule can also stop a VM while a developer is still working. Cloud Workstations provides standardized managed development environments and lifecycle controls with less administrative effort.</p><blockquote><strong><em>C is correct because Cloud Workstations lets the organization define one workstation configuration that produces consistent IDE and tool environments for all developers. A private workstation cluster prevents public internet access to the workstation control plane, and disabling public IP addresses on workstation VMs keeps the compute environment private. The idleTimeout automatically stops a workstation after it has received no user traffic for the configured period, reducing compute cost while allowing actively used workstations to continue running. This directly balances security, consistency, cost, and developer productivity.</em></strong></blockquote><p>D is incorrect because runningTimeout stops a workstation after a fixed amount of elapsed running time regardless of whether the workstation is idle. It can therefore terminate an active developer session simply because the maximum runtime has been reached. Running timeouts are useful for security-update and lifecycle policies, but idleTimeout is the more appropriate control when the primary goal is to reduce unused compute cost without disrupting active developers.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/workstations/docs/configure-vpc-service-controls-private-clusters">docs.cloud.google.com/workstations/docs/configure-vpc-service-controls-private-clusters</a></p><p><a href="http://docs.cloud.google.com/workstations/docs/reference/rest/v1/projects.locations.workstationClusters.workstationConfigs">docs.cloud.google.com/workstations/docs/reference/rest/v1/projects.locations.workstationClusters.workstationConfigs</a></p><p><a href="http://docs.cloud.google.com/workstations/docs/create-configuration">docs.cloud.google.com/workstations/docs/create-configuration</a></p><p><strong>Q24.</strong></p><p><strong>At Kalyntar Web, your company wants to implement a continuous-delivery pipeline in Cloud Deploy for a web service deployed to GKE. The web service currently has no automated testing. The Quality Assurance team must manually verify every new release in a non-production environment before any production traffic is processed. You need to design the delivery pipeline. What should you do?</strong></p><p><strong>A. Create a single pipeline stage and use a standard deployment strategy.</strong></p><p><strong>B. Create a single pipeline stage and use a canary deployment strategy.</strong></p><p><strong>C. Create two pipeline stages and use a canary deployment strategy.</strong></p><p><strong>D. Create two pipeline stages and use a standard deployment strategy.</strong></p><p>A is incorrect because one stage cannot provide a separate pre-production environment in which QA can test the release before it reaches the production environment. If the one stage represents production, the release is already deployed to the production target before manual verification occurs. The pipeline needs an earlier target that is isolated from production users.</p><p>B is incorrect because canary deployment limits the amount of production traffic sent to a new version, but the requirement states that QA must verify the release before any production traffic reaches it. A canary therefore exposes some production traffic too early if it is used as the substitute for a pre-production validation environment. A second stage is still required.</p><p>C is incorrect because two stages correctly provide a pre-production and production target, but a canary strategy is unnecessary when QA is expected to approve the release before any production users see it. Canarying is valuable when the organization wants progressive production exposure. Here, the cleaner design is to deploy normally to the QA or staging target, complete manual verification, and then promote to production.</p><blockquote><strong><em>D is correct because the first standard-deployment stage can deploy the complete candidate release to a staging or QA GKE target where the QA team can test it without production traffic. The second stage represents the production target. Cloud Deploy can configure the production target with requireApproval: true so promotion waits for human approval after QA succeeds. Once approved, a standard deployment releases the validated version to production. This directly matches the requirement and avoids introducing a production canary before manual verification.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/deploy/docs/create-pipeline-targets">docs.cloud.google.com/deploy/docs/create-pipeline-targets</a></p><p><a href="http://docs.cloud.google.com/deploy/docs/approve">docs.cloud.google.com/deploy/docs/approve</a></p><p><a href="http://docs.cloud.google.com/deploy/docs/deployment-strategies/canary">docs.cloud.google.com/deploy/docs/deployment-strategies/canary</a></p><p><strong>Q25.</strong></p><p><strong>At Fervanta Products, your company is actively developing a new web service using Site Reliability Engineering practices. You lead the SRE team and are preparing the service for production. You want to follow Google-recommended SRE practices during productionalization. Which two actions should you take next? (Choose two.)</strong></p><p><strong>A. Establish appropriate monitoring and alerting systems for the service.</strong></p><p><strong>B. Start documenting infrastructure and system guidelines.</strong></p><p><strong>C. Collaborate with the product team to design the service.</strong></p><p><strong>D. Implement redundancy measures.</strong></p><p><strong>E. Engage early with the development and product teams to review architecture and design choices in detail.</strong></p><blockquote><strong><em>A is correct because productionalization includes the operational capabilities needed to understand and safely run the service. Google SRE’s engagement model explicitly identifies monitoring, alerting, capacity planning, redundancy, spike and overload planning, and performance tuning as active-development productionalization work. Establishing observability before launch lets the teams validate behavior during testing and limited availability and provides the signals needed for incident response and SLO-based operation.</em></strong></blockquote><p>B is incorrect because documenting infrastructure and system guidelines is useful organizational work, but it is not one of the two strongest immediate productionalization actions for this specific service. The team first needs concrete service-level operational readiness, including observability, and direct SRE engagement with the application’s architecture and design. Guidelines can be developed in parallel but do not replace those service-specific activities.</p><p>C is incorrect because collaborating with the product team is valuable, but this option is broader and less specifically aligned with the SRE productionalization action than early-engagement consulting focused on architecture and design choices. The SRE contribution should deliberately surface reliability, scalability, dependency, operability, and failure-mode concerns rather than merely participate in product design in a generic role.</p><p>D is incorrect because redundancy can be an important production-readiness measure, but it should follow an understanding of the service’s SLOs, architecture, failure domains, dependencies, and capacity needs. Implementing redundancy as a generic next step is less appropriate than first establishing observability and engaging early on architecture and design so the team knows where redundancy actually provides reliability value.</p><blockquote><strong><em>E is correct because Google’s SRE engagement guidance emphasizes involving SRE during architecture and design. Early engagement allows SREs to challenge assumptions, review dependencies and failure modes, advise on scalability and operational behavior, and help ensure that reliability requirements are designed into the service instead of retrofitted shortly before or after launch. Fixing architectural mistakes becomes significantly more expensive later in the lifecycle.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://sre.google/workbook/engagement-model/">sre.google/workbook/engagement-model/</a></p><p><a href="http://sre.google/sre-book/evolving-sre-engagement-model/">sre.google/sre-book/evolving-sre-engagement-model/</a></p><p><a href="http://sre.google/sre-book/communication-and-collaboration/">sre.google/sre-book/communication-and-collaboration/</a></p><p><strong>Q26.</strong></p><p><strong>At Galvoryx Delivery, your company uses Cloud Build and Artifact Registry to deploy container images to Google Kubernetes Engine (GKE). Images are tagged with the source commit hash and are promoted through development and pre-production before production. A recent production deployment failed because of integration behavior that was not caught earlier, and the manual rollback was disruptive. You also notice many old unused images accumulating in Artifact Registry. You need to improve rollout and rollback management and automatically clean up old images. What should you do?</strong></p><p><strong>A. Build a custom rollback pipeline in Cloud Build and configure an Artifact Registry cleanup policy.</strong></p><p><strong>B. Adopt Cloud Deploy for release promotion and rollback management, and configure an Artifact Registry cleanup policy.</strong></p><p><strong>C. Deploy Cloud Service Mesh across all GKE clusters and manually delete old Artifact Registry images.</strong></p><p><strong>D. Adopt Cloud Deploy for release management and schedule a custom Cloud Build job that periodically deletes old images.</strong></p><p>A is incorrect because an Artifact Registry cleanup policy is appropriate, but building a custom rollback framework in Cloud Build requires the team to implement release history, target selection, rollback logic, permissions, failure handling, and auditing itself. Cloud Build is a flexible CI execution platform, but Cloud Deploy provides managed continuous-delivery resources and native rollback behavior for supported deployment targets.</p><blockquote><strong><em>B is correct because Cloud Deploy manages releases, promotions, rollouts, approvals, and deployment progression to GKE targets. If a production release fails, Cloud Deploy can create a new rollout based on the last successfully deployed release or another explicitly selected earlier release, giving the team a repeatable rollback workflow rather than an improvised manual procedure. Artifact Registry cleanup policies independently provide managed background deletion or retention of artifact versions based on age, tag state, prefixes, or keep-most-recent rules. Together, the services address both release lifecycle and accumulated-image cleanup with less custom automation.</em></strong></blockquote><p>C is incorrect because Cloud Service Mesh can provide service-to-service traffic management, security, and observability, but deploying a mesh does not replace a continuous-delivery release and rollback system. Manually deleting old images also creates recurring toil and increases the chance that an operator deletes an image that should have been retained for rollback.</p><p>D is incorrect because Cloud Deploy is appropriate for deployment management, but a scheduled custom cleanup build duplicates a managed Artifact Registry capability. Cleanup policies can automatically retain important recent or release artifacts and delete versions matching the configured criteria. Using the native policy requires less code, scheduling, permissions management, and maintenance.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/deploy/docs/overview">docs.cloud.google.com/deploy/docs/overview</a></p><p><a href="http://docs.cloud.google.com/deploy/docs/roll-back">docs.cloud.google.com/deploy/docs/roll-back</a></p><p><a href="http://docs.cloud.google.com/artifact-registry/docs/repositories/cleanup-policy">docs.cloud.google.com/artifact-registry/docs/repositories/cleanup-policy</a></p><p><strong>Q27.</strong></p><p><strong>At Fylora Platforms, your company uses one Google Kubernetes Engine (GKE) cluster for staging and another for production. You need to configure GKE release channels by following Google-recommended practices. You prioritize production stability over receiving the newest features and want staging to provide advance validation for production updates. What should you do?</strong></p><p><strong>A. Configure the production cluster to use the Stable release channel, and configure the staging cluster to use the Regular release channel.</strong></p><p><strong>B. Configure the production cluster to use the Extended release channel, and configure the staging cluster to use the Rapid release channel.</strong></p><p><strong>C. Configure the production cluster to use the Stable release channel, and configure the staging cluster to use the Stable release channel.</strong></p><p><strong>D. Configure the production cluster to use the Regular release channel, and configure the staging cluster to use the Regular release channel.</strong></p><blockquote><strong><em>A is correct because the Stable channel prioritizes maturity over feature freshness and receives versions only after they have accumulated validation in earlier release channels, making it the appropriate choice when production stability is the primary concern. Using Regular for staging gives the organization earlier exposure to Kubernetes and GKE changes that can later reach Stable, providing a realistic pre-production qualification window without using the least-mature Rapid channel for ordinary staging workloads.</em></strong></blockquote><p>B is incorrect because the Extended channel is intended primarily for workloads that need long-term support on a Kubernetes minor version, not as a general substitute for the Stable channel whenever production prioritizes maturity. Pairing it with Rapid also creates a very large version and release-cadence gap between staging and production, so staging is less representative of the changes that production will actually receive.</p><p>C is incorrect because Stable is appropriate for production, but putting staging on exactly the same release channel reduces its value as an advance-validation environment. Staging should normally encounter relevant GKE changes before production so the organization has time to detect compatibility or operational problems prior to production rollout.</p><p>D is incorrect because Regular provides a balanced mix of feature freshness and release maturity and is Google’s default recommendation for many clusters, but the question explicitly states that production prioritizes stability over access to newer features. Stable more directly satisfies that production requirement while Regular remains useful for pre-production validation.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/concepts/release-channels">docs.cloud.google.com/kubernetes-engine/docs/concepts/release-channels</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/how-to/release-channels">docs.cloud.google.com/kubernetes-engine/docs/how-to/release-channels</a></p><p><a href="http://docs.cloud.google.com/kubernetes-engine/docs/release-schedule">docs.cloud.google.com/kubernetes-engine/docs/release-schedule</a></p><p><strong>Q28.</strong></p><p><strong>At Heleris Analytics, you are investigating an ongoing anomaly involving a high number of requests. You created a custom Cloud Monitoring dashboard and determined that the requests originate from one type of client. You want the development team to use the same custom dashboard for further investigation while following Google-recommended practices and least privilege. What should you do?</strong></p><p><strong>A. Export the custom dashboard as a JSON file and share the file with the development team.</strong></p><p><strong>B. Export the dashboard time-series data in CSV format. Instruct the development team to use BigQuery to analyze the data.</strong></p><p><strong>C. Grant the development team the roles/monitoring.viewer IAM role at the project level and share the custom dashboard URL with the development team.</strong></p><p><strong>D. Grant the development team the roles/viewer IAM role at the project level and share the custom dashboard URL with the development team.