<?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 The Agentic Builder on Medium]]></title>
        <description><![CDATA[Stories by The Agentic Builder on Medium]]></description>
        <link>https://medium.com/@TheContainerCraftsman?source=rss-e08a82d83e2c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*l9NcrQujo3Kc31oTY3183A@2x.jpeg</url>
            <title>Stories by The Agentic Builder on Medium</title>
            <link>https://medium.com/@TheContainerCraftsman?source=rss-e08a82d83e2c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 18:04:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@TheContainerCraftsman/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[Building a Cloud Operations Assistant with Amazon Bedrock AgentCore Runtime]]></title>
            <link>https://aws.plainenglish.io/building-a-cloud-operations-assistant-with-amazon-bedrock-agentcore-runtime-bad482ece25f?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/bad482ece25f</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[aws-bedrock]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Sat, 28 Mar 2026 04:44:21 GMT</pubDate>
            <atom:updated>2026-03-28T04:44:21.348Z</atom:updated>
            <content:encoded><![CDATA[<p>Cloud operations today isn’t limited by data — it’s limited by how fast we can <strong>connect signals, reason about them, and take action.</strong></p><p>We already have everything:</p><p>•	logs</p><p>•	metrics</p><p>•	alerts</p><p>•	deployment events</p><p>Yet when something breaks, engineers still:</p><p>•	jump across dashboards</p><p>•	correlate signals manually</p><p>•	run scripts or playbooks</p><p>We are moving from dashboards and scripts to systems that can investigate incidents, correlate signals, and take action.</p><p>A Cloud Operations Assistant is an AI agent that can observe system signals, reason across them, and trigger controlled actions.</p><p><strong>From Chatbots to Real Agents</strong></p><p>Most AI assistants today are still chatbots:</p><p>•	stateless</p><p>•	passive</p><p>•	limited to answering questions</p><p>A real agent has three capabilities:</p><p><strong>Reasoning</strong></p><p>Understands intent and breaks down problems</p><p><strong>Action</strong></p><p>Interacts with real systems</p><p><strong>Memory</strong></p><p>Maintains context across steps and over time</p><p>Without action and memory, you don’t have an agent — you have a chatbot.</p><p><strong>Architecture: Production-Ready Design</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*qK93U3lCs_GOrArStq9-Nw@2x.jpeg" /></figure><p><strong>Core Stack</strong></p><p>•	<strong>Agent Framework</strong>: Strands Agent Framework</p><p>•	<strong>Execution Runtime</strong>: Amazon Bedrock AgentCore Runtime</p><p>•	<strong>Memory Layer</strong>: Amazon Bedrock AgentCore Memory</p><p>•	<strong>Execution Layer</strong>: Lambda / internal APIs</p><p>•	<strong>Interface</strong>: Web UI / CLI</p><p><strong>How the System Works</strong></p><p>1.	User sends a request</p><p>2.	Backend attaches identity + context</p><p>3.	Strands agent: defines tools	and filters relevant tools</p><p>4.	AgentCore Runtime: performs reasoning, orchestrates tool calls and maintains session state</p><p>5.	AgentCore Memory: retrieves past context, stores new insights</p><p>6.	Tools fetch data or trigger actions</p><p>7.	Response returned to user</p><p>Under the hood, the runtime leverages foundation models through Amazon Bedrock for reasoning and decision-making.</p><p>We use a multi-model approach — primarily Claude Sonnet for complex reasoning and analysis, and lighter models for simpler operational queries. Based on the query type, the system dynamically selects the most appropriate model to balance accuracy, latency, and cost.</p><p><strong>Data Sources: The Intelligence Layer</strong></p><p>The assistant continuously pulls signals from across your cloud stack.</p><p><strong>Observability</strong></p><p>•	Amazon CloudWatch</p><p>•	AWS X-Ray</p><p>•	Splunk</p><p><strong>Deployment &amp; Change</strong></p><p>•	AWS CodePipeline</p><p>•	AWS CodeBuild</p><p><strong>Runtime &amp; Infrastructure</strong></p><p>•	Amazon EKS</p><p>•	Amazon ECS</p><p>•	AWS EC2</p><p><strong>Incident Systems</strong></p><p>•	PagerDuty</p><p>•	Opsgenie</p><p><strong>Configuration</strong></p><p>•	AWS Systems Manager</p><p><strong>Key Insight</strong></p><p><em>The real power of the agent comes from correlating signals across logs, metrics, deployments, and infrastructure — not analyzing them in isolation.</em></p><p><strong>Real Scenarios</strong></p><ol><li><strong>Failed Deployment Investigation</strong></li></ol><p><em>User: “Why did the X service deployment fail?”</em></p><p>Agent:</p><p>•	pulls deployment events</p><p>•	analyzes logs</p><p>•	checks health signals</p><p>•	correlates changes</p><p>Output:</p><p><em>“Failure caused by readiness check timeout after configuration change.”</em></p><p><strong>2. Alert Triage</strong></p><p><em>User:“Why is CPU spiking in ABC Environment?”</em></p><p>Agent:</p><p>•	checks CloudWatch metrics</p><p>•	evaluates recent deploys</p><p>•	compares with past incidents</p><p>Output:</p><p><em>“Similar spike previously caused by scaling misconfiguration.”</em></p><p><strong>3. Controlled Remediation</strong></p><p><em>User:“Restart the unhealthy service”</em></p><p>Agent:</p><p>•	verifies intent</p><p>•	checks policy</p><p>•	triggers approved workflow</p><p>The agent never executes raw commands — it invokes approved tools, ensuring security and auditability.</p><p><strong>Memory: The Intelligence Multiplier</strong></p><p><strong>Short-Term Memory</strong></p><p>Handled by Amazon Bedrock AgentCore Memory</p><p>•	session context</p><p>•	reasoning chain</p><p>•	tool outputs</p><p><strong>Long-Term Memory</strong></p><p>Also handled by AgentCore Memory:</p><p>•	past incidents</p><p>•	system patterns</p><p>•	user behavior</p><p><strong>Key Insight</strong></p><p><em>We separate memory into session-level reasoning and persistent intelligence.</em></p><p><strong>Why Strands Matters</strong></p><p>Strands Agent Framework</p><p>In real systems, you don’t have 5 tools — you have 100 plus tools.</p><p><strong>The Real Challenge</strong></p><p>The challenge is not executing actions — it’s selecting the right action.</p><p>Strands:</p><p>•	filters tools</p><p>•	reduces noise</p><p>•	improves accuracy</p><p><strong>Platform-Agnostic by Design</strong></p><p>The assistant can run anywhere:</p><p>•	containers</p><p>•	VMs</p><p>•	on-prem</p><p>It only needs:</p><p>•	access to runtime APIs</p><p>Execution happens through defined tools — not direct command execution.</p><p><strong>Observability of the Agent</strong></p><p>In production, you must monitor the agent itself:</p><p>•	tool calls</p><p>•	reasoning steps</p><p>•	latency</p><p>•	failures</p><p>You don’t just monitor your systems — you monitor the agent reasoning about your systems.</p><p><strong>Real Engineering Lessons</strong></p><p>•	Tool design matters more than model choice</p><p>•	Limit tool space</p><p>•	Observability is critical</p><p>•	Costs grow with multi-step workflows</p><p><strong>Final Thoughts</strong></p><p>As we scaled toward 100+ tools, we started facing a different kind of challenge. The problem was no longer whether the agent could call a tool, but whether it could consistently choose the right one. Similar tools began to overlap, context became noisy, and accuracy dropped unless tool descriptions, boundaries, and instructions were made much clearer.</p><p>We addressed this by reducing what the model sees at runtime. Instead of exposing the full tool catalog, we pre-filter the list and pass only the most relevant subset to the agent. That keeps reasoning focused, improves tool selection, and scales much better. At that stage, strong agent performance depends not just on the model, but on prompt design, tool design, and filtering strategy working together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bad482ece25f" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/building-a-cloud-operations-assistant-with-amazon-bedrock-agentcore-runtime-bad482ece25f">Building a Cloud Operations Assistant with Amazon Bedrock AgentCore Runtime</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why SKILL.md Matters More Than You Think in Agentic AI]]></title>
            <link>https://medium.com/@TheContainerCraftsman/why-skill-md-matters-more-than-you-think-in-agentic-ai-419b54a3b9b6?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/419b54a3b9b6</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Mon, 09 Mar 2026 19:49:00 GMT</pubDate>
            <atom:updated>2026-03-09T19:49:00.825Z</atom:updated>
            <content:encoded><![CDATA[<p>A practical lesson from building a Kubernetes assistant and realizing that models and tools were not enough</p><p>When people talk about building agentic AI systems, the conversation usually goes in familiar directions. Which model should we use? How should we design the system prompt? What tools should the agent be allowed to call?</p><p>Those questions matter. But while thinking through an internal Kubernetes assistant, I started noticing another layer that felt less visible, yet just as important.</p><p>The assistant could talk to the cluster. It could collect useful signals. It could send that live context to an LLM and return an answer to the user. On paper, that sounded like the core architecture was already there.</p><p>And in some sense, it was.</p><p>The system could inspect cluster state, read events, look at YAML, review service mappings, and help explain what might be happening. But the more I thought about it as a serious internal engineering tool rather than a demo, the more one gap became obvious.</p><p>The problem was not that the model lacked intelligence.</p><p>The problem was that the workflow for specific tasks was still too undefined.</p><p>Sometimes the assistant would produce a sharp, structured answer. Other times it would be broad. Sometimes it would clearly prioritize the strongest evidence. Other times it would mention several technically possible causes without enough discipline or ranking. The output was often useful, but not always consistent in the way an internal engineering product needs to be.</p><p>That is what made the idea of SKILL.md click for me.</p><p><strong>The gap I kept running into</strong></p><p>A lot of LLM applications follow a similar pattern:</p><p>•	the user asks a question</p><p>•	the application gathers context</p><p>•	tools fetch additional data</p><p>•	the model analyzes the input</p><p>•	the answer is returned to the user</p><p>That is a strong starting point. In fact, it is the backbone of many good AI applications.