</strong></p><p>A is incorrect because exporting a dashboard definition as JSON is useful when another project needs an independent copy of the dashboard, but the development team needs to investigate the same live data and dashboard in the current project. A copied JSON definition is not necessary and does not by itself grant permission to read the underlying metric data.</p><p>B is incorrect because exporting data to CSV and moving the analysis to BigQuery creates a separate, static analysis workflow instead of letting developers inspect the same live Monitoring dashboard. It adds data export and query work and loses the convenience of the existing time-series visualization.</p><blockquote><strong><em>C is correct because Google Cloud recommends sharing a custom Monitoring dashboard through its URL and granting recipients the Monitoring Viewer role when they need read-only access. roles/monitoring.viewer provides the permissions needed to view Monitoring data and dashboard configuration without granting edit or broader project-management capabilities. This lets the development team investigate the current anomaly in the same dashboard while following least privilege more closely than a project-wide Viewer role.</em></strong></blockquote><p>D is incorrect because the basic project Viewer role grants broad read access to many Google Cloud resources outside Cloud Monitoring. The development team only needs monitoring visibility for this task. A product-specific Monitoring Viewer role provides the required dashboard and metric access with a narrower permission set.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/monitoring/charts/share-dashboards">docs.cloud.google.com/monitoring/charts/share-dashboards</a></p><p><a href="http://docs.cloud.google.com/monitoring/access-control">docs.cloud.google.com/monitoring/access-control</a></p><p><a href="http://docs.cloud.google.com/iam/docs/roles-permissions/monitoring">docs.cloud.google.com/iam/docs/roles-permissions/monitoring</a></p><p><strong>Q29.</strong></p><p><strong>At Zephryx Labs, you are working on a Google Cloud project that has been in production for several months, and you need to update an IAM policy to grant additional permissions to a team member. You need to proactively identify any potential risks associated with these IAM policy changes before applying them to the production environment. What should you do?</strong></p><p><strong>A. Implement the IAM policy changes. Use recent change recommendations to monitor for any unintended consequences.</strong></p><p><strong>B. Use change risk recommendations to assess the potential impact of the IAM policy changes and identify associated risks before applying them.</strong></p><p><strong>C. Grant the new team member the Project Owner role. Reevaluate permissions after the task is complete.</strong></p><p><strong>D. Use the Cloud deprecation and breaking changes recommender to assess the impact of the IAM policy changes on potentially deprecated or soon-to-change services.</strong></p><p>A is incorrect because the requirement is to identify risk before changing the production IAM policy. Monitoring recent changes after the policy is applied is reactive and can only help after access has already been modified. A risky IAM change could remove access required by an important workload or unexpectedly change a critical dependency before the team discovers the problem. Google Cloud change risk recommendations are designed to evaluate supported high-risk changes proactively rather than waiting for operational consequences.</p><blockquote><strong><em>B is correct because change risk recommendations are part of Google Cloud Recommender and are specifically designed to flag risky changes to important resources before they are committed. For IAM policies, the recommender evaluates signals such as recently exercised permissions, resource usage, and project importance to warn about changes that might break important access dependencies. This directly satisfies the requirement to assess the potential impact of the proposed IAM modification before applying it to production. The feature is currently a pre-GA capability, so organizations should also consider Policy Intelligence tools such as Policy Simulator when they need a more explicit access-impact simulation.</em></strong></blockquote><p>C is incorrect because the Project Owner role is a highly privileged basic role that grants broad administrative permissions far beyond the unspecified additional access that the team member needs. Granting Owner temporarily violates the principle of least privilege and materially increases the blast radius if the user’s account is compromised or an unintended action is performed. The correct approach is to grant only the required predefined or custom role after assessing the policy change.</p><p>D is incorrect because the deprecation and breaking-change recommender is concerned with product lifecycle changes that can affect workloads, such as deprecated features or service changes. It is not intended to model how an IAM allow-policy modification will affect principals or recently exercised permissions. Change risk recommendations and Policy Intelligence are the relevant tools for IAM policy-change risk.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/recommender/docs/change-risk-recommendations">docs.cloud.google.com/recommender/docs/change-risk-recommendations</a></p><p><a href="http://docs.cloud.google.com/policy-intelligence/docs/overview">docs.cloud.google.com/policy-intelligence/docs/overview</a></p><p><strong>Q30.</strong></p><p><strong>At Iveranta Payments, you manage a critical high-traffic application running on Google Kubernetes Engine (GKE). You are preparing a new release with significant changes to the payment-processing workflow. You need a deployment strategy that limits production risk, evaluates real telemetry, and automatically rolls back when the new release proves unstable. What should you do?</strong></p><p><strong>A. Use Cloud Deploy to execute a canary deployment beginning with a small traffic or workload percentage. Configure deploy analysis by using Google Cloud Observability alert-policy checks as release-health criteria, and configure Cloud Deploy automation to repair and roll back a failed rollout.</strong></p><p><strong>B. Configure a blue/green deployment manually in Kubernetes. Deploy the new version as a green Deployment beside the existing blue Deployment, and update the Service selector to gradually switch user traffic to the new version.</strong></p><p><strong>C. Edit the GKE Service manifest to select v2 Pods, create a second Deployment for the new version, and slowly scale the new Deployment up while scaling the old Deployment down.</strong></p><p><strong>D. Use the default Kubernetes rolling-update strategy. Monitor logs and metrics manually, and if an issue is detected run kubectl rollout undo.</strong></p><blockquote><strong><em>A is correct because Cloud Deploy supports progressive canary releases to GKE so only a limited portion of the workload or traffic receives the candidate version during early phases. Deploy analysis can evaluate production telemetry through Google Cloud Observability alert policies during a rollout and fail the analysis when the configured health criteria indicate a problem. Cloud Deploy automation can then use a repairRolloutRule to retry or automatically roll back a failed rollout. This combines controlled exposure, objective production telemetry, and automated recovery, which directly satisfies the requirement for a high-risk payment release.</em></strong></blockquote><p>B is incorrect because blue/green can provide rapid rollback, but manually updating a Kubernetes Service selector does not by itself provide the requested automated telemetry analysis and rollback. The option also says to switch traffic gradually through a Service selector, while a basic selector normally switches the selected backend set according to labels rather than providing a managed percentage-based progressive rollout. Additional traffic-management logic would be required.</p><p>C is incorrect because scaling two Deployments can change how many Pods of each version exist, but it does not guarantee a precise or stable traffic percentage and provides no automatic health analysis or rollback. It is a custom rollout procedure that the team would have to script, monitor, and maintain.</p><p>D is incorrect because a rolling update and kubectl rollout undo provide a basic manual rollback path, but the option depends on humans to watch telemetry and execute the rollback after the problem is detected. It does not meet the requirement for automatic release-health evaluation and automatic rollback, and a normal rolling update provides less controlled initial exposure than an explicit canary phase.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/deploy/docs/deployment-strategies/canary">docs.cloud.google.com/deploy/docs/deployment-strategies/canary</a></p><p><a href="http://docs.cloud.google.com/deploy/docs/analysis">docs.cloud.google.com/deploy/docs/analysis</a></p><p><a href="http://docs.cloud.google.com/deploy/docs/automation">docs.cloud.google.com/deploy/docs/automation</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03948d72d98c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GCP Google Cloud Professional Data Engineer Practice Questions]]></title>
            <link>https://medium.com/@kloudexams/gcp-google-cloud-professional-data-engineer-practice-questions-c4bb3c234744?source=rss-6ef48558799e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c4bb3c234744</guid>
            <category><![CDATA[gcp-certification]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gcp]]></category>
            <category><![CDATA[google-cloud]]></category>
            <dc:creator><![CDATA[KloudExams | Google Cloud Certification Prep]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 12:32:08 GMT</pubDate>
            <atom:updated>2026-10-05T12:32:08.787Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HS4TcY_nypRc7KUteIawKA.jpeg" /><figcaption>KloudExams GCP Google Cloud Professional Data Engineer Practice Questions</figcaption></figure><p>Preparing for the <strong>Google Cloud Professional Data Engineer certification</strong>? Knowing what data services do is a starting point — understanding how to design and operate systems that collect, process, store, and deliver data reliably is what turns that knowledge into practical skills.</p><p>This collection of practice questions from <strong>KloudExams</strong> is designed to help you test your understanding, identify knowledge gaps, and practise choosing data architectures and processing approaches that fit specific requirements. Each question includes detailed explanations for both correct and incorrect options, helping you learn the reasoning behind the answer rather than simply memorising a letter.</p><p>Before reading each explanation, take a moment to choose your answer and consider why the other options are less suitable. Even when you answer correctly, review the full explanation to check your reasoning. When you get a question wrong, use it as a starting point for revisiting the concept through Google Cloud documentation and hands-on data engineering practice.</p><p><strong>Focus on understanding, not memorisation. Let’s get started!</strong></p><p><strong>Q1.</strong></p><p><strong>At Nuvatrix Devices, you are choosing a NoSQL database to handle telemetry submitted by millions of Internet-of-Things devices. The volume of data is growing by approximately 100 TB per year, and each entry has about 100 attributes. The processing pipeline does not require ACID transactional semantics. High availability and low latency are required, and you need to analyze the data by querying individual fields. Which three databases best meet these requirements? (Choose three.)</strong></p><p><strong>A. Redis</strong></p><p><strong>B. HBase</strong></p><p><strong>C. MySQL</strong></p><p><strong>D. MongoDB</strong></p><p><strong>E. Cassandra</strong></p><p><strong>F. HDFS with Hive</strong></p><p>A is incorrect because Redis is primarily designed as an extremely low-latency in-memory data platform and key-value store. Modern Redis versions can support richer indexing and query capabilities, but using Redis as the primary durable store for a continuously growing telemetry corpus of roughly 100 TB per year would generally require a memory-heavy and operationally expensive architecture compared with databases designed around very large persistent datasets. The question asks for the three databases that best match massive durable NoSQL telemetry storage together with field-oriented querying, for which HBase, MongoDB, and Cassandra are more natural fits.</p><blockquote><strong><em>B is correct because HBase is a distributed wide-column NoSQL database built on the Hadoop ecosystem for very large sparse tables. It is designed to scale horizontally to huge datasets while providing low-latency random reads and writes by row key. Column families and qualifiers allow records to contain many attributes without requiring a rigid relational schema, and HBase filters and scans can operate on stored fields. Its architecture is therefore well aligned with large-scale telemetry where ACID behavior across arbitrary rows is not the primary requirement.</em></strong></blockquote><p>C is incorrect because MySQL is a relational database rather than a NoSQL database. It provides mature transactional semantics, indexes, SQL queries, and high-availability deployment patterns, but those features do not make it one of the requested NoSQL choices. Scaling an ingest-heavy telemetry corpus growing by 100 TB per year across millions of devices would also typically require substantial sharding or another distributed architecture rather than relying on a conventional relational MySQL design.</p><blockquote><strong><em>D is correct because MongoDB is a distributed document database that stores records as flexible documents whose individual fields can be queried and indexed. A telemetry record with around 100 attributes maps naturally to a document, and the schema can evolve as device firmware or event formats change. MongoDB supports replication for availability and sharding for horizontal scale, making it suitable for large operational datasets that require low-latency access and queries over document fields.</em></strong></blockquote><blockquote><strong><em>E is correct because Cassandra is a distributed wide-column NoSQL database designed for horizontal scalability, high availability, and high write throughput across clusters. Its architecture avoids a single master and can replicate data across failure domains, which makes it suitable for high-volume telemetry ingestion. With an appropriate primary-key model and, where suitable, supported indexing mechanisms, Cassandra can efficiently retrieve data based on modeled fields while scaling to very large datasets.</em></strong></blockquote><p>F is incorrect because HDFS with Hive is primarily a distributed storage and batch analytical query architecture, not a low-latency operational NoSQL database. Hive is well suited to large scans, ETL, and analytical SQL over data stored in distributed files, but query latency is generally far higher than systems intended for interactive or online reads and writes. It therefore does not satisfy the low-latency operational requirement even though it can store and analyze very large datasets.</p><p><strong>Reference links:</strong></p><p><a href="http://hbase.apache.org/book.html">hbase.apache.org/book.html</a></p><p><a href="http://mongodb.com/docs/manual/core/document">mongodb.com/docs/manual/core/document</a></p><p><a href="http://cassandra.apache.org/doc/latest/cassandra/architecture/overview.html">cassandra.apache.org/doc/latest/cassandra/architecture/overview.html</a></p><p><strong>Q2.</strong></p><p><strong>At Bidlume Auctions, your globally distributed auction application allows users to bid on items. Occasionally, users place identical bids at nearly identical times, and different application servers process those bids. Each bid event contains the item, amount, user, and event timestamp. You want to collate those bid events into a single location in real time so that the event timestamps can be compared to determine which user bid first. What should you do?</strong></p><p><strong>A. Create a file on a shared file system and have the application servers write all bid events to that file. Process the file with Apache Hadoop to identify which user bid first.</strong></p><p><strong>B. Have each application server publish bid events to Pub/Sub as they occur. Push the events from Pub/Sub to a custom endpoint that writes the bid-event information into Cloud SQL.