</p><p>But while thinking through the design of a Kubernetes assistant, I realized something important: giving the model access to data is not the same as giving it a method.</p><p>That distinction matters more than it first appears.</p><p>If the assistant is analyzing Kubernetes events, logs, YAML, service mappings, resource limits, or Helm values, it still needs a repeatable reasoning path for each class of task. Otherwise, the model is effectively deciding its own procedure every time.</p><p>And that is exactly where inconsistency shows up.</p><p>The model may be capable. The tools may be reliable. But if the task execution path is not well defined, the user experience will still feel uneven.</p><p>That was the hidden gap.</p><p><strong>What SKILL.md means in practice</strong></p><p>The way I think about SKILL.md is simple.</p><p>The model provides reasoning.</p><p>The tools provide access.</p><p>The system prompt provides global behavior.</p><p>But the skill file defines how one specific kind of task should be executed.</p><p>That is what makes it so useful.</p><p>To me, SKILL.md is not just another prompt file. It is a task-specific playbook. It gives the model a repeatable way to handle a category of work. It says, in effect: when this kind of problem appears, here is what matters, here is the order to think in, here is what you should not assume, and here is how the answer should be structured.</p><p>That is very different from a general system prompt.</p><p>A system prompt might tell the assistant to be accurate, grounded, safe, and clear. All of that is necessary. But it is still broad. It does not tell the model how to execute a pod scheduling diagnosis differently from a manifest review or a Helm values check.</p><p>A skill file does.</p><p>That difference becomes much more important in agentic systems than in simple chat experiences.</p><p><strong>Why this became obvious in a Kubernetes assistant</strong></p><p>Kubernetes is actually a very good environment for seeing why this matters.</p><p>If an assistant is connected to the cluster, it can gather many kinds of live signals: pod status, recent events, logs, service mappings, PVC state, Helm values, node summaries, and more. That gives it access to the truth of what is happening.</p><p>But Kubernetes is not one single reasoning problem. It is a collection of recurring problem types, and each one benefits from its own diagnostic flow.</p><p>A pod stuck in Pending needs one reasoning path.</p><p>A CrashLoopBackOff needs another.</p><p>A service connectivity issue needs another.</p><p>An Ingress problem needs another.</p><p>A manifest review needs another.</p><p>A Helm values review needs another.</p><p>A human platform engineer does not approach all of these with the same mental checklist. The workflow changes depending on the problem.</p><p>That was the key realization for me.</p><p>If the application already knows how to fetch the right context from the cluster, then the next challenge is not simply giving that context to the LLM. The next challenge is telling the LLM how to reason for that specific task.</p><p>Without that layer, the model is still improvising too much.</p><p><strong>Intelligence is not the same as procedure</strong></p><p>This was probably the biggest lesson.</p><p>In many AI conversations, we focus heavily on whether the model is powerful enough. That is an important question, but it is only part of the story. In an internal engineering assistant, what matters just as much is whether the model follows a disciplined procedure.</p><p>Take a simple example: a pod stuck in Pending.</p><p>A general model may list several possible causes: insufficient CPU, insufficient memory, taints, tolerations, affinity rules, PVC issues, maybe even image problems or probe problems. Some of those are valid Kubernetes issues in general, but they are not equally relevant in the moment.</p><p>A better reasoning path would start with scheduler events. If the pod has not even been scheduled yet, image pull failures and probe issues should not be the focus. Storage binding or placement constraints might matter, but only if the evidence points there.</p><p>That is not just a better answer. It is a better procedure.</p><p>And that is exactly the sort of logic a skill file can define.</p><p>Once I started looking at the problem through that lens, it became clear that one of the missing pieces in many agentic systems is not more intelligence. It is better-scoped execution logic.</p><p><strong>Why this is different from just expanding the system prompt</strong></p><p>One easy temptation in agent design is to keep adding more instructions into one larger and larger master prompt.</p><p>That works for a while. Then it starts to become messy.</p><p>As more capabilities are added, the prompt becomes crowded with overlapping rules, workflow-specific exceptions, and too many special cases. It becomes harder to maintain and harder to improve one capability without accidentally affecting another.</p><p>A skill-based approach feels much cleaner.</p><p>You can keep:</p><p>•	a global layer for overall assistant behavior</p><p>•	a tool layer for fetching data or taking actions</p><p>•	a skill layer for task-specific execution</p><p>That creates a healthier architecture.</p><p>If the manifest review flow needs improvement, I can refine that skill. If the pod troubleshooting flow needs better evidence prioritization, I can refine that skill. I do not need to rewrite the entire behavioral contract of the system.</p><p>From an engineering standpoint, that is a much more scalable way to build agentic applications.</p><p><strong>It is not only about troubleshooting</strong></p><p>Another reason I like this concept is that it is not limited to incident analysis.</p><p>Within a Kubernetes assistant, a skill layer can also help with app development workflows. The same internal assistant might support developers with:</p><p>•	reviewing manifests</p><p>•	validating probes</p><p>•	understanding Helm values</p><p>•	checking service-to-service connectivity</p><p>•	translating application requirements into Kubernetes resources</p><p>•	identifying risky configuration patterns</p><p>Those are very different tasks. They should not all be handled with the same reasoning style.</p><p>A manifest review skill should think like a configuration checker.</p><p>A runtime issue skill should think like an incident investigator.</p><p>An onboarding skill should think like a platform guide.</p><p>That is exactly why the SKILL.md idea feels so practical. It allows one assistant to support multiple high-value workflows without collapsing them into one vague analysis mode.</p><p><strong>Why this improves trust</strong></p><p>For internal engineering tools, trust matters more than polish.</p><p>The assistant does not just need to sound fluent. It needs to make it clear what was observed, what was inferred, and what still needs validation.</p><p>A strong skill file can enforce that structure. It can require the output to separate:</p><p>•	observed facts</p><p>•	likely interpretation</p><p>•	missing information</p><p>•	recommended next checks</p><p>That kind of clarity makes a huge difference.</p><p>When an assistant blends everything into one smooth paragraph, it can sound convincing without being easy to trust. But when it clearly says, in effect, this is what I observed, this is how I interpret it, and this is what still needs verification, the output becomes much more usable.</p><p>That kind of structure is important in Kubernetes, but it also applies to almost every serious enterprise AI workflow.</p><p><strong>The broader lesson for agentic AI</strong></p><p>What started as internal thinking around a Kubernetes assistant ended up feeling like a broader design lesson.</p><p>A lot of discussion in agentic AI focuses on prompts, tools, and models. Those are all important. But there is another layer that deserves much more attention: the task-specific execution layer.</p><p>That is what SKILL.md represents to me.</p><p>Not just a file. A design principle.</p><p>It is the recognition that real AI systems need more than intelligence and access. They also need reusable procedures for how specific kinds of work should be done.</p><p>That is what improves consistency. That is what makes outputs easier to trust. And that is what starts to separate a working demo from a genuinely useful internal product.</p><p><strong>Final thought</strong></p><p>While thinking through how to build a Kubernetes assistant inside an engineering organization, I came to a simple conclusion: the model was not the whole story, and the tools were not the whole story either.</p><p>What was missing was a structured way to tell the model how to perform a specific task well.</p><p>That is why SKILL.md matters more than it may first appear.</p><p>System prompts define general behavior.</p><p>Tools provide capability.</p><p>Skill files help turn that capability into a repeatable workflow.</p><p>And in real-world agentic systems, that may be one of the most important layers of all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=419b54a3b9b6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10 Kubernetes Pitfalls You Must Avoid in Production (with EKS-Specific Fixes)]]></title>
            <link>https://aws.plainenglish.io/10-kubernetes-pitfalls-you-must-avoid-in-production-with-eks-specific-fixes-bdd7a4a17d90?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/bdd7a4a17d90</guid>
            <category><![CDATA[aws-eks]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[aws]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Mon, 09 Mar 2026 12:36:13 GMT</pubDate>
            <atom:updated>2026-03-09T12:36:13.312Z</atom:updated>