</strong></p><p><strong>C. Set up a MySQL database for each application server to write bid events into. Periodically query each distributed MySQL database and update a master MySQL database with the bid-event information.</strong></p><p><strong>D. Have each application server publish bid events to Pub/Sub as they occur. Use a pull subscription to read the bid events with Dataflow. Give the bid for each item to the user whose bid event is processed first.</strong></p><p>A is incorrect because a shared file plus Hadoop is fundamentally a batch-oriented design and does not provide the required real-time collation. Multiple globally distributed application servers writing to a shared file also introduces coordination, locking, availability, and network concerns. Hadoop can process large historical datasets effectively, but waiting to process a file after events have accumulated is unnecessary when a managed messaging service can transport events continuously.</p><blockquote><strong><em>B is correct because Pub/Sub provides a scalable global ingestion layer for events produced by many application servers, and a push subscriber can deliver each event to an application endpoint as it arrives. Writing those events into one Cloud SQL database creates the requested central transactional location, where the application can compare the event timestamps associated with identical bids and determine the earlier bid. Pub/Sub delivery order itself should not be treated as the bid order; the correctness comes from persisting the event data and using the bid timestamp as the comparison value.</em></strong></blockquote><p>C is incorrect because maintaining a separate MySQL database for every application server creates unnecessary distributed state and then relies on periodic reconciliation. The periodic merge means the central view is not real time, and operational complexity grows as servers are added or removed. A messaging layer is a much better fit for continuously collecting events from distributed producers into a central system.</p><p>D is incorrect because Dataflow is a strong service for distributed stream processing, but this option awards the bid according to which event is processed first rather than according to the event timestamp. Pub/Sub does not provide global ordering by default, and current Dataflow documentation explicitly notes that the Pub/Sub connector and Apache Beam processing do not guarantee preservation of message-processing order. Network delay, retries, parallelism, or worker scheduling can therefore cause a later bid to be processed before an earlier one. A correct design must compare the bid timestamps or otherwise establish a valid ordering rule rather than equating processing order with event order.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/pubsub/docs/subscription-overview">docs.cloud.google.com/pubsub/docs/subscription-overview</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/streaming-with-cloud-pubsub">docs.cloud.google.com/dataflow/docs/concepts/streaming-with-cloud-pubsub</a></p><p><a href="http://docs.cloud.google.com/sql/docs/introduction">docs.cloud.google.com/sql/docs/introduction</a></p><p><strong>Q3.</strong></p><p><strong>At Ferluma Streaming, you stream data through Kafka into a Redis cluster, and both clusters run on Compute Engine VMs. You need encryption at rest with encryption keys whose lifecycle you control, including creation, rotation, disabling, and destruction. What should you do?</strong></p><p><strong>A. Create a dedicated service account and rely only on the default encryption at rest for the Compute Engine disks.</strong></p><p><strong>B. Create customer-managed encryption keys in Cloud Key Management Service and configure the Compute Engine persistent disks that store the Kafka and Redis data to use those keys.</strong></p><p><strong>C. Create encryption keys locally, import the key material into Cloud Key Management Service, and use the imported keys for every Compute Engine disk even though there is no requirement to manage key material outside Google Cloud.</strong></p><p><strong>D. Create encryption keys in Cloud Key Management Service and provide the keys in application API calls every time Kafka or Redis accesses data on the Compute Engine disks.</strong></p><p>A is incorrect because Compute Engine data is encrypted at rest by default with Google-owned and Google-managed keys, but those default keys do not satisfy the requirement that the organization control key creation, rotation, disabling, and destruction. A service account controls identity and authorization, not ownership of the disk-encryption key lifecycle. The requirement specifically calls for customer-controlled encryption keys.</p><blockquote><strong><em>B is correct because Compute Engine supports customer-managed encryption keys stored in Cloud KMS for persistent disks, snapshots, and related resources. Cloud KMS gives the organization control over key versions, rotation schedules, IAM permissions, disabling, destruction, auditability, and cryptographic location. The Compute Engine service agent receives permission to use the key, and the disk is configured to use that CMEK. This provides customer control over the key lifecycle without requiring Kafka or Redis to implement custom encryption-key handling.</em></strong></blockquote><p>C is incorrect because Cloud KMS supports imported key material, so this approach can be technically implemented, but it adds unnecessary external key-generation, import, custody, and operational procedures when the question does not require externally generated key material. If the requirement is simply to create, rotate, and destroy customer-managed keys, generating the key in Cloud KMS is simpler and reduces key-handling risk. Imported keys are appropriate only when an external key-origin requirement exists.</p><p>D is incorrect because customer-managed encryption for Compute Engine persistent disks is configured on the storage resource rather than supplied by Kafka or Redis on every data-access request. After the disk is configured with a CMEK and the Compute Engine service agent is authorized to use it, Compute Engine handles the encryption integration. Supplying a key with every application read or write resembles the operational model of customer-supplied keys, which is not the recommended current Compute Engine approach and is not how CMEK-protected disks are normally accessed.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/compute/docs/disks/disk-encryption">docs.cloud.google.com/compute/docs/disks/disk-encryption</a></p><p><a href="http://docs.cloud.google.com/compute/docs/disks/customer-managed-encryption">docs.cloud.google.com/compute/docs/disks/customer-managed-encryption</a></p><p><a href="http://docs.cloud.google.com/kms/docs/key-management-service">docs.cloud.google.com/kms/docs/key-management-service</a></p><p><strong>Q4.</strong></p><p><strong>At Trelvion Data, you are responsible for writing your company’s ETL pipelines to run on an Apache Hadoop cluster. The pipelines need to persist intermediate results for checkpoint-style processing and split one processing flow into multiple branches. Which method should you use to write the pipelines?</strong></p><p><strong>A. Pig Latin using Apache Pig</strong></p><p><strong>B. HiveQL using Apache Hive</strong></p><p><strong>C. Java using MapReduce</strong></p><p><strong>D. Python using MapReduce</strong></p><blockquote><strong><em>A is correct because Apache Pig provides the high-level Pig Latin data-flow language specifically for expressing multi-step data-processing pipelines on Hadoop. Pig Latin includes a SPLIT operator that partitions one relation into multiple output relations, which directly supports branching a pipeline. Pig also supports storing intermediate results and automatically writes intermediate data between MapReduce stages to distributed storage. These capabilities make it substantially easier to express checkpoint-style intermediate persistence and pipeline branching than implementing all orchestration manually with low-level MapReduce code.</em></strong></blockquote><p>B is incorrect because Apache Hive is primarily a SQL-oriented data-warehouse and query system. HiveQL is well suited to declarative operations such as SELECT, JOIN, GROUP BY, filtering, aggregation, and table manipulation, but the requirement emphasizes procedural data-flow branching and explicit intermediate pipeline stages. Hive can construct complex query plans and write intermediate or final tables, but Pig Latin provides a more natural data-flow abstraction and has a dedicated SPLIT operator for dividing one relation into multiple branches.</p><p>C is incorrect because Java MapReduce can implement the required behavior, but it is a substantially lower-level approach. The developer would need to define mapper and reducer classes, configure jobs, explicitly manage job dependencies, persist intermediate outputs, and orchestrate the branching of downstream jobs. Pig compiles its higher-level data-flow statements into distributed execution stages and therefore provides the required checkpointing and splitting behavior with much less implementation and orchestration code.</p><p>D is incorrect because Python can participate in Hadoop processing through mechanisms such as Hadoop Streaming, but using Python with MapReduce still leaves the developer responsible for building and coordinating the individual MapReduce stages. The language choice does not itself provide a high-level pipeline abstraction with built-in relation splitting and intermediate-result management. Pig Latin is specifically designed to describe this type of Hadoop data flow directly.</p><p><strong>Reference links:</strong></p><p><a href="http://pig.apache.org/docs/latest/start.html">pig.apache.org/docs/latest/start.html</a></p><p><a href="http://pig.apache.org/docs/latest/perf.html">pig.apache.org/docs/latest/perf.html</a></p><p><a href="http://hadoop.apache.org/docs/current/hadoop-mapreduce-client/hadoop-mapreduce-client-core/MapReduceTutorial.html">hadoop.apache.org/docs/current/hadoop-mapreduce-client/hadoop-mapreduce-client-core/MapReduceTutorial.html</a></p><p><strong>Q5.</strong></p><p><strong>At Avenquor Analytics, you migrated ETL jobs to BigQuery and need to verify that the migrated jobs produce the same output as the original jobs. You loaded the original output into one BigQuery table and the migrated output into another table. The tables do not contain a primary-key column that you can use to join corresponding rows. What should you do?</strong></p><p><strong>A. Select random samples from both tables with RAND() and compare the samples.</strong></p><p><strong>B. Select deterministic hash-based samples from both tables with FARM_FINGERPRINT and compare only those samples.</strong></p><p><strong>C. Use a Managed Service for Apache Spark cluster with the Spark BigQuery connector to read both tables, sort the comparable rows deterministically, calculate hashes from the relevant non-timestamp columns, and compare the resulting hashes.</strong></p><p><strong>D. Create stratified random samples with window functions and compare equivalent samples from both tables.</strong></p><p>A is incorrect because random sampling can provide evidence that two outputs look similar, but it cannot demonstrate that the full tables are identical. Rows that differ might simply not be selected in either sample, and independently generated random samples might not even contain corresponding records. The requirement is to validate the complete migrated result, so a deterministic comparison over all relevant rows is necessary.</p><p>B is incorrect because FARM_FINGERPRINT is a valid current BigQuery hash function and can be useful for deterministic row fingerprints, but comparing only sampled rows still leaves unexamined data. A hash-based sample can make repeatable spot checks easier than RAND(), but it cannot establish equality of the complete outputs. To validate migration correctness, all comparable rows must contribute to the comparison.</p><blockquote><strong><em>C is correct because the absence of a primary key prevents a straightforward row-by-row join, but the complete datasets can still be compared deterministically. The Spark BigQuery connector lets an Apache Spark workload read BigQuery tables directly. By selecting the same comparable columns, applying a deterministic ordering, and hashing the complete ordered contents, you can compare compact fingerprints representing the full outputs rather than relying on samples. Excluding processing timestamps is appropriate only when those timestamp values are intentionally nondeterministic and are not part of the business result being validated. Current Google Cloud documentation recommends the Spark BigQuery connector for Spark workloads rather than the legacy Hadoop MapReduce connector.</em></strong></blockquote><p>D is incorrect because stratified sampling is still sampling. Even if window functions are used to create comparable strata, differences outside the sampled rows can remain undetected. Stratification is useful for statistical analysis when a representative subset is sufficient, but the stated objective is to demonstrate that the migrated ETL output matches the original output, which requires a full deterministic comparison.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/managed-spark/docs/concepts/connectors/bigquery">docs.cloud.google.com/managed-spark/docs/concepts/connectors/bigquery</a></p><p><a href="http://docs.cloud.google.com/managed-spark/docs/guides/query-bigquery-tables">docs.cloud.google.com/managed-spark/docs/guides/query-bigquery-tables</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/reference/standard-sql/hash_functions">docs.cloud.google.com/bigquery/docs/reference/standard-sql/hash_functions</a></p><p><strong>Q6.</strong></p><p><strong>At Noveriq Archives, you keep historical structured files in Cloud Storage. You need to inspect the data for invalid entries and perform cleansing and transformation without requiring users to write application code or know SQL. What should you do?</strong></p><p><strong>A. Use Dataflow with Apache Beam transforms to detect invalid records and perform the transformations.</strong></p><p><strong>B. Create a BigQuery external table over the Cloud Storage data and use BigQuery data preparation’s visual editor and transformation suggestions to detect data-quality issues and prepare the data.</strong></p><p><strong>C. Use a Managed Service for Apache Spark cluster with a Hadoop or Spark job to detect invalid records and perform the transformations.</strong></p><p><strong>D. Use a BigQuery external table and manually write GoogleSQL queries to detect invalid records and perform the transformations.</strong></p><p>A is incorrect because Dataflow is a managed processing service, but custom Apache Beam pipelines require programming. A developer must define parsing, validation, branching, and transformation logic in a supported Beam SDK. That conflicts with the requirement that the data-cleaning workflow not require programming or SQL knowledge.</p><blockquote><strong><em>B is correct because BigQuery data preparation is the current visual data-cleaning capability in BigQuery. It provides a visual editor, sample-based previews, transformation suggestions, and data-quality assistance so users can prepare data without manually authoring a full transformation program. BigQuery data preparation supports external tables as sources, and BigQuery external tables can reference structured data stored in Cloud Storage. Therefore the historical objects can remain in Cloud Storage while being exposed to the visual preparation workflow. This is the current Google Cloud visual data-preparation workflow for this requirement.</em></strong></blockquote><p>C is incorrect because Managed Service for Apache Spark can run Hadoop and Spark processing at scale, but a custom job requires programming or scripting and operational knowledge of the processing framework. It is appropriate for engineered big-data pipelines, not for a user who specifically needs a visual preparation experience without coding.</p><p>D is incorrect because a BigQuery external table is an appropriate way to expose Cloud Storage data to BigQuery, but manually writing GoogleSQL directly violates the requirement that users not need SQL knowledge. BigQuery data preparation provides a visual layer over the cleansing and transformation process and is therefore the better current solution.