            <content:encoded><![CDATA[<p>Kubernetes gives teams a powerful way to run workloads at scale, but in production, even small oversights can snowball into outages, downtime, or spiraling cloud costs. Based on lessons learned running Amazon EKS in production, here are the top 10 Kubernetes pitfalls — and how to avoid them with EKS-ready solutions.</p><ol><li><strong>Running Pods Without CPU/Memory Requests and Limits</strong></li></ol><p>•	<strong>Pitfall</strong>: No requests = noisy neighbors. No limits = runaway pods. Autoscalers like HPA or Karpenter can’t make good decisions.</p><p>•	<strong>Fix</strong>: Define requests/limits in manifests. Enforce with Kyverno or OPA Gatekeeper.</p><pre>resources:<br>  requests:<br>    cpu: &quot;200m&quot;<br>    memory: &quot;256Mi&quot;<br>  limits:<br>    cpu: &quot;500m&quot;<br>    memory: &quot;512Mi&quot;</pre><p>•	<strong>EKS Tip</strong>: Combine with Vertical Pod Autoscaler (VPA) for right-sizing and let Karpenter scale nodes dynamically.</p><p><strong>2. Skipping Readiness and Liveness Probes</strong></p><p>•	<strong>Pitfall</strong>: Rolling updates route traffic to unready pods. Stuck containers live forever.</p><p>•	<strong>Fix</strong>: Always set readinessProbe. Use livenessProbe only for deadlock detection.</p><p>•	<strong>EKS Tip</strong>: When paired with ALB Ingress Controller, readiness probes ensure traffic only hits healthy pods.</p><p><strong>3. No PodDisruptionBudgets (PDBs)</strong></p><p>•	<strong>Pitfall</strong>: A single node drain can evict all pods of a Deployment.</p><p>•	<strong>Fix</strong>: Define PDBs to guarantee minimum availability.</p><pre>maxUnavailable: 1</pre><p>•	<strong>EKS Tip</strong>: Use PDBs alongside Managed Node Group upgrades or Fargate evictions to keep apps online.</p><p><strong>4. Running in a Single AZ or Node Pool</strong></p><p>•	<strong>Pitfall</strong>: One AZ outage = full downtime.</p><p>•	<strong>Fix</strong>: Spread workloads with multi-AZ Managed Node Groups. Add topologySpreadConstraints for fine-grained balancing.</p><p>•	<strong>EKS Tip</strong>: With Karpenter, define multiple NodePools across AZs to guarantee resilience.</p><p><strong>5. Overlooking NetworkPolicies</strong></p><p>•	<strong>Pitfall</strong>: By default, every pod can talk to every other pod. Flat networks are a huge security risk.</p><p>•	<strong>Fix</strong>: Apply a namespace default-deny policy and explicitly allow required communications.</p><p>•	<strong>EKS Tip</strong>: Use Calico or Cilium as the CNI plugin to enforce NetworkPolicies at scale. Pair with Wiz Security for runtime drift detection.</p><p><strong>6. Using the latest Tag in Production</strong></p><p>•	<strong>Pitfall</strong>: :latest pulls a different image tomorrow → non-reproducible deploys.</p><p>•	<strong>Fix</strong>: Pin to digests (image@sha256:…) and sign them with Cosign.</p><p>•	<strong>EKS Tip</strong>: Integrate ECR image scanning to block vulnerable builds before deploy.</p><p><strong>7. Hardcoding Secrets in Plain Text</strong></p><p>•	<strong>Pitfall</strong>: Secrets in env vars or Git leak into logs and dumps.</p><p>•	<strong>Fix</strong>: Store in AWS Secrets Manager or Conjur. Use External Secrets Operator (ESO) to sync securely into pods.</p><p>•	<strong>EKS Tip</strong>: Always mount secrets as files, not environment variables.</p><p><strong>8. Ignoring Graceful Shutdowns</strong></p><p>•	<strong>Pitfall</strong>: Pods killed without drain = 502s, dropped connections.</p><p>•	<strong>Fix</strong>: Handle SIGTERM, set terminationGracePeriodSeconds, and use preStop hooks.</p><p>•	<strong>EKS Tip</strong>: Match app drain time with ALB deregistration delay (default 300s) for smooth rolling updates.</p><p><strong>9. Forgetting Logging and Cost Controls</strong></p><p>•	<strong>Pitfall</strong>: Chatty apps → unbounded CloudWatch bills and node disk exhaustion.</p><p>•	<strong>Fix</strong>: Rotate logs, set retention policies, enable sampling.</p><p>•	<strong>EKS Tip</strong>: Use Fluent Bit DaemonSets + CloudWatch Logs with filters. For cost control, monitor with CloudZero or AWS Cost Explorer.</p><p><strong>10. Treating Persistent Volumes as Disposable</strong></p><p>•	<strong>Pitfall</strong>: PVC deletion without backup = instant data loss.</p><p>•	<strong>Fix</strong>: Set correct reclaimPolicy, enable EBS snapshots, and define backup/restore runbooks.</p><p>•	<strong>EKS Tip</strong>: Use EBS CSI driver with snapshots for stateful workloads. For shared storage, enable Amazon EFS CSI driver with backups.</p><p><strong>Final Thoughts</strong></p><p>Running Kubernetes in production is less about YAML and more about discipline:</p><p>•	Automate guardrails with Kyverno.</p><p>•	Scale smart with Karpenter + HPA/VPA.</p><p>•	Secure workloads with IRSA, Secrets Manager, and NetworkPolicies.</p><p>•	Spread across multi-AZ node groups for resilience.</p><p>•	Control observability and cost with the right logging strategy.</p><p>Avoiding these 10 pitfalls will keep your EKS workloads stable, secure, and cost-efficient — so your team can focus on delivering value instead of firefighting.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bdd7a4a17d90" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/10-kubernetes-pitfalls-you-must-avoid-in-production-with-eks-specific-fixes-bdd7a4a17d90">10 Kubernetes Pitfalls You Must Avoid in Production (with EKS-Specific Fixes)</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My CKAD Exam Experience (After CKA 2021)]]></title>
            <link>https://medium.com/codetodeploy/my-ckad-exam-experience-after-cka-2021-97a0a75f70e4?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/97a0a75f70e4</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[container-orchestration]]></category>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[aws-eks]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Mon, 19 Jan 2026 03:57:47 GMT</pubDate>
            <atom:updated>2026-01-19T03:57:47.129Z</atom:updated>
            <content:encoded><![CDATA[<h4>What Changed, What Didn’t, and What Truly Matters</h4><p><a href="https://medium.com/codetodeploy">CodeToDeploy</a></p><p>I cleared the <strong>Certified Kubernetes Application Developer (CKAD)</strong> exam last weekend with a score of 89. I had previously taken CKA in 2021, and I’ve been working with Kubernetes for around four years in real-world environments.</p><p>Most of the concepts in the CKAD exam were familiar to me. What surprised me was <strong>how much the exam now emphasizes speed, efficiency, and discipline</strong>, even for people with solid Kubernetes experience.</p><p>This is not a syllabus recap. This is what genuinely matters during the exam.</p><p><strong>What Has Changed Since CKA 2021</strong></p><p>The biggest change is not the complexity of Kubernetes concepts, but the emphasis on speed and execution.</p><p>•	Questions are short and very task-oriented</p><p>•	Time pressure is real</p><p>•	Over-engineering costs marks</p><p>•	Efficiency matters more than elegance</p><p>Even with strong Kubernetes experience, <strong>time management becomes the hardest part.</strong></p><p><strong>The Most Important Lesson: Just Solve the Task</strong></p><p>This is the single biggest takeaway from my exam:</p><p><strong><em>Do exactly what the question asks. Don’t try to be perfect.</em></strong></p><p>If the task says:</p><p>•	Create a Service → don’t add extra labels</p><p>•	Mount a Secret → don’t redesign the Pod</p><p>•	Add a sidecar → don’t refactor the application</p><p>CKAD does not reward best-practice perfection.</p><p><strong><em>It rewards correct results delivered quickly.</em></strong></p><h4>🚀 Top Remote Tech Jobs — $50–$120/hr</h4><p>🔥 <em>Multiple Roles Open</em><br> Hiring E<strong>xperienced Talent (3+ years)</strong> Only.</p><ul><li>Frontend / Backend / Full Stack</li><li>Mobile (iOS/Android)</li><li>AI / ML</li><li>DevOps &amp; Cloud</li></ul><p>⏳ <strong>Opportunities Fill FAST — Early Applicants Get Priority! </strong><br> 👉 <a href="https://app.usebraintrust.com/r/code6/"><strong>Apply Here</strong></a></p><p><strong>kubectl Is Your Speed Tool</strong></p><p>In CKAD, kubectl is not optional — it’s your biggest advantage.</p><p>Use it to <strong>generate YAML quickly</strong>, then edit only what’s required.</p><pre>kubectl create deployment nginx \<br>  --image=nginx \<br>  --dry-run=client -o yaml &gt; deploy.yaml</pre><p>This approach:</p><p>•	saves time</p><p>•	avoids syntax errors</p><p>•	reduces mental load under pressure</p><p><strong>Commands That Save the Most Time</strong></p><p>Get comfortable with:</p><p><em>•	kubectl run</em></p><p><em>•	kubectl create</em></p><p><em>•	kubectl expose</em></p><p><em>•	kubectl set</em></p><p>They help you move fast without writing YAML from scratch.</p><p><strong>When Time Is Tight: Use kubectl edit</strong></p><p>If a resource already exists, don’t recreate it.</p><p><em>Use kubectl edit for Deployments, Services, and any resource that supports it.</em></p><p>This is often the fastest way to:</p><p>•	fix ports</p><p>•	update env vars</p><p>•	add volumes or mounts</p><p>•	tweak selectors</p><p>•	add sidecar or init containers</p><p><strong>Double-Check Context and Namespace (Non-Negotiable)</strong></p><p>This sounds basic, but it’s one of the easiest ways to lose time.</p><p>Before starting every task, always check:</p><pre>kubectl config current-context or ssh node-name<br>kubectl get ns</pre><p>Many questions use non-default namespaces.</p><p>Forgetting this can make you think something is broken when it isn’t.</p><p><strong>Troubleshooting: Your Best Friends</strong></p><p>When an application doesn’t work, don’t guess.</p><p>Your best friends are:</p><p><em>•	kubectl describe</em></p><p><em>•	kubectl logs</em></p><pre>kubectl describe pod &lt;pod&gt;<br>kubectl logs &lt;pod&gt; -c &lt;container&gt;<br></pre><p>Events and logs usually tell you exactly what’s wrong — image pull errors, probe failures, config issues, permission problems. This saves minutes per task.</p><p><strong>Topics You Must Be Very Comfortable With</strong></p><p><strong>NodePort Services (Very Important)</strong></p><p>You must clearly understand:</p><p>•	how NodePort exposes applications</p><p><em>•	port vs targetPort vs nodePort</em></p><p>•	traffic flow from outside the cluster to the Pod</p><p>This is a common place people lose time under pressure.</p><p><strong>ConfigMaps and Secrets with Containers</strong></p><p>Expect practical tasks involving:</p><p>•	environment variables</p><p>•	volume mounts</p><p>•	specific keys</p><p>•	mixing ConfigMaps and Secrets</p><p>Be comfortable with:</p><p>•	env vs envFrom</p><p>•	mounting as files vs environment variables</p><p><strong>Sidecars and Init Containers (Logging Use Case)</strong></p><p>One important insight:</p><p><em>Sidecars used for logging often belong in initContainers.</em></p><p>Why:</p><p>•	directories may need to exist first</p><p>•	permissions may need to be set</p><p>•	one-time setup should happen before the main container starts</p><p>Understanding when and why matters more than syntax.