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/data-prep-introduction">docs.cloud.google.com/bigquery/docs/data-prep-introduction</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/external-data-cloud-storage">docs.cloud.google.com/bigquery/docs/external-data-cloud-storage</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/transform-intro">docs.cloud.google.com/bigquery/docs/transform-intro</a></p><p><strong>Q7.</strong></p><p><strong>At Raventrix Health, you need to copy millions of sensitive patient records from an on-premises relational database to BigQuery. The database contains approximately 10 TB of data. You need a secure and time-efficient solution. What should you do?</strong></p><p><strong>A. Export the records from the database as Avro files. Upload the files securely to Cloud Storage using the current gcloud storage tooling, and then load the Avro data into BigQuery from the Google Cloud console or an automated load job.</strong></p><p><strong>B. Export the records from the database as Avro files. Copy the files onto a Transfer Appliance, physically ship the appliance to Google, and then load the data into BigQuery.</strong></p><p><strong>C. Export the records from the database as CSV files. Create public URLs for the CSV files, use Storage Transfer Service to move them to Cloud Storage, and then load them into BigQuery.</strong></p><p><strong>D. Export the records from the database as Avro files. Create public URLs for the Avro files, use Storage Transfer Service to move them to Cloud Storage, and then load them into BigQuery.</strong></p><blockquote><strong><em>A is correct because Avro is a self-describing binary format that preserves schema information and is directly supported by BigQuery load jobs. Cloud Storage uploads are encrypted in transit, can use resumable and parallel transfer behavior, and avoid deliberately exposing sensitive patient files through public URLs. With adequate network bandwidth, an online transfer also avoids the shipping and handling delay of an offline appliance. For very large production migrations, current Google Cloud guidance recommends evaluating Storage Transfer Service for on-premises transfers above 1 TB, but among these choices the secure online Avro-to-Cloud-Storage path is the appropriate design because no poor-connectivity constraint is stated.</em></strong></blockquote><p>B is incorrect because Transfer Appliance is primarily intended when the dataset would take too long to transfer over the available network, especially when connectivity is poor or insufficient to meet a deadline. A 10 TB dataset can fit on a Transfer Appliance, but the question does not state that the network is inadequate. Physical delivery, data capture, return shipping, and ingestion add logistical latency, so an appliance is not automatically the fastest choice merely because the dataset is large.</p><p>C is incorrect because creating a public URL for sensitive patient information is an unacceptable security design. Storage Transfer Service can ingest from public URLs, but making regulated health data publicly reachable just to facilitate the transfer unnecessarily exposes the data. CSV also lacks the schema-rich characteristics of Avro and can require more explicit type handling during load.</p><p>D is incorrect because Avro is an appropriate file format, but publishing sensitive patient files through public URLs still violates the security requirement. The fact that Storage Transfer Service can retrieve publicly accessible objects does not justify making confidential records public. Secure authenticated transfer to Cloud Storage should be used instead.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/storage-transfer/docs/transfer-options">docs.cloud.google.com/storage-transfer/docs/transfer-options</a></p><p><a href="http://docs.cloud.google.com/storage/docs/parallel-composite-uploads">docs.cloud.google.com/storage/docs/parallel-composite-uploads</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/loading-data-cloud-storage-avro">docs.cloud.google.com/bigquery/docs/loading-data-cloud-storage-avro</a></p><p><strong>Q8.</strong></p><p><strong>At Brinovex Partners, you use a BigQuery dataset for analysis and want to provide third-party companies with access to the same dataset. You need to keep data-sharing costs low and ensure that subscribers see current data. Which solution should you choose?</strong></p><p><strong>A. Use BigQuery sharing to control data access and provide the third-party companies with subscriptions to the shared dataset.</strong></p><p><strong>B. Use Cloud Scheduler to export the data regularly to Cloud Storage and provide the third-party companies with access to the bucket.</strong></p><p><strong>C. Create a separate BigQuery dataset containing copies of the relevant data and provide the third-party companies with access to the copied dataset.</strong></p><p><strong>D. Create a Dataflow job that frequently reads the source data and writes copies to separate BigQuery datasets or Cloud Storage buckets for the third-party companies.</strong></p><blockquote><strong><em>A is correct because BigQuery sharing is designed for secure data exchange across project and organizational boundaries without requiring the publisher to replicate the underlying dataset for every subscriber. Subscribers receive linked, read-only access to the shared data, so they can query current publisher data while storage remains centrally managed. Avoiding duplicated datasets reduces storage and pipeline costs and also eliminates synchronization delays that would otherwise make copied data stale.</em></strong></blockquote><p>B is incorrect because regularly exporting BigQuery data to Cloud Storage creates a separate copy that must be scheduled, stored, secured, and refreshed. Third parties would see data only as current as the most recent successful export, and the publisher would incur additional storage and operational overhead. BigQuery sharing provides direct controlled access without this replication step.</p><p>C is incorrect because creating a separate BigQuery dataset with copied data adds duplicate storage and requires a mechanism to keep the copy synchronized with the original dataset. If the copy is refreshed periodically, it can be stale between refreshes. BigQuery sharing provides the same cross-organizational consumption model without duplicating the source data.</p><p>D is incorrect because a Dataflow replication pipeline adds continuous compute cost, pipeline operations, duplicate storage, and failure-handling complexity. Dataflow is valuable when records need transformation or streaming processing, but it is unnecessary when the requirement is simply to share an existing BigQuery dataset securely and keep the shared data current.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/analytics-hub-introduction">docs.cloud.google.com/bigquery/docs/analytics-hub-introduction</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/analytics-hub-view-subscribe-listings">docs.cloud.google.com/bigquery/docs/analytics-hub-view-subscribe-listings</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/analytics-hub-manage-listings">docs.cloud.google.com/bigquery/docs/analytics-hub-manage-listings</a></p><p><strong>Q9.</strong></p><p><strong>At Elarion Holdings, you need to configure BigQuery access for several departments. Each department must access only its own data. Department leads must be able to create and update tables, while department analysts must be able to query but not modify the data. How should you configure access?</strong></p><p><strong>A. Create one dataset per department. Grant department leads BigQuery Data Owner and grant analysts BigQuery Data Editor on their department’s dataset.</strong></p><p><strong>B. Create one dataset per department. Grant department leads BigQuery Data Editor and grant analysts BigQuery Data Viewer on their department’s dataset.</strong></p><p><strong>C. Create one table per department. Grant department leads the Owner basic role and analysts the Editor basic role on the project containing the tables.</strong></p><p><strong>D. Create one table per department. Grant department leads the Editor basic role and analysts the Viewer basic role on the project containing the tables.</strong></p><p>A is incorrect because BigQuery Data Owner gives the leads more control than the requirement calls for, while BigQuery Data Editor gives analysts write permissions that they must not have. Analysts need read/query access only. The role pairing therefore violates the separation between data maintainers and read-only analysts.</p><blockquote><strong><em>B is correct because separate datasets provide a clean IAM boundary for each department. BigQuery Data Editor at the dataset level lets department leads create, update, and delete tables and modify table data, while BigQuery Data Viewer lets analysts read and query the dataset without modifying its tables. Query execution also requires permission to create BigQuery jobs, typically through BigQuery Job User on an appropriate project, but the dataset-level data roles in this option correctly implement the requested data-access separation.</em></strong></blockquote><p>C is incorrect because project-level Owner and Editor are broad basic IAM roles that grant permissions far beyond the department’s BigQuery tables. They also make it difficult to enforce the requirement that each department access only its own data. Google Cloud recommends predefined, narrowly scoped IAM roles rather than broad basic roles for production access control.</p><p>D is incorrect because project-level Editor and Viewer permissions are broader than necessary and operate at the wrong resource boundary for departmental isolation. A project Viewer can receive visibility into many resources, and a project Editor can modify far more than one department’s BigQuery data. Dataset-level predefined BigQuery roles provide a more precise least-privilege design.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/control-access-to-resources-iam">docs.cloud.google.com/bigquery/docs/control-access-to-resources-iam</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/access-control">docs.cloud.google.com/bigquery/docs/access-control</a></p><p><a href="http://docs.cloud.google.com/iam/docs/using-iam-securely">docs.cloud.google.com/iam/docs/using-iam-securely</a></p><p><strong>Q10.</strong></p><p><strong>At Gavorix Streaming, you operate a Dataflow streaming pipeline that aggregates events from a Pub/Sub subscription within windows and writes the resulting aggregates to a Cloud Storage bucket. The source normally has consistent throughput. You want Cloud Monitoring in Google Cloud Observability to alert you when the pipeline appears to stop making progress. Which alert pattern should you create?</strong></p><p><strong>A. Alert when the Pub/Sub subscription’s num_undelivered_messages metric decreases while the Cloud Storage bucket’s total-bytes growth rate increases.</strong></p><p><strong>B. Alert when the Pub/Sub subscription’s num_undelivered_messages metric increases while the Cloud Storage bucket’s total-bytes growth rate decreases or stops.</strong></p><p><strong>C. Alert when a Cloud Storage source-size metric decreases and a Pub/Sub destination-backlog metric increases.</strong></p><p><strong>D. Alert when a Cloud Storage source-size metric increases and a Pub/Sub destination-backlog metric decreases.</strong></p><p>A is incorrect because these signals generally describe a healthy pipeline. A falling or controlled Pub/Sub backlog means messages are being consumed, and continuing growth in the destination bucket is evidence that the pipeline is producing output. An alert intended to detect stalled processing should watch for the opposite trend.</p><blockquote><strong><em>B is correct because a rising pubsub.googleapis.com/subscription/num_undelivered_messages value indicates that the source subscription backlog is growing, while a falling or near-zero rate of growth in Cloud Storage output indicates that the sink is no longer receiving expected results. When input volume is normally consistent, those two signals together strongly suggest that the streaming pipeline is stuck, severely delayed, or no longer processing at the expected rate. Dataflow system-lag and job-status metrics can also be added to a modern monitoring policy for faster diagnosis.</em></strong></blockquote><p>C is incorrect because the resource directions are reversed. In this pipeline Pub/Sub is the source and Cloud Storage is the destination, so a Cloud Storage source-size metric and Pub/Sub destination backlog do not describe the actual topology.</p><p>D is incorrect because it likewise reverses the source and destination roles. Even though the mathematical trend could describe activity in some other architecture, Pub/Sub is the incoming source here and Cloud Storage is the output sink. Monitoring must use metrics associated with the actual source and sink.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/dataflow/docs/guides/using-cloud-monitoring">docs.cloud.google.com/dataflow/docs/guides/using-cloud-monitoring</a></p><p><a href="http://docs.cloud.google.com/pubsub/docs/monitoring">docs.cloud.google.com/pubsub/docs/monitoring</a></p><p><a href="http://docs.cloud.google.com/storage/docs/getting-bucket-size">docs.cloud.google.com/storage/docs/getting-bucket-size</a></p><p><strong>Q11.</strong></p><p><strong>At Lavorent Commerce, you need to create a Spanner transaction table that stores product-sales records. You are deciding which primary-key strategy provides the best write distribution and avoids hotspots. Which strategy should you choose?</strong></p><p><strong>A. Use the current epoch timestamp as the primary key.</strong></p><p><strong>B. Use a concatenation of the product name followed by the current epoch timestamp.</strong></p><p><strong>C. Use a random version 4 universally unique identifier (UUID) as the primary key.</strong></p><p><strong>D. Use the original sales-system order number, which is a monotonically increasing integer, as the primary key.</strong></p><p>A is incorrect because an ever-increasing timestamp creates a monotonic primary key. Spanner stores rows in primary-key order and distributes contiguous key ranges among servers. New inserts with increasing keys therefore concentrate at the end of the key space, directing a disproportionate amount of write traffic to one range and creating a hotspot.</p><p>B is incorrect because placing product name first can distribute writes across products, but the distribution depends heavily on the number and popularity of products. A very popular product can still receive a large sequence of writes whose timestamp suffix increases monotonically inside the same product prefix, creating a hot key range. It is safer to use a key whose high-order bits are well distributed when no natural evenly distributed key exists.</p><blockquote><strong><em>C is correct because a version 4 UUID contains random high-order bits and spreads newly inserted rows broadly across Spanner’s key space. Google Cloud explicitly recommends UUID version 4 as one method for avoiding hotspots when generating surrogate keys. The random distribution allows writes to be served by multiple splits instead of concentrating all new rows in one sequential key range.</em></strong></blockquote><p>D is incorrect because a monotonically increasing order ID has the same hotspot problem as an epoch timestamp. Every new sale is inserted near the end of the key range, so a high insert rate can overload the server currently responsible for that range. If the original order ID must be retained for business purposes, it can remain a separate indexed column while the primary key uses a better-distributed value or a hash/bit-reversed strategy.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/spanner/docs/schema-design">docs.cloud.google.com/spanner/docs/schema-design</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/schema-and-data-model">docs.cloud.google.com/spanner/docs/schema-and-data-model</a></p><p><a href="http://docs.cloud.google.com/spanner/docs/whitepapers/optimizing-schema-design">docs.cloud.google.com/spanner/docs/whitepapers/optimizing-schema-design</a></p><p><strong>Q12.