</p><p><strong>Container Behavior Matters</strong></p><p>You should understand:</p><p>•	container lifecycle</p><p>•	shared volumes between containers</p><p>•	startup order</p><p>•	restart behavior</p><p>CKAD tests how you think as an application developer, not a cluster admin.</p><p><strong>Mock Tests Are Non-Negotiable</strong></p><p>Mock exams made a big difference for me.</p><p>I strongly recommend:</p><p>•	<strong>KodeKloud CKAD mock tests</strong> to get comfortable with task patterns</p><p>•	<strong>At least a couple of killer.sh CKAD mocks</strong> for timing accuracy and confidence</p><p>Killer.sh is harder than the real exam — and that’s exactly why it helps.</p><p><strong>Final Advice</strong></p><p>If you already work with Kubernetes:</p><p>•	you don’t need more theory</p><p>•	you need <strong>speed, discipline, and accuracy</strong></p><p>Focus on:</p><p>•	kubectl — dry-run=client -o yaml</p><p>•	fast YAML editing</p><p>•	kubectl run, kubectl expose, kubectl set</p><p>•	kubectl edit when time is limited</p><p>Deliver the <strong>minimum correct solution</strong>, then move on.</p><p><strong>Final Thoughts</strong></p><p>CKAD is a fair exam — but it punishes overthinking.</p><p>I scored 89 by:</p><p>•	solving exactly what was asked</p><p>•	staying efficient</p><p>•	trusting hands-on experience</p><p>If you already use Kubernetes in real projects, train for <strong>execution</strong>, not memorization.</p><p>Good luck — and stay calm during the exam.</p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d9QTaaaxboQP_gKSLedW_w.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://www.linkedin.com/in/code-to-deploy-3784391b9/"><strong>Linkedin</strong></a>| <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p><strong>Note:</strong> This Post may contain affiliate links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97a0a75f70e4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/my-ckad-exam-experience-after-cka-2021-97a0a75f70e4">My CKAD Exam Experience (After CKA 2021)</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Kubernetes v1.35: A Quiet Release That Fundamentally Improves Production Clusters]]></title>
            <link>https://aws.plainenglish.io/kubernetes-v1-35-a-quiet-release-that-fundamentally-improves-production-clusters-7a7dc4042ccd?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a7dc4042ccd</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[site-reliability-engineer]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[cloud-native]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Sat, 03 Jan 2026 15:28:51 GMT</pubDate>
            <atom:updated>2026-01-12T07:17:26.340Z</atom:updated>
            <content:encoded><![CDATA[<p>Released in December 2025, Kubernetes v1.35 focuses on production stability, efficiency, and operational maturity rather than flashy new abstractions.</p><p>Not every Kubernetes release is about new APIs or shiny features.</p><p><strong>Kubernetes v1.35 </strong>is one of those releases.</p><p><strong>In-Place Pod Resource Resize Goes GA</strong></p><p>For a long time, changing CPU or memory in Kubernetes meant one thing: <strong>the Pod had to restart.</strong></p><p>That single limitation forced teams to:</p><p>•	Over-provision resources</p><p>•	Accept downtime</p><p>•	Avoid tuning workloads in production</p><p>With v1.35,<strong> in-place Pod resource resize is now GA.</strong></p><p><strong>What changed</strong></p><p>•	CPU and memory can be updated on running Pods</p><p>•	No restarts, no evictions</p><p>•	The kubelet applies resource changes live</p><p><strong>Why it matters</strong></p><p>•	Stateful workloads stay online</p><p>•	Long-running services no longer fear tuning</p><p>•	Over-provisioning becomes optional, not mandatory</p><p>This is a fundamental shift. Kubernetes finally behaves more like a real operating system for applications.</p><p><strong>Vertical Pod Autoscaler Finally Becomes Practical</strong></p><p>Vertical Pod Autoscaler (VPA) has existed for years, but it always came with a tradeoff: it relied on <strong>Pod restarts.</strong></p><p>In v1.35:</p><p>•	VPA works seamlessly with in-place resizing</p><p>•	No forced Pod evictions</p><p>•	Resource recommendations are more stable</p><p>For the first time, teams can safely use:</p><p>•	<strong>HPA</strong> to scale replicas</p><p>•	<strong>VPA</strong> to tune CPU and memory</p><p>Together — without chaos.</p><p>This is a major win for <strong>cost optimization and performance stability.</strong></p><p><strong>Scheduler and Control Plane Efficiency Improvements</strong></p><p>v1.35 also delivers meaningful improvements under the hood — the kind you don’t see in YAML, but absolutely feel in production.</p><p>Key benefits include:</p><p>•	Faster Pod scheduling during burst workloads</p><p>•	Lower scheduler CPU usage</p><p>•	Reduced control-plane pressure in large clusters</p><p>If you operate job-heavy platforms, CI/CD clusters, or large multi-node environments, this release quietly improves reliability.</p><p><strong>Security Defaults Continue to Tighten</strong></p><p>Kubernetes continues its steady move toward <strong>secure-by-default</strong> behavior.</p><p>In v1.35:</p><p>•	Admission validation is stricter</p><p>•	Unsafe defaults are further reduced</p><p>•	Pod Security Standards are enforced more consistently</p><p>Some older manifests or Helm charts may break — but that’s intentional.</p><p>The platform is actively removing foot-guns instead of preserving risky defaults.</p><p><strong>API Cleanup and Maturity</strong></p><p>This release continues Kubernetes’ long-term cleanup strategy:</p><p>•	Beta APIs graduating to GA</p><p>•	Deprecated APIs removed</p><p>•	Clearer validation errors during apply time</p><p>It’s not exciting — and that’s exactly why it’s good.</p><p>Cleaner APIs today mean fewer painful upgrades tomorrow.</p><p><strong>Better Autoscaling Signals and Observability</strong></p><p>Autoscaling is only as good as the metrics behind it.</p><p>v1.35 improves:</p><p>•	Resource metric consistency</p><p>•	Kubelet-to-metrics pipeline stability</p><p>•	Autoscaler decision accuracy</p><p>The result:</p><p>•	Less scaling oscillation</p><p>•	More predictable behavior</p><p>•	Better cost control</p><p><strong>What Kubernetes v1.35 Is Not</strong></p><p>Let’s set expectations clearly.</p><p>•	Not a YAML syntax overhaul</p><p>•	Not a CRD revolution</p><p>•	Not an urgent upgrade for every cluster</p><p>This is a <strong>maturity release</strong>, not a marketing release.</p><p><strong>Kubernetes v1.35 — What’s Added vs What’s Removed (At a Glance)</strong></p><p><strong>✅ What’s Added / Improved in v1.35</strong></p><p>•	In-place Pod resource resize (GA)</p><p>•	Vertical Pod Autoscaler becomes production-ready</p><p>•	Better scheduler performance</p><p>•	Improved control-plane efficiency</p><p>•	More reliable autoscaling(HPA/VPA) signals</p><p>•	Stronger security defaults</p><p>•	API maturity and stability</p><p><strong>❌ What’s Removed / Disabled / Cleaned Up in v1.35</strong></p><p>•	cgroup v1 support effectively removed</p><p>•	Obsolete feature gates removed</p><p>•	Deprecated APIs fully removed</p><p>•	Legacy behaviors tightened</p><p>•	Networking legacy paths moving out</p><p><strong>Final Thoughts</strong></p><p><strong>Kubernetes v1.35 doesn’t try to impress — it tries to help.</strong></p><p>It reduces downtime, improves efficiency, and makes autoscaling usable in real production environments.</p><p><strong>And before you go, don’t forget to clap and follow the writer️! ❤️</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7a7dc4042ccd" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/kubernetes-v1-35-a-quiet-release-that-fundamentally-improves-production-clusters-7a7dc4042ccd">Kubernetes v1.35: A Quiet Release That Fundamentally Improves Production Clusters</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Load Balancing in Amazon EKS: ALB, NLB, and Gateway API Explained]]></title>
            <link>https://aws.plainenglish.io/load-balancing-in-amazon-eks-alb-nlb-and-gateway-api-explained-bcf7a795a3f2?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcf7a795a3f2</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[kubernetes-networking]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[load-balancing]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Wed, 24 Sep 2025 20:46:50 GMT</pubDate>
            <atom:updated>2025-10-14T05:45:07.460Z</atom:updated>
            <content:encoded><![CDATA[<p>Kubernetes on Amazon EKS gives us the flexibility to run containerized workloads at scale — but workloads need a reliable way to receive and route traffic. That’s where load balancers come in.</p><p>On AWS, EKS integrates natively with three load balancing options:</p><p>•	Network Load Balancer (NLB) — high-performance Layer 4</p><p>•	Application Load Balancer (ALB) — feature-rich Layer 7</p><p>•	Gateway API with VPC Lattice — modern service networking</p><p>But which one should you choose, and when? In this blog, we’ll explore each option with use cases, YAML examples, setup steps, and cost considerations.</p><p><strong>Why Load Balancing Matters in EKS</strong></p><p>EKS workloads need to handle both:</p><p>•	North — South traffic → traffic entering or leaving the cluster (e.g., users accessing your app).</p><p>•	East — West traffic → service-to-service communication inside or across clusters.</p><p>AWS provides controllers that let Kubernetes resources (Service, Ingress, Gateway) directly provision cloud-native load balancers.</p><p><strong>Option 1: Network Load Balancer (NLB)</strong></p><p>When to use:</p><p>•	gRPC services (fintech, ML inference)</p><p>•	Gaming/chat apps with long-lived TCP connections</p><p>•	Ultra-high throughput TCP/UDP traffic</p><p>•	Services needing static IPs</p><p>Pros:</p><p>✅ Extremely fast, L4 only</p><p>✅ Handles millions of requests/sec</p><p>✅ Cheaper at scale than ALB</p><p>Cons:</p><p>❌ No HTTP routing</p><p>❌ No WAF or auth integration</p><p>Example: gRPC Service via NLB</p><pre>apiVersion: v1<br>kind: Service<br>metadata:<br>  name: grpc-service<br>  annotations:<br>    service.beta.kubernetes.io/aws-load-balancer-type: external   # NLB<br>    service.beta.kubernetes.io/aws-load-balancer-nlb-target-type: ip<br>spec:<br>  type: LoadBalancer<br>  ports:<br>    - name: grpc<br>      port: 443<br>      targetPort: 50051<br>      protocol: TCP<br>  selector:<br>    app: grpc-app</pre><p>Use case: A payment service exposing gRPC to external clients needing TLS passthrough.