</strong></p><p><strong>At Solviora Dataworks, you want to migrate an on-premises Hadoop system to Managed Service for Apache Spark. Apache Hive is the primary tool in use, and the data format is Optimized Row Columnar (ORC). All ORC files have already been copied to a Cloud Storage bucket. You need to replicate some of the data to the cluster’s local Hadoop Distributed File System (HDFS) to maximize performance. What are two ways to start using Hive in Managed Service for Apache Spark? (Choose two.)</strong></p><p><strong>A. Use gcloud storage copy operations to copy the ORC files directly from the Cloud Storage bucket to an hdfs:// destination, and then create Hive tables over the copied files.</strong></p><p><strong>B. Use gcloud storage copy operations to copy the ORC files from Cloud Storage to the local file system of any worker node, and create Hive tables over those node-local files.</strong></p><p><strong>C. Use gcloud storage copy operations to copy the ORC files from Cloud Storage to the local file system of the cluster’s driver node, then use Hadoop file-system commands to copy the files into HDFS and create Hive tables over the HDFS data.</strong></p><p><strong>D. Use the Cloud Storage connector for Hadoop to create external Hive tables directly over the ORC files in Cloud Storage, and then copy selected data from those external tables into native Hive tables backed by HDFS.</strong></p><p><strong>E. Load the ORC files into BigQuery, use the Hive-BigQuery connector to expose the BigQuery tables to Hive, and then replicate the data from BigQuery-backed Hive tables into native HDFS-backed Hive tables.</strong></p><p>A is incorrect because gcloud storage commands operate on Cloud Storage and local file-system paths; they are not Hadoop file-system commands that directly write an hdfs:// destination. Managed Service for Apache Spark includes an HDFS-compatible Cloud Storage connector, so Hadoop tools can work with gs:// paths, but a Cloud Storage CLI copy command is not the direct Cloud-Storage-to-HDFS mechanism described in this option. A valid local-staging workflow requires an additional Hadoop copy step as described in option C.</p><p>B is incorrect because copying data to the local disk of an arbitrary worker does not place that data in HDFS. Node-local files are visible only on that VM unless explicitly distributed, so Hive jobs running on other workers cannot reliably treat the local path as a distributed table location. HDFS or Cloud Storage should provide the distributed storage namespace that all cluster nodes can access.</p><blockquote><strong><em>C is correct because the files can first be staged onto a cluster node’s local file system with current Cloud Storage tooling and then copied into HDFS using Hadoop file-system commands. Once the files reside in HDFS, Hive can define managed or external tables over those HDFS locations. This explicitly satisfies the requirement to replicate selected source data to cluster-local HDFS for workloads where local distributed storage provides better performance.</em></strong></blockquote><blockquote><strong><em>D is correct because Managed Service for Apache Spark automatically provides the Cloud Storage connector for Hadoop, which allows Hadoop and Hive workloads to address Cloud Storage objects through gs:// paths without first loading all of the data into HDFS. Hive can therefore create an external table over the ORC files in Cloud Storage. The organization can then select only the data that benefits from local placement and write that subset into a native Hive table stored in HDFS, preserving Cloud Storage as the durable source while accelerating selected workloads locally.</em></strong></blockquote><p>E is incorrect because current Hive-BigQuery integration can allow Hive workloads to access data stored in BigQuery, so the individual technologies are compatible. However, loading the ORC data into BigQuery solely to move it back through Hive and into HDFS adds an unnecessary storage system, data-load step, and cost when the original ORC objects are already directly accessible through the Cloud Storage connector. The two direct migration paths are to copy selected files into HDFS or to expose the Cloud Storage files as external Hive data and materialize only the required subset locally.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/managed-spark/docs/concepts/connectors/cloud-storage">docs.cloud.google.com/managed-spark/docs/concepts/connectors/cloud-storage</a></p><p><a href="http://docs.cloud.google.com/managed-spark/docs/concepts/hdfs">docs.cloud.google.com/managed-spark/docs/concepts/hdfs</a></p><p><a href="http://docs.cloud.google.com/managed-spark/docs/guides/hive-lineage">docs.cloud.google.com/managed-spark/docs/guides/hive-lineage</a></p><p><strong>Q13.</strong></p><p><strong>At Varelium Insights, you are migrating a data warehouse to BigQuery. All data has been migrated into tables in one dataset. Multiple teams in your organization will use the data, but each team should be able to read only the specific whole tables assigned to that team. Using current BigQuery access-control capabilities, how should you set permissions with the least additional data-modeling overhead?</strong></p><p><strong>A. Grant the appropriate users or groups BigQuery Data Viewer access directly on each table they are allowed to read.</strong></p><p><strong>B. Create normal SQL views for each team in the same dataset as the source tables, and grant users or groups BigQuery Data Viewer access to those views.</strong></p><p><strong>C. Create authorized views for each team in the same dataset as the source tables, and grant users or groups BigQuery Data Viewer access to the authorized views.</strong></p><p><strong>D. Create authorized views for each team in separate datasets, authorize those views to access the source dataset, and grant users or groups BigQuery Data Viewer access to the datasets that contain their views.</strong></p><blockquote><strong><em>A is correct because current BigQuery IAM supports fine-grained access policies directly on individual tables and views. The BigQuery Data Viewer role can be granted on a specific table to a user or group, allowing that principal to read that table without automatically granting access to every other table in the dataset. Because the requirement is to expose complete tables rather than selected rows or columns, direct table-level IAM is the simplest current solution and avoids creating and maintaining redundant views. Users also need the ability to create BigQuery query jobs in the project where their queries execute.</em></strong></blockquote><p>B is incorrect because a normal view does not independently authorize access to underlying tables. A principal querying a standard view generally needs the necessary permissions on the data referenced by that view unless the view is configured as an authorized resource. Creating ordinary views therefore does not solve the access-isolation requirement by itself and adds unnecessary objects when direct table-level IAM already exists.</p><p>C is incorrect because BigQuery documentation recommends creating authorized views in a different dataset from the source data. If the team is granted dataset-level access to a dataset that also contains the protected source tables, the intended separation can be defeated through inherited access. In addition, authorized views are primarily useful when you need to expose a filtered or projected subset of source data rather than an entire table that can now be permissioned directly.</p><p>D is incorrect because authorized views in separate datasets are a valid and important BigQuery pattern when users should see only selected rows or columns from underlying tables without receiving direct source access. However, the requirement here is only to control which complete tables each team can read. Current table-level IAM does that directly with less administrative overhead, so constructing team-specific authorized-view datasets is unnecessary for this scenario.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/control-access-to-resources-iam">docs.cloud.google.com/bigquery/docs/control-access-to-resources-iam</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/access-control">docs.cloud.google.com/bigquery/docs/access-control</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/authorized-views">docs.cloud.google.com/bigquery/docs/authorized-views</a></p><p><strong>Q14.</strong></p><p><strong>At Zenvorix IoT, you operate an IoT pipeline built around Apache Kafka that normally receives approximately 5,000 messages per second. You want to create an alert as soon as the moving average over the most recent one hour drops below 4,000 messages per second. What should you do?</strong></p><p><strong>A. Consume the Kafka stream in Dataflow using KafkaIO. Use a sliding event-time window with a one-hour window size and a five-minute period. Compute the rate for each window and send an alert when the one-hour average falls below 4,000 messages per second.</strong></p><p><strong>B. Consume the Kafka stream in Dataflow using KafkaIO. Use fixed one-hour windows, compute the rate when each fixed window closes, and send an alert when the average falls below 4,000 messages per second.</strong></p><p><strong>C. Use Kafka Connect to send Kafka records to Pub/Sub, use Dataflow to write them to Bigtable, and use Cloud Scheduler once per hour to count recently created rows and alert when the count falls below 4,000.</strong></p><p><strong>D. Use Kafka Connect to send Kafka records to Pub/Sub, use Dataflow to write them to BigQuery, and use Cloud Scheduler every five minutes to count rows from the last hour and alert when the count falls below 4,000.</strong></p><blockquote><strong><em>A is correct because the requirement is a moving one-hour average rather than an average over disjoint one-hour periods. Apache Beam sliding windows allow each element to participate in overlapping one-hour windows that begin at a regular interval, such as every five minutes. Running the pipeline on Dataflow with KafkaIO provides continuous stream processing, while the five-minute window period means the system can reevaluate the most recent one-hour interval much more frequently than once per hour. The calculation should convert the number of events in each one-hour window into a per-second rate before comparing it with the 4,000-messages-per-second threshold.</em></strong></blockquote><p>B is incorrect because fixed windows divide the stream into non-overlapping one-hour blocks. The system would normally produce a new one-hour aggregate only as each hourly window is completed according to its trigger and watermark behavior, so a degradation that begins shortly after a window boundary might not be detected until close to an hour later. Sliding windows are the appropriate abstraction for a continuously updated moving average.</p><p>C is incorrect because it introduces Pub/Sub, Bigtable, Cloud Scheduler, and custom polling even though Dataflow can compute the streaming metric directly. The proposed threshold is also dimensionally wrong: 4,000 is a messages-per-second rate, while counting rows created during an entire hour produces a total number of messages. A healthy 4,000-messages-per-second hour corresponds to approximately 14.4 million events, not 4,000 rows.</p><p>D is incorrect because although querying the preceding hour every five minutes resembles a moving-window evaluation, the architecture needlessly persists every message to BigQuery and repeatedly scans recent data just to calculate a streaming rate. It also states the threshold in terms of row count rather than converting the one-hour count to messages per second. Dataflow windowing computes the metric continuously in the streaming pipeline with lower latency and fewer moving parts.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines">docs.cloud.google.com/dataflow/docs/concepts/streaming-pipelines</a></p><p><a href="http://beam.apache.org/documentation/programming-guide">beam.apache.org/documentation/programming-guide</a></p><p><a href="http://beam.apache.org/releases/javadoc/current/org/apache/beam/sdk/io/kafka/KafkaIO.html">beam.apache.org/releases/javadoc/current/org/apache/beam/sdk/io/kafka/KafkaIO.html</a></p><p><strong>Q15.</strong></p><p><strong>At Maventra Cloud Services, a customer requests daily reports showing use of Google Cloud compute resources and the identities responsible for that activity. You need to make the relevant Cloud Logging data easy to query and turn it into repeatable daily reports as quickly as possible. What should you do?</strong></p><p><strong>A. Create a Cloud Logging Log Router sink that routes the relevant logs to BigQuery, and create BigQuery views that filter or aggregate the data by project, log type, resource, and principal.</strong></p><p><strong>B. Manually filter the data in Cloud Logging by project, resource, and user every day and export the visible results as CSV files.</strong></p><p><strong>C. Manually filter Cloud Logging by project, log type, resource, and user, export the results, and then import the exported data into BigQuery each day.</strong></p><p><strong>D. Route the logs to Cloud Storage as files and use BigQuery data preparation to cleanse and filter the files by project, resource, and user before creating each report.</strong></p><blockquote><strong><em>A is correct because Cloud Logging can route matching log entries continuously to a BigQuery dataset through a Log Router sink. Once the logs are available in BigQuery, reusable views can extract fields such as resource information and authenticated principal identity and can aggregate the activity needed for the customer’s daily report. This removes manual daily export and import steps and lets the reporting queries run repeatedly over a structured analytical destination. For new log-analytics designs, Google Cloud also recommends considering log buckets with Observability Analytics and a linked BigQuery dataset, but among the listed choices a managed Logging-to-BigQuery route plus views is the direct automated reporting solution.</em></strong></blockquote><p>B is incorrect because manually filtering and exporting CSV files every day is labor intensive, difficult to reproduce consistently, and poorly suited to ongoing reporting. It also separates each day’s output into files that require additional processing if the customer later wants trend analysis or comparisons across reporting periods.</p><p>C is incorrect because this approach eventually places the logs in BigQuery, but the daily manual filter-export-import cycle is unnecessary. Cloud Logging sinks can route matching records directly into BigQuery as they arrive, eliminating the operational delay and repeated human work.</p><p>D is incorrect because Cloud Storage is useful for long-term log archives, and BigQuery data preparation can help cleanse datasets, but this design adds file-based staging and preparation work before every report. When the goal is fast repeatable analytics over structured logging fields, direct routing to BigQuery is simpler and allows views and scheduled queries to perform the reporting transformations.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/logging/docs/export/configure_export_v2">docs.cloud.google.com/logging/docs/export/configure_export_v2</a></p><p><a href="http://docs.cloud.google.com/logging/docs/export/bigquery">docs.cloud.google.com/logging/docs/export/bigquery</a></p><p><a href="http://docs.cloud.google.com/logging/docs/audit">docs.cloud.google.com/logging/docs/audit</a></p><p><strong>Q16.</strong></p><p><strong>At Asteronix Logistics, you need to migrate 1 PB of data from an on-premises data center to Google Cloud. The data-transfer phase of the migration must complete within only a few hours. You want to follow Google-recommended practices for transferring the data at very high throughput over a secure connection. What should you do?</strong></p><p><strong>A. Establish a high-capacity Cloud Interconnect connection between the on-premises data center and Google Cloud, and use Storage Transfer Service to transfer the data.</strong></p><p><strong>B. Use Transfer Appliance and have engineers manually encrypt, decrypt, and verify the data.</strong></p><p><strong>C. Establish a Cloud VPN connection, start gcloud compute scp jobs in parallel, and run checksums to verify the data.</strong></p><p><strong>D. Reduce the data into 3 TB batches, transfer the data with gcloud storage commands, and run checksums to verify the data.</strong></p><blockquote><strong><em>A is correct because transferring 1 PB in only a few hours requires aggregate throughput far beyond what a typical internet or VPN connection can provide. Cloud Interconnect provides dedicated or partner connectivity between an on-premises network and Google’s network, and current Dedicated Interconnect configurations can scale to very high aggregate bandwidth by using multiple high-capacity circuits. Storage Transfer Service is optimized for transfers larger than 1 TiB and supports highly parallelized, reliable transfer of petabyte-scale data with automatic retries and integrity validation. Its file-system transfer agents communicate securely with Google Cloud over TLS. The Interconnect capacity must be sized appropriately for the required transfer window, and a production design should also use the required redundancy and encryption controls for the organization’s security policy.</em></strong></blockquote><p>B is incorrect because Transfer Appliance is designed for very large offline migrations when available network bandwidth is inadequate, but physical appliance workflows include data capture, shipping, ingestion, and return logistics. The requirement says that the data-transfer phase should complete in only a few hours, so an offline appliance workflow is not the appropriate mechanism when sufficiently high-bandwidth online connectivity can be provisioned. Google also provides encryption and integrity mechanisms as part of the managed transfer workflow, so having engineers manually implement all encryption, decryption, and verification is unnecessary.</p><p>C is incorrect because Cloud VPN is not intended to provide the aggregate throughput needed to move 1 PB in a few hours, and parallel gcloud compute scp operations are not the Google-recommended bulk data-transfer mechanism for a petabyte-scale migration. SCP is primarily a VM file-copy mechanism and would require custom orchestration, retries, and verification. Storage Transfer Service is specifically designed to move large datasets reliably and in parallel.</p><p>D is incorrect because dividing the data into 3 TB batches does not solve the fundamental bandwidth requirement. Current gcloud storage commands can upload objects efficiently, but manually managing hundreds of multi-terabyte batches, retries, and checksums is unnecessary for a petabyte-scale migration. Storage Transfer Service provides managed parallelization, retries, and integrity checking, while Cloud Interconnect provides the high-capacity network path required to meet the very short transfer window.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/network-connectivity/docs/interconnect/concepts/dedicated-overview">docs.cloud.google.com/network-connectivity/docs/interconnect/concepts/dedicated-overview</a></p><p><a href="http://docs.cloud.google.com/network-connectivity/docs/interconnect/quotas">docs.cloud.google.com/network-connectivity/docs/interconnect/quotas</a></p><p><a href="http://docs.cloud.google.com/storage-transfer/docs/overview">docs.cloud.google.com/storage-transfer/docs/overview</a></p><p><strong>Q17.</strong></p><p><strong>At Cendorva Commerce, you are developing a deep-learning model that predicts a customer’s likelihood of purchasing from your ecommerce site. Evaluation on the training data and a separate test dataset shows that the model is overfitting. You want to improve the model’s ability to generalize to new data. What should you do?</strong></p><p><strong>A. Increase the size of the training dataset and increase the number of input features.</strong></p><p><strong>B. Increase the size of the training dataset and decrease the number of input features.</strong></p><p><strong>C. Reduce the size of the training dataset and increase the number of input features.</strong></p><p><strong>D. Reduce the size of the training dataset and decrease the number of input features.</strong></p><p>A is incorrect because adding more representative training examples is a standard way to reduce overfitting, but increasing the feature count can increase model complexity and create additional opportunities to fit noise or accidental correlations. New features can improve a model when they add meaningful predictive signal, but blindly adding features is not a standard corrective action for a model that already has excessive variance.</p><blockquote><strong><em>B is correct because both changes generally move an overfit model toward better generalization. More representative training examples give the model more evidence about the underlying population and make it harder to memorize idiosyncrasies of a small sample. Reducing the feature set can lower effective model complexity by removing noisy, redundant, or weakly predictive dimensions. Together these actions reduce the model’s ability to fit incidental patterns while preserving the opportunity to learn signals that repeat across many examples.</em></strong></blockquote><p>C is incorrect because reducing the number of training examples gives the model less evidence about the true data distribution, while adding features increases the space of patterns that the model can fit. Both changes tend to increase the risk of overfitting rather than reduce it.</p><p>D is incorrect because decreasing the number of features can help simplify an overfit model, but simultaneously reducing the amount of training data removes information that is useful for generalization. With fewer examples, even a smaller model can become more sensitive to sampling noise. If more representative training data is available, increasing rather than decreasing the training set is the better direction.</p><p><strong>Reference links:</strong></p><p><a href="http://developers.google.com/machine-learning/crash-course/overfitting/overfitting">developers.google.com/machine-learning/crash-course/overfitting/overfitting</a></p><p><a href="http://developers.google.com/machine-learning/crash-course/overfitting/model-complexity">developers.google.com/machine-learning/crash-course/overfitting/model-complexity</a></p><p><a href="http://developers.google.com/machine-learning/crash-course/overfitting/data-characteristics">developers.google.com/machine-learning/crash-course/overfitting/data-characteristics</a></p><p><strong>Q18.</strong></p><p><strong>At Kendoris Insights, you use BigQuery and Looker Studio to create a customer-facing dashboard that displays large quantities of aggregated data. You expect many concurrent dashboard users and need the visualizations to respond with minimal latency. What should you do?</strong></p><p><strong>A. Use BigQuery BI Engine together with materialized views that precompute the frequently requested aggregations.</strong></p><p><strong>B. Use BigQuery BI Engine together with logical views for the dashboard queries.</strong></p><p><strong>C. Use BigQuery BI Engine and rely only on the fact that the underlying tables receive streaming data.</strong></p><p><strong>D. Use BigQuery BI Engine together with authorized views solely to accelerate the dashboard.</strong></p><blockquote><strong><em>A is correct because BI Engine is BigQuery’s in-memory analytical acceleration service and is specifically intended to improve the latency of business-intelligence queries. Materialized views precompute and incrementally maintain results for suitable recurring queries, such as the aggregates repeatedly requested by a dashboard. Current BigQuery guidance explicitly recommends both BI Engine and materialized views for performance optimization, and BI Engine works best with pre-joined or pre-aggregated data. Combining these features reduces repeated computation and provides a strong architecture for a highly concurrent customer-facing dashboard.</em></strong></blockquote><p>B is incorrect because a logical view stores only its SQL definition and normally executes the underlying query when the view is queried. It can simplify the semantic layer, but it does not precompute the large aggregates in the way a materialized view does. BI Engine can accelerate many supported queries involving views, but a logical view alone does not provide the same reduction in repeated aggregation work.</p><p>C is incorrect because streaming data is an ingestion characteristic, not a dashboard-performance optimization by itself. Continuously arriving rows do not automatically precompute dashboard metrics or guarantee lower query latency. BI Engine can accelerate supported queries over fresh data, but the repeatedly requested aggregates should also be optimized with materialization when appropriate.</p><p>D is incorrect because authorized views are primarily an access-control feature that lets users query a controlled subset of source data without receiving direct access to the underlying tables. BI Engine can coexist with authorized views, but authorization does not itself precompute repeated aggregates or reduce the work required by a high-concurrency dashboard. Materialized views target the stated performance problem directly.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/bi-engine-intro">docs.cloud.google.com/bigquery/docs/bi-engine-intro</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/materialized-views-intro">docs.cloud.google.com/bigquery/docs/materialized-views-intro</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/best-practices-performance-compute">docs.cloud.google.com/bigquery/docs/best-practices-performance-compute</a></p><p><strong>Q19.</strong></p><p><strong>At Westralis Data, your organization is implementing a data-mesh operating model. Multiple independent data-engineering teams build separate data products. Each product follows a common lifecycle: source files first land in Cloud Storage, raw data is transformed in Cloud Storage or BigQuery, and a final curated BigQuery dataset is shared with consumers. You want to organize the products in Knowledge Catalog so that each team can manage only the assets for its own data product and can expose its curated product to other teams. What should you do?</strong></p><p><strong>A. Create one Knowledge Catalog lake for the entire organization and use a single zone for all landing, raw, and curated assets. Grant each engineering team access to the whole lake.</strong></p><p><strong>B. Create one Knowledge Catalog lake for the entire organization and one zone containing landing, raw, and curated assets for all products. Create separate assets for each product and assign permissions only at the common zone level.</strong></p><p><strong>C. Create one Knowledge Catalog lake for each data product, use one zone in each lake for all landing, raw, and curated assets, and grant the owning team access to the lake.</strong></p><p><strong>D. Create one Knowledge Catalog lake for each data product, create separate zones that reflect the landing/raw and curated stages, and grant each engineering team control over its own lake while granting consumers access to the appropriate curated zone or assets.</strong></p><p>A is incorrect because putting every product and every lifecycle stage into one lake and one zone creates an overly broad administrative and access boundary. Teams would either receive more access than they need or require extensive asset-by-asset exceptions. It also fails to model the separate ownership boundaries that are central to a data-mesh approach.</p><p>B is incorrect because separate assets help identify individual resources, but one common lake and one common zone still make zone-level governance too coarse. If permissions are assigned at that shared zone, teams can inherit access to assets belonging to other data products. The design does not provide clean domain or product ownership.</p><p>C is incorrect because using one lake per data product establishes a useful ownership boundary, but collapsing landing, raw, and curated assets into one zone prevents the organization from assigning different governance and consumer-access policies to data at different maturity levels. Raw producer data and curated consumable data normally have different trust and sharing requirements.</p><blockquote><strong><em>D is correct because a lake can represent a data domain or product boundary, while zones can represent stages with different governance characteristics, such as raw/landing data and curated data. Giving each engineering team control of its own lake isolates product ownership, and granting consumers access only to the curated zone or specific curated assets supports controlled data-product sharing. This aligns the catalog structure with the organizational structure and with the transition from raw producer data to governed consumer-ready data.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/architecture/design-self-service-data-platform-data-mesh">docs.cloud.google.com/architecture/design-self-service-data-platform-data-mesh</a></p><p><a href="http://cloud.google.com/discover/what-is-data-mesh">cloud.google.com/discover/what-is-data-mesh</a></p><p><strong>Q20.</strong></p><p><strong>At Bravion Partners, you have a BigQuery dataset protected with customer-managed encryption keys (CMEK). You need to share the data with a partner organization, and the partner should not receive direct access to your encryption keys. Which option provides a straightforward sharing design while preserving key separation?</strong></p><p><strong>A. Give the partner organization a copy of your customer-managed encryption keys so that it can decrypt the shared BigQuery tables.</strong></p><p><strong>B. Export the BigQuery data to Parquet files in Cloud Storage and grant the partner the Storage Insights Viewer role on the bucket.</strong></p><p><strong>C. Copy the tables into a separate BigQuery dataset whose tables do not use your restricted CMEK, and publish that dataset to the partner through BigQuery sharing.</strong></p><p><strong>D. Create an authorized view over the encrypted tables and embed the CMEK or key material in the view definition so that the partner can decrypt the source tables.</strong></p><p>A is incorrect because customer-managed encryption keys should not be copied or disclosed to a partner merely to enable data consumption. CMEK is specifically intended to keep key lifecycle and authorization under the data owner’s control. Sharing raw key material destroys that separation of duties and creates unnecessary compromise risk.</p><p>B is incorrect because exporting Parquet files creates another physical copy and changes the consumption model from BigQuery sharing to object access. More importantly, the Storage Insights Viewer role provides access to storage-insight metadata rather than ordinary read access to the bucket’s objects, so it does not give the partner the data access described by the option.</p><blockquote><strong><em>C is correct because the publisher can create a dedicated shareable dataset whose encryption configuration is independent of the restricted source-key policy and then publish that dataset through BigQuery sharing. BigQuery sharing lets a provider expose datasets to subscribers without granting those subscribers direct access to the publisher’s broader project. Current BigQuery sharing also supports many CMEK-protected sharing scenarios, so a copy is not universally required today; however, when the security requirement is specifically to keep the partner completely independent of the restricted source-key authorization model, a separately prepared shareable dataset is the valid listed design.</em></strong></blockquote><p>D is incorrect because an authorized view can expose selected data without granting direct access to the underlying source tables, but a view definition cannot contain or distribute a customer-managed encryption key. BigQuery’s service identities use Cloud KMS according to configured IAM permissions; decryption keys are not embedded in SQL. Authorized views solve data-access scoping, not key-material delivery.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/customer-managed-encryption">docs.cloud.google.com/bigquery/docs/customer-managed-encryption</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/analytics-hub-introduction">docs.cloud.google.com/bigquery/docs/analytics-hub-introduction</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/authorized-views">docs.cloud.google.com/bigquery/docs/authorized-views</a></p><p><strong>Q21.</strong></p><p><strong>At Elarvix Security, you want to encrypt customer data stored in BigQuery. You need to implement per-user crypto-deletion so that the encrypted data belonging to one user can be rendered unrecoverable without destroying the ability to decrypt other users’ data. You want to use native Google Cloud capabilities and avoid a custom cryptographic implementation. What should you do?