</p><p><strong>Option 2: Application Load Balancer (ALB)</strong></p><p>When to use:</p><p>•	Public websites &amp; APIs</p><p>•	Host/path-based routing (/shop, /checkout)</p><p>•	SSL termination, OIDC authentication, WAF</p><p>Pros:</p><p>✅ Full HTTP/S features</p><p>✅ Works natively with Ingress</p><p>✅ Supports WAF, Cognito, SSL offloading</p><p>Cons:</p><p>❌ Slightly pricier than NLB per request</p><p>❌ Not for raw TCP/UDP</p><p>Example: Multi-Path ALB Ingress</p><pre>apiVersion: networking.k8s.io/v1<br>kind: Ingress<br>metadata:<br>  name: ecommerce-ingress<br>  annotations:<br>    kubernetes.io/ingress.class: alb<br>    alb.ingress.kubernetes.io/scheme: internet-facing<br>    alb.ingress.kubernetes.io/listen-ports: &#39;[{&quot;HTTP&quot;:80,&quot;HTTPS&quot;:443}]&#39;<br>    alb.ingress.kubernetes.io/certificate-arn: arn:aws:acm:us-east-1:123456789012:certificate/abc<br>spec:<br>  rules:<br>  - host: shop.example.com<br>    http:<br>      paths:<br>      - path: /cart<br>        pathType: Prefix<br>        backend:<br>          service:<br>            name: cart-svc<br>            port: { number: 80 }<br>      - path: /checkout<br>        pathType: Prefix<br>        backend:<br>          service:<br>            name: checkout-svc<br>            port: { number: 80 }</pre><p>Use case: An e-commerce site routing /cart and /checkout to separate microservices.</p><p><strong>Option 3: Gateway API with VPC Lattice</strong></p><p>The Kubernetes Gateway API is the evolution of Ingress. It introduces new CRDs like:</p><p>•	GatewayClass → defines the type of load balancer (e.g., AWS VPC Lattice)</p><p>•	Gateway → the actual instance of a load balancer/proxy</p><p>•	HTTPRoute / GRPCRoute / TCPRoute → routes that app teams own to send traffic to their services</p><p>On AWS, the AWS Gateway API Controller translates these CRDs into Amazon VPC Lattice resources.</p><p>This unlocks a lot more than just ingress. Let’s see how.</p><p>Example 1: Multi-Service Internal Routing</p><pre>apiVersion: gateway.networking.k8s.io/v1<br>kind: GatewayClass<br>metadata:<br>  name: amazon-vpc-lattice<br>spec:<br>  controllerName: gateway.api.aws/controller<br>---<br>apiVersion: gateway.networking.k8s.io/v1<br>kind: Gateway<br>metadata:<br>  name: internal-gateway<br>  namespace: platform<br>spec:<br>  gatewayClassName: amazon-vpc-lattice<br>  listeners:<br>  - name: http<br>    protocol: HTTP<br>    port: 80<br>---<br>apiVersion: gateway.networking.k8s.io/v1<br>kind: HTTPRoute<br>metadata:<br>  name: orders-route<br>  namespace: app<br>spec:<br>  parentRefs:<br>  - name: internal-gateway<br>    namespace: platform<br>  rules:<br>  - matches:<br>    - path:<br>        type: PathPrefix<br>        value: /orders<br>    backendRefs:<br>    - name: orders-svc<br>      port: 80<br>---<br>apiVersion: gateway.networking.k8s.io/v1<br>kind: HTTPRoute<br>metadata:<br>  name: billing-route<br>  namespace: app<br>spec:<br>  parentRefs:<br>  - name: internal-gateway<br>    namespace: platform<br>  rules:<br>  - matches:<br>    - path:<br>        type: PathPrefix<br>        value: /billing<br>    backendRefs:<br>    - name: billing-svc<br>      port: 80</pre><p><strong>Outcome</strong>: One Gateway, multiple services routed by path. Platform team owns the Gateway, app teams own Routes.</p><p>Example 2: Cross-Cluster Service-to-Service</p><p>•	Cluster A hosts orders-svc.</p><p>•	Cluster B hosts billing-svc.</p><p>With VPC Lattice, they discover and call each other securely without VPC peering or exposing ALBs.</p><p><strong>Outcome</strong>: Simplified multi-cluster service networking.</p><p>Example 3: Traffic Policies with Retries &amp; Timeouts</p><pre>apiVersion: gateway.networking.k8s.io/v1<br>kind: HTTPRoute<br>metadata:<br>  name: resilient-orders<br>spec:<br>  parentRefs:<br>  - name: lattice-gw<br>  rules:<br>  - backendRefs:<br>    - name: orders-svc<br>      port: 80<br>  timeouts:<br>    request: 5s<br>  retries:<br>    attempts: 3<br>    perTryTimeout: 2s<br>    retryOn: [&quot;5xx&quot;, &quot;gateway-error&quot;]</pre><p><strong>Outcome</strong>: Built-in resilience without coding retry logic in clients.</p><p><strong>Gateway API Use Cases</strong></p><p>1.	Multi-cluster microservices → connect apps across EKS clusters.</p><p>2.	Platform vs app separation → platform defines Gateways, apps add Routes.</p><p>3.	Service-to-service (east — west) → AWS-managed alternative to Istio/Linkerd.</p><p>4.	Gradual migration → run ALB Ingress for north — south + Gateway API for east — west.</p><p>5.	Security → enforce IAM/mTLS between services.</p><p>6.	Traffic shaping → retries, timeouts, canary releases with weighted routes.</p><p><strong>Hybrid Approach: Best of All Worlds</strong></p><p>Most teams combine load balancers:</p><p>•	ALB Ingress → for public HTTP/S traffic</p><p>•	NLB → for gRPC or TCP/UDP workloads</p><p>•	Gateway API (VPC Lattice) → for service-to-service networking</p><p><strong>Setup Guide</strong></p><p><strong>Install AWS Load Balancer Controller (ALB + NLB)</strong></p><pre>helm repo add eks https://aws.github.io/eks-charts<br>helm repo update<br>kubectl create ns kube-system<br><br>helm upgrade --install aws-load-balancer-controller eks/aws-load-balancer-controller \<br>  -n kube-system \<br>  --set clusterName=&lt;your-cluster-name&gt; \<br>  --set serviceAccount.create=false \<br>  --set serviceAccount.name=aws-load-balancer-controller</pre><p><strong>Install AWS Gateway API Controller (VPC Lattice)</strong></p><pre>kubectl apply -k &quot;github.com/kubernetes-sigs/gateway-api/config/crd?ref=v1.0.0&quot;<br><br>helm repo add eks https://aws.github.io/eks-charts<br>helm repo update<br>kubectl create ns aws-gateway-controller<br><br>helm upgrade --install aws-gateway-controller eks/aws-gateway-controller \<br>  -n aws-gateway-controller</pre><p>✅ Now you can deploy NLB Services, ALB Ingresses, and Gateway API Routes.</p><p><strong>Cost Comparison</strong></p><p><strong>Network Load Balancer (NLB)</strong></p><p>•	Pricing: charged per LCU-hour (Load Balancer Capacity Unit) + data processed</p><p>•	Operates at Layer 4 (TCP/UDP)</p><p>•	Best for: gRPC, TCP/UDP apps, static IPs</p><p>•	No WAF or auth integration</p><p>•	No multi-cluster support</p><p><strong>Application Load Balancer (ALB)</strong></p><p>•	Pricing: charged per LCU-hour + data processed</p><p>•	Operates at Layer 7 (HTTP/S)</p><p>•	Best for: websites, APIs, path/host-based routing</p><p>•	Supports WAF, SSL termination, OIDC authentication</p><p>•	No native multi-cluster support</p><p><strong>Gateway API with VPC Lattice</strong></p><p>•	Pricing: per service-hour + request units</p><p>•	Operates at both L4 and L7 using Gateway API</p><p>•	Best for: service-to-service networking, multi-cluster</p><p>•	Supports IAM/mTLS policies, retries, timeouts</p><p>•	Provides multi-cluster and cross-VPC connectivity</p><p><strong>Tip</strong>:</p><p>•	For very high request volumes → NLB is cheapest.</p><p>•	For feature-rich HTTP routing → ALB is the sweet spot.</p><p>•	For multi-cluster networking → VPC Lattice (Gateway API) is the only option.</p><p><strong>Conclusion</strong></p><p>Amazon EKS doesn’t lock you into one option — you can (and should) combine load balancers based on workload needs:</p><p>•	Use ALB for public web apps and REST APIs.</p><p>•	Use NLB for high-performance TCP/UDP or gRPC workloads.</p><p>•	Use Gateway API with VPC Lattice for scalable, multi-cluster service networking.</p><p><strong>The future of Kubernetes traffic management on AWS is Gateway API — but ALB and NLB remain the foundation for most clusters today.</strong></p><p>And before you go, don’t forget to clap and follow the writer️! ❤️</p><h3>A message from our Founder</h3><p><strong>Hey, </strong><a href="https://linkedin.com/in/sunilsandhu"><strong>Sunil</strong></a><strong> here.</strong> I wanted to take a moment to thank you for reading until the end and for being a part of this community.</p><p>Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? <strong>We don’t receive any funding, we do this to support the community. ❤️</strong></p><p>If you want to show some love, please take a moment to <strong>follow me on </strong><a href="https://linkedin.com/in/sunilsandhu"><strong>LinkedIn</strong></a><strong>, </strong><a href="https://tiktok.com/@messyfounder"><strong>TikTok</strong></a>, <a href="https://instagram.com/sunilsandhu"><strong>Instagram</strong></a>. You can also subscribe to our <a href="https://newsletter.plainenglish.io/"><strong>weekly newsletter</strong></a>.</p><p>And before you go, don’t forget to <strong>clap</strong> and <strong>follow</strong> the writer️!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcf7a795a3f2" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/load-balancing-in-amazon-eks-alb-nlb-and-gateway-api-explained-bcf7a795a3f2">Load Balancing in Amazon EKS: ALB, NLB, and Gateway API Explained</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How We Combined EKS MCP Server with Amazon Q to Slash RTA and DevOps Time]]></title>
            <link>https://medium.com/codetodeploy/how-we-combined-eks-mcp-server-with-amazon-q-to-slash-rta-and-devops-time-819e53e3775c?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/819e53e3775c</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[amazon-q]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Wed, 24 Sep 2025 01:04:54 GMT</pubDate>
            <atom:updated>2025-09-27T18:03:50.869Z</atom:updated>
            <content:encoded><![CDATA[<p>Every DevOps team knows the pain: incidents happen at odd hours, triage eats up precious time, and repetitive fixes consume bandwidth. We wanted a way to <strong>shorten RTA (Response Time to Acknowledge) and reduce DevOps cycles spent on firefighting</strong> without compromising safety.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*9Y9BkqujiYAGijbGgm8m8w.jpeg" /></figure><p>Our solution? We combined an <strong>MCP (Model Context Protocol) server for EKS with Amazon Q</strong>, giving our operations team an AI “co-pilot” that inspects clusters, diagnoses issues, and suggests guarded remediations. The result: <strong>40 — 60% faster triage</strong>, fewer escalations, and more sleep for our on-call engineers.</p><p><strong>Why We Did This</strong></p><p>•	<strong>Faster triage:</strong> Reduce the time between alert → root cause.</p><p>•	<strong>Consistency</strong>: Encode tribal knowledge into reusable AI playbooks.</p><p>•	<strong>Safety:</strong> Ensure any AI-driven action is gated by approval and rollback.</p><p>•	<strong>Cost optimization</strong>: Spot noisy pods and runaway logs before the bill spikes.</p><p><strong>High-Level Architecture</strong></p><p>•	<strong>Amazon Q</strong>: Conversational interface for troubleshooting.</p><p>•	<strong>MCP Server:</strong> Safe, RBAC-scoped tool layer for EKS.</p><p>•	<strong>Observability hooks:</strong> Prometheus + CloudWatch for logs/metrics.</p><p>•	<strong>Guardrails:</strong> Policies, approvals, rollback plans.</p><p><strong>Day-to-Day in Action</strong></p><ol><li><strong>Triage</strong></li></ol><p>Instead of digging into kubectl manually:</p><p>“Q, show top failing pods in prod for the last 30 minutes.”</p><p>•	Q calls MCP: kubectl get/describe + logs.</p><p>•	Summarizes root causes (CrashLoopBackOff, OOMKilled).</p><p>•	Suggests next steps.</p><p><strong>2. Explainability</strong></p><p>Q explains its reasoning in plain English:</p><p>“fluent-bit is OOMKilled. Current limit 128Mi, observed peak 210Mi. Propose raising to 220Mi requests / 250Mi limits.”</p><p><strong>3. Safe Remediation</strong></p><p>•	Q prepares a manifest/Helm patch.</p><p>•	Posts a diff preview and rollback command.</p><p>•	Waits for /approve.</p><p>Example patch:</p><pre>spec:<br>  template:<br>    spec:<br>      containers:<br>      - name: orders-api<br>        resources:<br>          requests:<br>            cpu: &quot;400m&quot;<br>            memory: &quot;512Mi&quot;<br>          limits:<br>            cpu: &quot;800m&quot;<br>            memory: &quot;768Mi&quot;<br>        readinessProbe:<br>          httpGet: { path: /healthz, port: 8080 }<br>          periodSeconds: 5<br>          failureThreshold: 2</pre><p><strong>4. Verification</strong></p><p>After applying, Q checks:</p><p>•	Pod restarts decreased.</p><p>•	Error rate/5xx dropped.</p><p>•	SLO probe is green.</p><p><strong>Guardrails We Enforced</strong></p><p>•	RBAC read-only by default, write only on approval.</p><p>•	Namespace allow-list (no prod writes without Incident Commander).</p><p>•	Change windows respected.</p><p>•	Audit trail stored in Slack + S3.</p><p>•	Automated rollback if health doesn’t recover.</p><p><strong>What We Automated with Playbooks</strong></p><p>•	<strong>Runaway logs:</strong> Apply Fluent Bit filters + log sampling.</p><p>•	<strong>OOMKilled pods:</strong> Tune requests/limits to observed P95.</p><p>•	<strong>ALB 5xx spikes:</strong> Identify bad rollouts → rollback to stable.</p><p>•	<strong>HPA flapping:</strong> Detect over-aggressive scaling → propose sane thresholds.</p><p><strong>Outcomes</strong></p><p>•	<strong>Triage time:</strong> 20 — 30 min → 8 — 12 min.</p><p>•	<strong>RTA</strong>: Cut by 30 — 50% in noisy incidents.</p><p>•	<strong>Escalations</strong>: Reduced L3 involvement for repetitive fixes.</p><p>•	<strong>Costs</strong>: CloudWatch spend lowered after log-volume tuning.</p><p><strong>How You Can Replicate</strong></p><p>1.	Deploy MCP server with Kubernetes tool adapters.</p><p>2.	Connect Amazon Q and define approval workflows.</p><p>3.	Lock down access with scoped RBAC.</p><p>4.	Encode top incident classes into Q playbooks.</p><p>5.	Add observability (Prometheus/CloudWatch).</p><p>6.	Track metrics: RTA, MTTR, change success rate.</p><p><strong>Lessons Learned</strong></p><p>•	Start <strong>read-only</strong> first to build trust.</p><p>•	Always demand diff + rollback from AI.</p><p>•	Post-change health checks save time and avoid false positives.</p><p>•	Prefer Helm value patches if workloads are Helm-managed (avoid drift).</p><p><strong>Conclusion</strong></p><p>By putting <strong>Amazon Q + EKS MCP server</strong> into the hands of our Cloud Ops team, we moved from manual runbooks to AI-assisted, guardrailed automation. The biggest win wasn’t just speed — it was turning tribal knowledge into repeatable workflows that anyone on-call can execute with confidence.</p><p>The payoff? <strong>More uptime, less stress, and fewer 2 a.m. firefights.</strong></p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*efJSs2jm59kfBoXs9j0YBA.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://x.com/Bhuwanchet67277"><strong>X</strong></a> | <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p>👉 <strong>Follow our publication, </strong><a href="https://medium.com/codetodeploy"><strong>CodeToDeploy</strong></a></p><p><strong>Note:</strong> This Post may contain affiliate links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=819e53e3775c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/how-we-combined-eks-mcp-server-with-amazon-q-to-slash-rta-and-devops-time-819e53e3775c">How We Combined EKS MCP Server with Amazon Q to Slash RTA and DevOps Time</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Terraform AWS Provider 6.0: Change the Whole Game]]></title>
            <link>https://medium.com/codetodeploy/terraform-aws-provider-6-0-change-the-whole-game-10a4e20aa886?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/10a4e20aa886</guid>
            <category><![CDATA[infrastructure-as-code]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cloud-engineering]]></category>
            <category><![CDATA[terraform]]></category>
            <category><![CDATA[aws]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Tue, 23 Sep 2025 08:47:03 GMT</pubDate>
            <atom:updated>2025-09-27T18:05:16.815Z</atom:updated>
            <content:encoded><![CDATA[<p>Terraform has been the backbone of Infrastructure as Code (IaC) for years, and the AWS Provider is its most widely used plugin. With the release of Terraform AWS Provider 6.0, HashiCorp and AWS have introduced breaking changes, performance optimizations, and a refreshed approach to resource management. This update isn’t just a version bump — it reshapes how teams provision and manage AWS infrastructure at scale.</p><p>In this blog, we’ll break down what’s new in AWS Provider 6.0, the challenges it introduces, and how you can adapt your workflows to harness its full power.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bgEj3Fud55VAQ-ZPrVcVVg.jpeg" /></figure><p><strong>What’s New in AWS Provider 6.0</strong></p><ol><li><strong>Breaking Resource and Data Source Changes</strong></li></ol><p>•	Refactoring of Core Resources: Several AWS resources (like VPC, Subnets, and IAM policies) have been updated for consistency with AWS APIs. Attributes that were previously optional may now be required.</p><p>•	Deprecated Fields Removed: Backward-compatibility is no longer guaranteed for older field names or behaviors. Expect some refactoring when upgrading existing Terraform states.</p><p><strong>2. Identity and Access Management (IAM) Improvements</strong></p><p>•	Granular Policy Handling: The new provider streamlines IAM role and policy creation, aligning closer to AWS best practices.</p><p>•	Stricter Validation: Misconfigured JSON policies that used to “pass” will now fail validation, forcing better governance.</p><p><strong>3. Enhanced Performance</strong></p><p>•	Parallelism Improvements: Large infrastructure plans (think 1,000+ resources) see reduced execution times thanks to optimized API calls.</p><p>•	Retry and Throttling Controls: New built-in mechanisms handle AWS API throttling gracefully, reducing flaky runs.</p><p><strong>4. Regional + Multi-Account Workflows</strong></p><p>•	Better Multi-Region Support: Managing resources across AWS regions is simpler with new data source enhancements.</p><p>•	Provider Aliases Streamlined: Cleaner syntax when managing multiple accounts or organizations.</p><p><strong>Migration Considerations</strong></p><p>Upgrading to AWS Provider 6.0 isn’t “plug and play.” Here’s what you need to check:</p><p>•	State Files: Run terraform plan in a safe environment first. Look out for resource recreation due to field renames or schema changes.</p><p>•	IAM Policy Updates: If you’ve been relying on loose policy JSON, expect failures until you update them.</p><p>•	Module Compatibility: Community modules (like VPC or EKS) may need version bumps — check their changelogs before applying.</p><p>•	CI/CD Pipelines: Update your pipeline lock files (.terraform.lock.hcl) to pin the correct version.</p><p><strong>Example: VPC Resource Update</strong></p><pre># Old (Terraform AWS Provider 5.x)<br>resource &quot;aws_vpc&quot; &quot;main&quot; {<br>  cidr_block = &quot;10.0.0.0/16&quot;<br>  enable_dns_support = true<br>  enable_dns_hostnames = true<br>}<br><br># New (Terraform AWS Provider 6.x)<br>resource &quot;aws_vpc&quot; &quot;main&quot; {<br>  cidr_block = &quot;10.0.0.0/16&quot;<br>  dns_support   = true<br>  dns_hostnames = true<br>}</pre><p>Notice how enable_dns_support and enable_dns_hostnames are renamed to shorter, cleaner fields in v6.0.</p><p><strong>S3 Bucket Ownership Controls</strong></p><pre># Old (v5.x)<br>resource &quot;aws_s3_bucket&quot; &quot;example&quot; {<br>  bucket = &quot;my-app-bucket&quot;<br>  acl    = &quot;private&quot;<br>}<br><br># New (v6.0)<br>resource &quot;aws_s3_bucket&quot; &quot;example&quot; {<br>  bucket = &quot;my-app-bucket&quot;<br>}<br><br>resource &quot;aws_s3_bucket_ownership_controls&quot; &quot;example&quot; {<br>  bucket = aws_s3_bucket.example.id<br>  rule {<br>    object_ownership = &quot;BucketOwnerEnforced&quot;<br>  }<br>}</pre><p><strong>IAM Role Inline Policy</strong></p><pre># Old (v5.x)<br>resource &quot;aws_iam_role_policy&quot; &quot;inline&quot; {<br>  name = &quot;inline-policy&quot;<br>  role = aws_iam_role.example.id<br><br>  policy = &lt;&lt;EOF<br>{<br>  &quot;Version&quot;: &quot;2012-10-17&quot;,<br>  &quot;Statement&quot;: [<br>    {<br>      &quot;Action&quot;: [&quot;s3:ListBucket&quot;],<br>      &quot;Effect&quot;: &quot;Allow&quot;,<br>      &quot;Resource&quot;: &quot;*&quot;<br>    }<br>  ]<br>}<br>EOF<br>}<br><br># New (v6.0)<br>resource &quot;aws_iam_role_policy&quot; &quot;inline&quot; {<br>  name   = &quot;inline-policy&quot;<br>  role   = aws_iam_role.example.id<br>  policy = jsonencode({<br>    Version = &quot;2012-10-17&quot;<br>    Statement = [<br>      {<br>        Action   = [&quot;s3:ListBucket&quot;]<br>        Effect   = &quot;Allow&quot;<br>        Resource = &quot;*&quot;<br>      }<br>    ]<br>  })<br>}</pre><p><strong>EKS Node Group Scaling</strong></p><pre># Old (v5.x)<br>resource &quot;aws_eks_node_group&quot; &quot;example&quot; {<br>  cluster_name    = aws_eks_cluster.demo.name<br>  node_group_name = &quot;demo-nodes&quot;<br>  node_role_arn   = aws_iam_role.node.arn<br>  subnet_ids      = aws_subnet.private[*].id<br><br>  scaling_config {<br>    desired_size = 2<br>    max_size     = 