</strong></p><p><strong>A. Use BigQuery Authenticated Encryption with Associated Data (AEAD) functions with separate keysets for individual users or appropriate user-level key domains.</strong></p><p><strong>B. Create one customer-managed encryption key in Cloud Key Management Service and associate that CMEK with the BigQuery table.</strong></p><p><strong>C. Create one customer-managed encryption key in Cloud Key Management Service and use that single key to encrypt all data before storing it in BigQuery.</strong></p><p><strong>D. Encrypt all user data during ingestion with a custom cryptographic library in the ETL pipeline.</strong></p><blockquote><strong><em>A is correct because BigQuery provides native GoogleSQL AEAD and deterministic-encryption functions together with keyset functions. Separate keysets can be generated or wrapped for different users, and the encrypted values can be associated with the corresponding user identifier. If a particular user’s keyset or required wrapping key is securely destroyed, the ciphertext protected only by that key can no longer be decrypted while other users’ independently keyed ciphertext remains recoverable. This gives the application a practical building block for fine-grained crypto-deletion without implementing encryption primitives itself.</em></strong></blockquote><p>B is incorrect because a table-level CMEK controls encryption of the BigQuery storage resource as a whole. Disabling or destroying the key affects all data protected by that table’s key rather than only the rows belonging to one user. CMEK is valuable for customer control of encryption at rest but does not by itself provide per-user cryptographic deletion inside one table.</p><p>C is incorrect because using one Cloud KMS key to protect every customer’s data has the same granularity problem. Destroying that one key makes all data encrypted under it unrecoverable, while retaining it leaves every user’s ciphertext decryptable to principals with appropriate access. Per-user crypto-deletion requires independent cryptographic key material or key domains.</p><p>D is incorrect because a custom cryptographic library can theoretically implement per-user keying, but the requirement explicitly prefers native Google Cloud functionality and wants to avoid a custom solution. Implementing encryption correctly also introduces key-format, rotation, storage, auditing, and library-maintenance responsibilities that BigQuery’s AEAD and keyset functions already address.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/reference/standard-sql/aead_encryption_functions">docs.cloud.google.com/bigquery/docs/reference/standard-sql/aead_encryption_functions</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/aead-encryption-concepts">docs.cloud.google.com/bigquery/docs/aead-encryption-concepts</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/customer-managed-encryption">docs.cloud.google.com/bigquery/docs/customer-managed-encryption</a></p><p><strong>Q22.</strong></p><p><strong>At Kelvorin Business Analytics, business users need a way to clean and prepare data before analyzing it. The users are not highly technical and prefer graphical interfaces for defining transformations. After the data is transformed, they want to analyze it directly in a spreadsheet. Which solution should you recommend?</strong></p><p><strong>A. Use BigQuery data preparation to clean and transform the data, write the prepared results to BigQuery, and analyze the data with Connected Sheets.</strong></p><p><strong>B. Use BigQuery data preparation to clean and transform the data, write the prepared results to BigQuery, and analyze the data with Looker Studio.</strong></p><p><strong>C. Use Dataflow to clean and transform the data, write the results to BigQuery, and analyze the data with Connected Sheets.</strong></p><p><strong>D. Use Dataflow to clean and transform the data, write the results to BigQuery, and analyze the data with Looker Studio.</strong></p><blockquote><strong><em>A is correct because BigQuery data preparation provides a visual interface for profiling data, receiving transformation suggestions, and applying cleansing and transformation steps without requiring users to hand-code an Apache Beam pipeline. Prepared output can be materialized in BigQuery. Connected Sheets then lets business users work with BigQuery data from the familiar Google Sheets interface, including pivot tables, formulas, charts, and scheduled refreshes. This combination matches both the graphical preparation requirement and the explicit desire to perform analysis in a spreadsheet.</em></strong></blockquote><p>B is incorrect because BigQuery data preparation is an appropriate graphical cleansing tool, but Looker Studio is a dashboarding and visualization product rather than a spreadsheet. It is useful when users want interactive reports and dashboards, but the requirement specifically says that the transformed data should be analyzed directly in a spreadsheet.</p><p>C is incorrect because Connected Sheets satisfies the spreadsheet requirement, but Dataflow is primarily a code-driven Apache Beam execution service. Nontechnical users would need a developer to implement or modify the transformation logic unless a prebuilt template happened to meet the use case. BigQuery data preparation provides a more suitable self-service graphical experience.</p><p>D is incorrect because neither part of the combination matches the user’s preferred interaction model as closely as option A. Dataflow requires substantially more technical pipeline-development skill, while Looker Studio is a reporting interface rather than a spreadsheet environment.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/data-prep-introduction">docs.cloud.google.com/bigquery/docs/data-prep-introduction</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/data-prep-get-suggestions">docs.cloud.google.com/bigquery/docs/data-prep-get-suggestions</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/connected-sheets">docs.cloud.google.com/bigquery/docs/connected-sheets</a></p><p><strong>Q23.</strong></p><p><strong>At Maventra Warehousing, you want to migrate an existing Teradata data warehouse to BigQuery. You want to move historical data with the most efficient managed method that requires the least custom programming, but free local disk space on the existing data-warehouse environment is limited. What should you do?</strong></p><p><strong>A. Use BigQuery Data Transfer Service with the Teradata migration agent configured for extraction through the JDBC driver with a FastExport connection.</strong></p><p><strong>B. Write a Teradata Parallel Transporter export script manually, export the historical data, and load it into BigQuery with the bq command-line tool.</strong></p><p><strong>C. Use BigQuery Data Transfer Service with the Teradata Parallel Transporter tbuild extraction method.</strong></p><p><strong>D. Write a custom export script, upload the historical data to Cloud Storage in batches, and configure a separate BigQuery Data Transfer Service transfer from Cloud Storage to BigQuery.</strong></p><blockquote><strong><em>A is correct because the BigQuery Teradata migration workflow uses BigQuery Data Transfer Service together with a migration agent to automate schema conversion and historical data movement. Current Google Cloud documentation specifically recommends JDBC extraction with a FastExport connection when local storage is constrained or when TPT cannot be used. In FastExport mode, the migration agent extracts data into Avro files, uploads those files to Cloud Storage, deletes the local files after upload, and allows a maximum local-storage limit to be configured. This directly addresses the limited disk space while minimizing custom migration code.</em></strong></blockquote><p>B is incorrect because manually writing Teradata Parallel Transporter scripts and orchestrating BigQuery loads gives the team more responsibility for extraction, staging, schema handling, retries, and job coordination. The managed migration agent and BigQuery Data Transfer Service already automate much of this workflow and are the lower-programming choice.</p><p>C is incorrect because TPT is a supported and often high-performance extraction method, but Google Cloud documentation states that the TPT method requires enough local free space for the largest table or table partition being extracted. The question explicitly says local storage is limited. FastExport lets the migration agent enforce a local-storage limit by uploading and deleting staged files as it proceeds.</p><p>D is incorrect because a custom script plus a second Cloud Storage-to-BigQuery transfer configuration creates more code and operational steps than necessary. The Teradata migration agent already stages extracted data and coordinates transfer into BigQuery. Reimplementing those functions manually conflicts with the requirement for the least programming effort.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/migration/teradata">docs.cloud.google.com/bigquery/docs/migration/teradata</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/migration/teradata-overview">docs.cloud.google.com/bigquery/docs/migration/teradata-overview</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/migration/teradata-migration-intro">docs.cloud.google.com/bigquery/docs/migration/teradata-migration-intro</a></p><p><strong>Q24.</strong></p><p><strong>At Brinorix Data, you are designing a Dataflow pipeline that reads from Cloud Storage and writes to BigQuery. The networking team has provided a Shared VPC network and subnetwork in a host project for the pipeline workers. You need to allow the Dataflow job to deploy and create worker interfaces in that Shared VPC subnetwork. What should you do?</strong></p><p><strong>A. Grant the Compute Network User role to the Dataflow service agent only.</strong></p><p><strong>B. Grant the Compute Network User role on the Shared VPC subnetwork to the service account that executes the Dataflow worker workload, together with the other required Shared VPC service-account grants.</strong></p><p><strong>C. Grant the Dataflow Admin role to the Dataflow service agent.</strong></p><p><strong>D. Grant the Dataflow Admin role to the service account that executes the Dataflow pipeline.</strong></p><p>A is incorrect because the Dataflow service agent and the Dataflow worker service account perform different functions. The service agent allows the Dataflow service to manage job resources, while the workers use the worker service account to create and use resources needed by the job. Current Shared VPC guidance requires Compute Network User access for the relevant service accounts on the host project’s shared network or subnetwork; granting it only to the Dataflow service agent does not provide all of the permissions required by the worker identity.</p><blockquote><strong><em>B is correct because a Dataflow job that uses a Shared VPC must be allowed to use the host project’s shared subnetwork. The worker service account needs the Compute Network User role on the specified Shared VPC subnetwork so that Dataflow workers can attach network interfaces there. Current guidance also requires the appropriate Shared VPC grants for the host project’s Compute Engine service account and the service account that submits the job, depending on the exact deployment arrangement. Among the listed choices, granting Compute Network User to the workload’s service account is the required action that directly enables worker use of the shared subnet.</em></strong></blockquote><p>C is incorrect because Dataflow Admin grants broad Dataflow control-plane permissions but does not grant permission to use a Shared VPC subnetwork. Network attachment is governed by Compute Engine networking permissions such as compute.subnetworks.use, which are included in Compute Network User rather than Dataflow Admin.</p><p>D is incorrect because Dataflow Admin controls creation and management of Dataflow resources, not permission to attach worker VMs to a subnetwork owned by another project. The workload identity still requires the relevant Compute Network User permission in the Shared VPC host project.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/dataflow/docs/guides/specifying-networks">docs.cloud.google.com/dataflow/docs/guides/specifying-networks</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/access-control">docs.cloud.google.com/dataflow/docs/concepts/access-control</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/concepts/security-and-permissions">docs.cloud.google.com/dataflow/docs/concepts/security-and-permissions</a></p><p><strong>Q25.</strong></p><p><strong>At Fynora Data, you are configuring networking for a Dataflow job. The pipeline uses custom container images with all required transformation libraries preinstalled, reads data from Cloud Storage, and writes data to BigQuery. You need secure and cost-effective communication between Dataflow workers and Google APIs and services, and the workers should not require external IP addresses. What should you do?</strong></p><p><strong>A. Disable external IP addresses on the worker VMs and enable Private Google Access on the Dataflow subnetwork.</strong></p><p><strong>B. Leave external IP addresses assigned to the worker VMs while relying on firewall rules to restrict traffic.</strong></p><p><strong>C. Disable external IP addresses and configure only a Private Service Connect endpoint IP address without enabling the required Google API access for the worker subnet.</strong></p><p><strong>D. Enable Cloud NAT to provide general outbound internet connectivity while enforcing firewall rules.</strong></p><blockquote><strong><em>A is correct because Dataflow supports workers that use internal IP addresses only. When external worker IPs are disabled, the selected VPC subnetwork must provide Private Google Access so those internal-only worker VMs can reach Google APIs and services such as Cloud Storage and BigQuery. This keeps the worker instances from requiring public IP addresses, avoids the cost associated with external worker IPs, and provides the private Google API connectivity required by the pipeline.</em></strong></blockquote><p>B is incorrect because keeping public worker IP addresses allows the workers to reach Google APIs, but it creates a broader network exposure and incurs the cost associated with external IP usage. Firewall rules are still important, but they do not provide the same private-only posture as disabling public worker addresses.</p><p>C is incorrect because Private Service Connect can provide private endpoints for supported Google APIs and services in appropriate architectures, but simply configuring one endpoint address does not automatically satisfy all of Dataflow’s service and worker connectivity requirements. The standard documented design for Dataflow workers with public IPs disabled is to enable Private Google Access on the subnetwork.</p><p>D is incorrect because Cloud NAT is useful when internal-only VMs must reach arbitrary public internet destinations. The pipeline’s required dependencies are already included in its custom container image, and the primary external destinations are Google APIs such as Cloud Storage and BigQuery. Private Google Access provides that access without giving the workers general internet egress through NAT, making it simpler and more tightly scoped.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/dataflow/docs/reference/pipeline-options">docs.cloud.google.com/dataflow/docs/reference/pipeline-options</a></p><p><a href="http://docs.cloud.google.com/dataflow/docs/guides/troubleshoot-networking">docs.cloud.google.com/dataflow/docs/guides/troubleshoot-networking</a></p><p><a href="http://docs.cloud.google.com/vpc/docs/private-google-access">docs.cloud.google.com/vpc/docs/private-google-access</a></p><p><strong>Q26.</strong></p><p><strong>At Gavorin Automation, you use Workflows to call an API that returns a 1 KB JSON response, apply complex business logic to that response, wait until the logic is complete, and then start a load from a Cloud Storage file into BigQuery. Workflows expressions and its standard library are not sufficient for the complex logic, and you want to implement that logic with Python’s standard library while optimizing for simplicity and fast execution. What should you do?</strong></p><p><strong>A. Create a Managed Service for Apache Airflow environment and run the Python logic as an Airflow task.</strong></p><p><strong>B. Create a Managed Service for Apache Spark cluster and use PySpark to apply the logic to the 1 KB JSON response.