5<br>    min_size     = 1<br>  }<br>}<br><br># New (v6.0)<br>resource &quot;aws_eks_node_group&quot; &quot;example&quot; {<br>  cluster_name    = aws_eks_cluster.demo.name<br>  node_group_name = &quot;demo-nodes&quot;<br>  node_role_arn   = aws_iam_role.node.arn<br>  subnet_ids      = aws_subnet.private[*].id<br><br>  scaling_config {<br>    desired_size = 2<br>    max_size     = 5<br>    min_size     = 1<br>  }<br><br>  update_config {<br>    max_unavailable = 1<br>  }<br>}</pre><p><strong>Multi-Region Provider Aliases</strong></p><pre># Old (v5.x)<br>provider &quot;aws&quot; {<br>  region = &quot;us-east-1&quot;<br>}<br><br>provider &quot;aws&quot; {<br>  alias  = &quot;west&quot;<br>  region = &quot;us-west-2&quot;<br>}<br><br># New (v6.0)<br>provider &quot;aws&quot; {<br>  region = &quot;us-east-1&quot;<br>  alias  = &quot;east&quot;<br>}<br><br>provider &quot;aws&quot; {<br>  region = &quot;us-west-2&quot;<br>  alias  = &quot;west&quot;<br>}</pre><p><strong>Single configuration, multi-Region EKS</strong></p><pre># One provider block; no aliases needed<br>provider &quot;aws&quot; {<br>  region = &quot;us-east-1&quot; # default; can be overridden per resource<br>}<br><br># Primary cluster in us-east-1<br>resource &quot;aws_eks_cluster&quot; &quot;primary&quot; {<br>  name     = &quot;eks-primary&quot;<br>  role_arn = aws_iam_role.eks_control_plane.arn<br>  version  = &quot;1.33&quot;<br><br>  region = &quot;us-east-1&quot; # &lt;- new in v6: per-resource region<br>  vpc_config {<br>    subnet_ids = [aws_subnet.a.id, aws_subnet.b.id]<br>  }<br>}<br><br># DR cluster in us-west-2 using the same provider, overridden at resource-level<br>resource &quot;aws_eks_cluster&quot; &quot;dr&quot; {<br>  name     = &quot;eks-dr&quot;<br>  role_arn = aws_iam_role.eks_control_plane.arn<br>  version  = &quot;1.33&quot;<br><br>  region = &quot;us-west-2&quot; # &lt;- no provider alias required<br>  vpc_config {<br>    subnet_ids = [aws_subnet.c.id, aws_subnet.d.id]<br>  }<br>}</pre><p>Why it matters: v6 lets you place region directly on most resources, eliminating many provider aliases and cutting memory overhead.</p><p><strong>Managed add-ons pinned &amp; upgrade-ready (CoreDNS, VPC CNI, kube-proxy)</strong></p><pre># CoreDNS<br>resource &quot;aws_eks_addon&quot; &quot;coredns&quot; {<br>  cluster_name             = aws_eks_cluster.primary.name<br>  addon_name               = &quot;coredns&quot;<br>  addon_version            = &quot;v1.11.4-eksbuild.1&quot; # pin exact version<br>  resolve_conflicts        = &quot;OVERWRITE&quot;<br>  service_account_role_arn = null<br>  region                   = &quot;us-east-1&quot;<br>}<br><br># VPC CNI<br>resource &quot;aws_eks_addon&quot; &quot;vpc_cni&quot; {<br>  cluster_name      = aws_eks_cluster.primary.name<br>  addon_name        = &quot;vpc-cni&quot;<br>  addon_version     = &quot;v1.18.2-eksbuild.1&quot;<br>  resolve_conflicts = &quot;OVERWRITE&quot;<br>  region            = &quot;us-east-1&quot;<br>}<br><br># kube-proxy<br>resource &quot;aws_eks_addon&quot; &quot;kube_proxy&quot; {<br>  cluster_name      = aws_eks_cluster.primary.name<br>  addon_name        = &quot;kube-proxy&quot;<br>  addon_version     = &quot;v1.33.0-eksbuild.1&quot;<br>  resolve_conflicts = &quot;OVERWRITE&quot;<br>  region            = &quot;us-east-1&quot;<br>}</pre><p>Tip: Pinned versions make control-plane upgrades predictable; upgrade add-ons after the control plane, then roll your nodes. (Basic sequence aligned with AWS guidance.)</p><p><strong>IRSA the modern, strict way (OIDC + aws_iam_role trust)</strong></p><pre># Discover the cluster OIDC issuer<br>data &quot;aws_eks_cluster&quot; &quot;this&quot; {<br>  name   = aws_eks_cluster.primary.name<br>  region = &quot;us-east-1&quot;<br>}<br><br>data &quot;aws_eks_cluster_auth&quot; &quot;this&quot; {<br>  name   = aws_eks_cluster.primary.name<br>  region = &quot;us-east-1&quot;<br>}<br><br>data &quot;tls_certificate&quot; &quot;oidc&quot; {<br>  url = data.aws_eks_cluster.this.identity[0].oidc[0].issuer<br>}<br><br># OIDC provider for IRSA<br>resource &quot;aws_iam_openid_connect_provider&quot; &quot;eks&quot; {<br>  url             = data.tls_certificate.oidc.url<br>  client_id_list  = [&quot;sts.amazonaws.com&quot;]<br>  thumbprint_list = [data.tls_certificate.oidc.certificates[0].sha1_fingerprint]<br>}<br><br># Role for a service account (e.g., external-dns)<br>data &quot;aws_iam_policy_document&quot; &quot;irsa_trust&quot; {<br>  statement {<br>    actions = [&quot;sts:AssumeRoleWithWebIdentity&quot;]<br>    principals {<br>      type        = &quot;Federated&quot;<br>      identifiers = [aws_iam_openid_connect_provider.eks.arn]<br>    }<br>    condition {<br>      test     = &quot;StringEquals&quot;<br>      variable = &quot;${replace(data.aws_eks_cluster.this.identity[0].oidc[0].issuer, &quot;https://&quot;, &quot;&quot;)}:sub&quot;<br>      values   = [&quot;system:serviceaccount:kube-system:external-dns&quot;]<br>    }<br>  }<br>}<br><br>resource &quot;aws_iam_role&quot; &quot;external_dns&quot; {<br>  name               = &quot;eks-external-dns&quot;<br>  assume_role_policy = data.aws_iam_policy_document.irsa_trust.json<br>}<br><br># Attach minimal policy…</pre><p><strong>Why This Release Changes the Game</strong></p><p><strong>1.	Cleaner Syntax = Less Cognitive Load</strong></p><p>Terraform configurations now map more closely to AWS’s own API and console terminology, reducing confusion for newcomers.</p><p><strong>2.	Performance Optimizations</strong></p><p>Enterprises managing thousands of resources will see tangible reductions in plan and apply times.</p><p><strong>3.	Security by Default</strong></p><p>Stricter IAM policy validation means fewer misconfigurations slip through, aligning with the “shift-left security” mindset.</p><p><strong>4.	Future-Proofing</strong></p><p>With old cruft removed, the provider is more aligned with AWS’s evolving API surface, ensuring smoother adoption of new AWS services.</p><p><strong>Best Practices for Adopting v6.0</strong></p><p>•	Test in Non-Prod First: Use workspaces or isolated accounts to validate upgrades.</p><p>•	Pin Versions: Lock the provider in your versions.tf to avoid surprise updates.</p><p>•	Use terraform plan + terraform state mv: Don’t accept unnecessary resource recreation blindly.</p><p>•	Follow Module Authors: Stay updated with community module maintainers who are patching for v6 compatibility.</p><p><strong>Conclusion</strong></p><p>Terraform AWS Provider 6.0 is more than an incremental upgrade — it’s a paradigm shift. By enforcing better security, streamlining syntax, and delivering faster performance, it sets a new standard for IaC on AWS.</p><p>Yes, the migration may cause a few headaches initially, but the long-term benefits of cleaner code, reduced risk, and better scalability make the move worthwhile.</p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*efJSs2jm59kfBoXs9j0YBA.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://x.com/Bhuwanchet67277"><strong>X</strong></a> | <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p>👉 <strong>Follow our publication, </strong><a href="https://medium.com/codetodeploy"><strong>CodeToDeploy</strong></a></p><p><strong>Note:</strong> This Post may contain affiliate links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10a4e20aa886" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/terraform-aws-provider-6-0-change-the-whole-game-10a4e20aa886">Terraform AWS Provider 6.0: Change the Whole Game</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ExternalDNS on Amazon EKS: Automating Route 53 Records from Ingress to Multi-Region Failover]]></title>
            <link>https://aws.plainenglish.io/externaldns-on-amazon-eks-automating-route-53-records-from-ingress-to-multi-region-failover-4768f8330cda?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/4768f8330cda</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[external-dns]]></category>
            <category><![CDATA[kubernetes]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Thu, 11 Sep 2025 00:52:22 GMT</pubDate>
            <atom:updated>2025-09-27T18:09:51.780Z</atom:updated>
            <content:encoded><![CDATA[<p>Managing DNS records for Kubernetes services can quickly become tedious and error-prone. Every time you deploy a new application or expose a service through a LoadBalancer or Ingress, you often need to manually update DNS records in Route 53 or other DNS providers.</p><p>This is where ExternalDNS comes in. It automates DNS record management, keeping your DNS zones in sync with the resources running inside your Amazon EKS cluster. In this blog, we’ll explore how ExternalDNS works, why it matters, and how to set it up with EKS + AWS Route 53.</p><p><strong>How ExternalDNS Works</strong></p><p>ExternalDNS runs as a Kubernetes controller that:</p><p>1.	Watches Kubernetes resources such as Services and Ingresses.</p><p>2.	Reads their annotations for desired DNS names.</p><p>3.	Automatically creates, updates, or deletes DNS records in your provider (like Route 53).</p><p>This means developers only need to define DNS requirements in Kubernetes manifests. ExternalDNS handles the rest.</p><p><strong>Why Use ExternalDNS on EKS</strong></p><p>•	<strong>Automation</strong>: no more manual Route 53 updates.</p><p>•	<strong>GitOps-friendly</strong>: DNS changes are declarative, tracked in Git.</p><p>•	<strong>Multi-provider support:</strong> works with Route 53, Cloudflare, Google DNS, and more.</p><p>•	<strong>Dynamic environments:</strong> perfect for CI/CD and ephemeral environments.</p><p><strong>Setups:</strong></p><p><strong>Step 1: Create IAM Policy and Role for ExternalDNS</strong></p><p>ExternalDNS needs permissions to change DNS records.