</strong></p><p><strong>C. Invoke a Cloud Run function from Workflows that uses Python to execute the complex logic and return the result before the workflow continues.</strong></p><p><strong>D. Invoke a Workflows subworkflow and implement the complex logic only with Workflows syntax.</strong></p><p>A is incorrect because Managed Service for Apache Airflow is a capable workflow orchestrator, but introducing an entire Airflow environment solely to process a 1 KB JSON response is unnecessarily heavy. The existing orchestration is already implemented in Workflows, so replacing or nesting it inside another orchestration platform increases cost and operational complexity.</p><p>B is incorrect because Managed Service for Apache Spark is designed for distributed Spark processing over substantially larger datasets. Starting a Spark environment to execute Python logic on a 1 KB JSON object would add high startup and runtime overhead relative to the tiny payload and would not simplify the existing workflow.</p><blockquote><strong><em>C is correct because Google Cloud explicitly recommends invoking a Cloud Run service or Cloud Run function when Workflows needs reusable or computationally complex logic that is easier to implement in code than in Workflows expressions. A lightweight Python Cloud Run function can receive the 1 KB JSON input, use Python’s standard library to perform the required business logic, and return the result synchronously. Workflows can make an authenticated HTTP call, wait for the response, and then proceed to the BigQuery load step. This keeps orchestration in Workflows while moving only the unsupported logic into a small managed serverless component.</em></strong></blockquote><p>D is incorrect because a subworkflow is still written with Workflows syntax and uses the same expression language and standard-library capabilities. Splitting the complex logic into another Workflows definition can improve organization but does not provide access to Python’s standard library, which is the explicit requirement.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/workflows/docs/calling-run-functions">docs.cloud.google.com/workflows/docs/calling-run-functions</a></p><p><a href="http://docs.cloud.google.com/workflows/docs/http-requests">docs.cloud.google.com/workflows/docs/http-requests</a></p><p><a href="http://docs.cloud.google.com/run/docs/functions/overview">docs.cloud.google.com/run/docs/functions/overview</a></p><p><strong>Q27.</strong></p><p><strong>At Doventra Knowledge Systems, your company wants to implement a Retrieval-Augmented Generation (RAG) system that allows employees to query an extensive knowledge base of internal documents such as policy manuals and project reports. You need to prepare the unstructured text for semantic retrieval in the RAG system. What should you do to help the system retrieve the most relevant information?</strong></p><p><strong>A. Use Sensitive Data Protection to scan and redact sensitive information within the documents before processing.</strong></p><p><strong>B. Store the documents as compressed files in a traditional relational database to enable more efficient storage and retrieval.</strong></p><p><strong>C. Convert the unstructured document text, typically in appropriately sized chunks, into high-dimensional numerical embedding vectors that capture semantic meaning and relationships.</strong></p><p><strong>D. Index each individual word from the documents in a search engine to support only keyword-based search.</strong></p><p>A is incorrect because Sensitive Data Protection is useful for discovering, classifying, masking, and de-identifying sensitive content before documents are made available to a RAG system. That can be an important governance step, but it does not create the semantic representation required for similarity-based retrieval. Redaction solves a privacy problem rather than the core information-retrieval problem described by the question.</p><p>B is incorrect because compression and relational storage affect how the source documents are stored, but they do not create a representation that lets the retrieval system identify passages that are semantically similar to a user’s question. A relational database can store document metadata, but semantic RAG retrieval relies on embedding representations or another semantic-search mechanism.</p><blockquote><strong><em>C is correct because text embeddings convert language into dense numerical vectors whose geometry represents semantic meaning. Semantically related passages are placed closer together in vector space even when they do not use exactly the same words. A typical RAG preparation process divides long documents into useful chunks, generates a retrieval-document embedding for each chunk, and stores those vectors in a vector index. At query time, the user question is embedded as a retrieval query and the system searches for nearby document vectors. This allows the generator to receive context that is relevant by meaning rather than only by exact keyword overlap.</em></strong></blockquote><p>D is incorrect because keyword indexes are valuable for lexical search, but indexing individual words alone does not capture the richer semantic relationships between phrases, sentences, and document passages. A user can ask a question using vocabulary different from the relevant source document. Dense embeddings are designed to improve retrieval in precisely those semantic-matching situations, and hybrid retrieval can combine embeddings with lexical search when both are useful.</p><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings">docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/get-text-embeddings">docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/get-text-embeddings</a></p><p><a href="http://docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/use-embedding-models">docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/use-embedding-models</a></p><p><strong>Q28.</strong></p><p><strong>At Kelvora Platforms, you need to orchestrate a simple sequential pipeline consisting of a batch Dataflow job, followed by a BigQuery query job, followed by a Gemini Enterprise Agent Platform batch inference job. You want a lightweight serverless orchestration service with minimal operational overhead. Which service should you use?</strong></p><p><strong>A. Managed Service for Apache Airflow</strong></p><p><strong>B. Compute Engine with cron</strong></p><p><strong>C. Managed Service for Apache Spark with Apache Oozie</strong></p><p><strong>D. Workflows</strong></p><p>A is incorrect because Managed Service for Apache Airflow can orchestrate all of these services and is a strong choice for complex DAGs, large data-platform estates, backfills, and workflows that benefit from the Apache Airflow ecosystem. However, the requirement describes a simple sequential chain and explicitly prioritizes a lightweight, serverless orchestration mechanism. A dedicated Airflow environment introduces more orchestration infrastructure than this workflow requires.</p><p>B is incorrect because a Compute Engine VM running cron requires the organization to provision, secure, patch, monitor, and keep the VM available. Custom scripts would also need to submit jobs, poll long-running operations, implement retries, and persist workflow state. This is substantially more operational work than using a managed orchestration platform.</p><p>C is incorrect because Managed Service for Apache Spark is a data-processing platform rather than a general-purpose lightweight orchestrator for Dataflow, BigQuery, and AI batch jobs. Apache Oozie is associated with Hadoop workflow orchestration and would add unnecessary cluster-oriented complexity for a pipeline whose services already expose managed APIs.</p><blockquote><strong><em>D is correct because Workflows is a fully managed serverless orchestration service designed to execute Google Cloud services and HTTP APIs in an explicit sequence. It can call Google APIs, retain workflow state, implement conditions and retries, and wait for long-running operations without requiring a persistent server or cluster. The Dataflow job can be started and monitored through its API, the BigQuery job through the BigQuery connector or API, and the Gemini Enterprise Agent Platform batch inference through its API. This is the lowest-overhead fit for a small sequential service-orchestration workflow.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/workflows/docs">docs.cloud.google.com/workflows/docs</a></p><p><a href="http://docs.cloud.google.com/workflows/docs/overview">docs.cloud.google.com/workflows/docs/overview</a></p><p><a href="http://docs.cloud.google.com/workflows/docs/reference/googleapis">docs.cloud.google.com/workflows/docs/reference/googleapis</a></p><p><strong>Q29.</strong></p><p><strong>At Pryntoria Compliance, your organization stores highly sensitive employee data in BigQuery. Privacy policy requires the encrypted values belonging to an individual employee to be crypto-shreddable so that, when that employee leaves the organization and the required retention conditions permit deletion, the employee’s protected values can be made permanently unreadable without making every other employee’s data unreadable. What should you do?</strong></p><p><strong>A. Use column-level access controls with policy tags and revoke viewer permissions when the employee leaves the organization.</strong></p><p><strong>B. Use dynamic data masking and revoke masked or raw-data permissions when the employee leaves the organization.</strong></p><p><strong>C. Use one customer-managed encryption key for the BigQuery table and destroy that key when an employee leaves the organization.</strong></p><p><strong>D. Encrypt sensitive values with BigQuery AEAD functions by using separately managed keysets or cryptographic key domains for individual employees, and securely destroy the applicable key material when an employee’s data must become unreadable.</strong></p><p>A is incorrect because column-level security controls who is authorized to query protected columns, but revoking one employee’s permissions does not cryptographically destroy the sensitive values stored for that employee. Other authorized principals can still access the data, and an administrator could later regrant access. Authorization revocation and cryptographic deletion are different controls.</p><p>B is incorrect because dynamic data masking changes what a principal sees when querying a protected column, but the underlying original value remains stored and can still be returned to principals with raw-data access. Removing an employee’s masking or query permission prevents that user from seeing the value but does not render the stored value permanently unrecoverable.</p><p>C is incorrect because a table-level customer-managed encryption key protects the entire BigQuery table or other configured resource. Destroying that single key would affect all data protected by it, including records for employees who remain active and whose data must remain usable. CMEK is valuable for customer-controlled encryption at rest, but one table key does not provide per-employee crypto-deletion granularity.</p><blockquote><strong><em>D is correct because BigQuery AEAD functions and keyset functions let an application encrypt individual values with application-defined cryptographic keysets rather than relying solely on one storage-level key for the entire table. If different employees or deletion domains use independently controlled key material, destroying the key required for one domain makes that ciphertext undecryptable while leaving data protected by other keysets usable. In a robust design, keysets should themselves be protected with an appropriate key-management hierarchy, and the organization must ensure that backups or historical key material cannot reintroduce decryptability after the mandated crypto-shredding event.</em></strong></blockquote><p><strong>Reference links:</strong></p><p><a href="http://docs.cloud.google.com/bigquery/docs/reference/standard-sql/aead_encryption_functions">docs.cloud.google.com/bigquery/docs/reference/standard-sql/aead_encryption_functions</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/aead-encryption-concepts">docs.cloud.google.com/bigquery/docs/aead-encryption-concepts</a></p><p><a href="http://docs.cloud.google.com/kms/docs/destroy-restore">docs.cloud.google.com/kms/docs/destroy-restore</a></p><p><strong>Q30.</strong></p><p><strong>At Joventra Customer Data, you are designing a managed data lake on Google Cloud for large volumes of customer-interaction data arriving from websites, mobile applications, and social-media sources in a variety of structured, semi-structured, and unstructured formats. Analysts need consistent metadata discovery, and the platform must support automated data-quality checks and managed transformations before downstream consumption. What should you do?</strong></p><p><strong>A. Use Cloud Storage as the central repository and use Gemini Enterprise Agent Platform alone to classify, transform, catalog, and validate every type of lake data.</strong></p><p><strong>B. Stream all data directly into native BigQuery tables and assume that all data will be cataloged, governed, and validated automatically without additional governance configuration.</strong></p><p><strong>C. Use Cloud Storage and BigQuery as the storage and analytical layers, use Knowledge Catalog for automated discovery, metadata governance, and data-quality scans, and use managed BigQuery data preparation or appropriate data-processing pipelines for the required transformations.</strong></p><p><strong>D. Use Cloud Storage as the central repository and write a custom Cloud Run function that catalogs every object, performs all transformations, and implements the data-quality checks.</strong></p><p>A is incorrect because Gemini Enterprise Agent Platform provides powerful AI and machine-learning capabilities but is not the general metadata-governance system for a heterogeneous enterprise data lake. Using an AI service alone does not create the required centralized catalog, automated storage discovery, lineage, governance metadata, and systematic quality framework across Cloud Storage and BigQuery.</p><p>B is incorrect because BigQuery is an excellent analytical store and automatically contributes technical metadata to Knowledge Catalog, but forcing every incoming data type directly into native BigQuery tables removes the flexibility of object storage for raw and unstructured content. In addition, merely loading data into BigQuery does not automatically define the organization’s data-quality rules, domain metadata, or required preprocessing logic.</p><blockquote><strong><em>C is correct because Cloud Storage can retain raw and unstructured lake data while BigQuery provides the analytical and structured data layer. Knowledge Catalog is Google Cloud’s current managed governance and discovery platform and can automatically discover supported Cloud Storage content, catalog BigQuery assets, capture metadata, and run managed data-quality and profiling scans. Transformation can then use managed BigQuery data preparation for SQL-oriented visual preparation or another appropriate managed processing pipeline while the catalog retains governance metadata about the resulting assets. This architecture supplies managed discovery, quality, governance, and transformation capabilities without requiring the organization to build its own catalog.</em></strong></blockquote><p>D is incorrect because a custom Cloud Run function would require the organization to implement file discovery, schema inference, catalog updates, validation rules, transformation logic, retries, state management, scaling, and reporting. That recreates a large portion of the capabilities offered by Knowledge Catalog and managed Google Cloud data-processing services and therefore creates unnecessary engineering and maintenance work.</p><p><strong>Reference links:</strong></p><p><a href="http://cloud.google.com/products/knowledge-catalog">cloud.google.com/products/knowledge-catalog</a></p><p><a href="http://docs.cloud.google.com/dataplex/cloud-storage-discovery">docs.cloud.google.com/dataplex/cloud-storage-discovery</a></p><p><a href="http://docs.cloud.google.com/bigquery/docs/data-prep-introduction">docs.cloud.google.com/bigquery/docs/data-prep-introduction</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c4bb3c234744" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>