</p><pre>cat &lt;&lt;EOF &gt; externaldns-policy.json<br>{<br>  &quot;Version&quot;: &quot;2012-10-17&quot;,<br>  &quot;Statement&quot;: [<br>    {<br>      &quot;Effect&quot;: &quot;Allow&quot;,<br>      &quot;Action&quot;: [<br>        &quot;route53:ChangeResourceRecordSets&quot;,<br>        &quot;route53:ListResourceRecordSets&quot;<br>      ],<br>      &quot;Resource&quot;: [<br>        &quot;arn:aws:route53:::hostedzone/YOUR_HOSTED_ZONE_ID&quot;<br>      ]<br>    },<br>    {<br>      &quot;Effect&quot;: &quot;Allow&quot;,<br>      &quot;Action&quot;: &quot;route53:ListHostedZones&quot;,<br>      &quot;Resource&quot;: [&quot;*&quot;]<br>    }<br>  ]<br>}<br>EOF<br><br>aws iam create-policy \<br>  --policy-name AllowExternalDNSUpdates \<br>  --policy-document file://externaldns-policy.json</pre><p>Then, create an IAM role for the ExternalDNS service account using IAM OIDC provider:</p><pre>eksctl utils associate-iam-oidc-provider --cluster your-cluster --approve<br><br>eksctl create iamserviceaccount \<br>  --name external-dns \<br>  --namespace kube-system \<br>  --cluster your-cluster \<br>  --attach-policy-arn arn:aws:iam::ACCOUNT_ID:policy/AllowExternalDNSUpdates \<br>  --approve</pre><p><strong>Step 2: Deploy ExternalDNS</strong></p><p>You can install ExternalDNS using Helm:</p><pre>helm repo add external-dns https://kubernetes-sigs.github.io/external-dns/<br>helm upgrade --install external-dns external-dns/external-dns \<br>  --namespace kube-system \<br>  --set provider=aws \<br>  --set aws.zoneType=public \<br>  --set registry=txt \<br>  --set txtOwnerId=external-dns \<br>  --set serviceAccount.name=external-dns</pre><p><strong>Step 3: Annotate Services or Ingress</strong></p><p><strong>Example: Service</strong></p><pre>apiVersion: v1<br>kind: Service<br>metadata:<br>  name: my-app<br>  namespace: default<br>  annotations:<br>    external-dns.alpha.kubernetes.io/hostname: myapp.example.com<br>spec:<br>  type: LoadBalancer<br>  ports:<br>    - port: 80<br>      targetPort: 8080<br>  selector:<br>    app: my-app</pre><p><strong>Example: Ingress</strong></p><pre>apiVersion: networking.k8s.io/v1<br>kind: Ingress<br>metadata:<br>  name: my-app-ingress<br>  annotations:<br>    kubernetes.io/ingress.class: alb<br>    external-dns.alpha.kubernetes.io/hostname: myapp.example.com<br>spec:<br>  rules:<br>    - host: myapp.example.com<br>      http:<br>        paths:<br>        - path: /<br>          pathType: Prefix<br>          backend:<br>            service:<br>              name: my-app<br>              port:<br>                number: 80</pre><p><strong>How It Works with an Ingress Annotation</strong></p><p>1.	You apply an Ingress with the annotation external-dns.alpha.kubernetes.io/hostname: shop.example.com.</p><p>2.	The Ingress controller provisions an ALB and writes the ALB’s hostname into the Ingress status.loadBalancer.</p><p>3.	ExternalDNS detects the annotation (desired DNS) and the target (ALB hostname from status).</p><p>4.	ExternalDNS updates Route 53:</p><p>•	An A/ALIAS record for shop.example.com pointing to the ALB.</p><p>•	A TXT record for ownership tracking.</p><p>5.	DNS goes live: shop.example.com resolves to the ALB.</p><p>6.	Continuous reconciliation ensures Route 53 stays updated as cluster state changes.</p><p><strong>Failover with ExternalDNS</strong></p><p>For active/passive setups (multi-region or multi-cluster), you can define PRIMARY and SECONDARY records for the same hostname. ExternalDNS will create the corresponding Route 53 failover records automatically.</p><p><strong>Primary Ingress (with Health Check)</strong></p><pre>apiVersion: networking.k8s.io/v1<br>kind: Ingress<br>metadata:<br>  name: shop-primary<br>  annotations:<br>    kubernetes.io/ingress.class: alb<br>    external-dns.alpha.kubernetes.io/hostname: shop.example.com<br>    external-dns.alpha.kubernetes.io/set-identifier: shop-east<br>    external-dns.alpha.kubernetes.io/aws-failover: &quot;PRIMARY&quot;<br>    external-dns.alpha.kubernetes.io/aws-health-check-id: &quot;HC1234567890&quot;<br>spec:<br>  rules:<br>    - host: shop.example.com<br>      http:<br>        paths:<br>          - path: /<br>            pathType: Prefix<br>            backend:<br>              service:<br>                name: shop-svc<br>                port:<br>                  number: 80</pre><p><strong>Secondary Ingress</strong></p><pre>apiVersion: networking.k8s.io/v1<br>kind: Ingress<br>metadata:<br>  name: shop-secondary<br>  annotations:<br>    kubernetes.io/ingress.class: alb<br>    external-dns.alpha.kubernetes.io/hostname: shop.example.com<br>    external-dns.alpha.kubernetes.io/set-identifier: shop-west<br>    external-dns.alpha.kubernetes.io/aws-failover: &quot;SECONDARY&quot;<br>spec:<br>  rules:<br>    - host: shop.example.com<br>      http:<br>        paths:<br>          - path: /<br>            pathType: Prefix<br>            backend:<br>              service:<br>                name: shop-svc<br>                port:<br>                  number: 80</pre><p>With this setup:</p><p>•	Route 53 directs traffic to the PRIMARY as long as it’s healthy.</p><p>•	If the PRIMARY becomes unhealthy, Route 53 automatically switches DNS resolution to the SECONDARY.</p><p><strong>Why Failover Requires Health Checks</strong></p><p>Route 53 will only fail over if it has a health signal for the PRIMARY record. Without this signal, Route 53 will keep returning the PRIMARY, even if it is down. There are two ways to provide health information:</p><p><strong>Case A — ALIAS with Evaluate Target Health</strong></p><p>If the record is an ALIAS to an AWS Load Balancer (ALB or NLB), Route 53 can infer health from the load balancer. With Evaluate Target Health enabled:</p><p>•	If no healthy targets remain, the ALIAS record is considered unhealthy.</p><p>•	Route 53 then stops returning the PRIMARY and switches to the SECONDARY.</p><p><strong>Case B — Explicit Route 53 Health Check</strong></p><p>If you want application-level visibility, create a Route 53 Health Check (for example, HTTPS GET /healthz) and attach its ID to the PRIMARY via annotation:</p><pre>metadata:<br>  annotations:<br>    external-dns.alpha.kubernetes.io/hostname: shop.example.com<br>    external-dns.alpha.kubernetes.io/set-identifier: shop-east<br>    external-dns.alpha.kubernetes.io/aws-failover: &quot;PRIMARY&quot;<br>    external-dns.alpha.kubernetes.io/aws-health-check-id: &quot;HC1234567890&quot;</pre><p>Route 53 will probe that endpoint at regular intervals. If it fails, the PRIMARY is marked unhealthy, and Route 53 directs traffic to the SECONDARY.</p><p><strong>What Happens at Query Time</strong></p><p>1.	Client requests shop.example.com.</p><p>2.	Route 53 checks PRIMARY’s health:</p><p>•	Healthy → returns PRIMARY (east ALB).</p><p>•	Unhealthy → returns SECONDARY (west ALB).</p><p>3.	DNS TTL (e.g., 60 — 120 seconds) determines how quickly clients see the switch.</p><p><strong>Quick Checklist</strong></p><p>•	PRIMARY must have a health signal:</p><p>•	Either Evaluate Target Health for ALIAS records.</p><p>•	Or a Route 53 Health Check ID attached.</p><p>•	SECONDARY does not need a health check.</p><p>•	Keep DNS TTL low for faster switchover.</p><p>•	Use a dedicated app endpoint (/healthz or /readyz) for reliability.</p><p><strong>Multi-Region with ExternalDNS (us-east-1 and us-west-2)</strong></p><p>Running two EKS clusters across us-east-1 and us-west-2 is common in production. With ExternalDNS + Route 53, you can manage DNS declaratively and achieve either failover or active/active routing.</p><p><strong>Pattern A: Active/Passive Failover</strong></p><p>•	East = PRIMARY, West = SECONDARY.</p><p>•	PRIMARY must have a health signal:</p><p>•	Use Evaluate Target Health if the record is an ALIAS to an ALB/NLB.</p><p>•	Or attach a Route 53 Health Check ID for explicit application-level probes (/healthz).</p><p><strong>Primary</strong>:</p><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-east<br>external-dns.alpha.kubernetes.io/aws-failover: &quot;PRIMARY&quot;<br>external-dns.alpha.kubernetes.io/aws-health-check-id: &quot;HC1234567890&quot;<br></pre><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-west<br>external-dns.alpha.kubernetes.io/aws-failover: &quot;SECONDARY&quot;</pre><p>When the health check for PRIMARY fails, Route 53 automatically switches DNS answers to SECONDARY.</p><p><strong>Pattern B: Active/Active Weighted</strong></p><p>Split traffic between east and west (e.g., 70% east, 30% west).</p><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-east<br>external-dns.alpha.kubernetes.io/aws-weight: &quot;70&quot;</pre><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-west<br>external-dns.alpha.kubernetes.io/aws-weight: &quot;30&quot;</pre><p>Use health checks here if you want Route 53 to automatically exclude an unhealthy region.</p><p><strong>Pattern C: Latency-Based Routing</strong></p><p>Route 53 directs users to the closest cluster</p><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-east<br>external-dns.alpha.kubernetes.io/aws-latency: &quot;us-east-1&quot;</pre><pre>external-dns.alpha.kubernetes.io/set-identifier: shop-west<br>external-dns.alpha.kubernetes.io/aws-latency: &quot;us-west-2&quot;</pre><p><strong>Data Considerations</strong></p><p>•	Stateless services scale easily across regions.</p><p>•	Use Aurora Global Database or DynamoDB Global Tables for multi-region data.</p><p>•	Replicate S3 buckets with Cross-Region Replication.</p><p>•	Externalize sessions with Redis or DynamoDB.</p><p><strong>DNS Strategy Cheat-Sheet</strong></p><p>•	set-identifier must be unique per record.</p><p>•	txtOwnerId unique per cluster.</p><p>•	aws-failover for active/passive.</p><p>•	aws-weight for traffic splits.</p><p>•	aws-latency for latency-based routing.</p><p>•	TTL: keep low (60 — 120 seconds).</p><p><strong>Cross-Account Best Practice</strong></p><p>•	Host the public zone in a central DNS account.</p><p>•	Each app account/cluster gets an IRSA role scoped to that hosted zone ARN.</p><p>•	ExternalDNS assumes its role and manages only the intended zone.</p><p><strong>Runbook for Incidents</strong></p><p>1.	Check health of primary cluster and ALB.</p><p>2.	Verify ExternalDNS logs.</p><p>3.	Adjust routing (weights or failover).</p><p>4.	Rollback by restoring health or weights.</p><p><strong>When to Use AWS Global Accelerator</strong></p><p>DNS failover works, but caching can delay recovery. If you need static global IPs, sub-second failover, or TCP/UDP anycast, put AWS Global Accelerator in front of your ALBs. ExternalDNS points to GA, and GA handles health-based routing.</p><p><strong>Best Practices</strong></p><p>•	Use txtOwnerId to prevent conflicts across clusters.</p><p>•	Always use a unique set-identifier when creating multiple records for the same hostname.</p><p>•	Scope ExternalDNS with domainFilters or zone-id-filter.</p><p>•	Start with policy=upsert-only in shared zones.</p><p>•	Keep TTLs low (60 — 120 seconds) for faster failover.</p><p>•	Use kubectl logs -n kube-system deploy/external-dns for debugging.</p><p><strong>Conclusion</strong></p><p>ExternalDNS bridges the gap between Kubernetes and DNS, making your applications accessible without manual intervention. On Amazon EKS, it integrates seamlessly with Route 53, giving you a fully automated and declarative DNS management workflow.</p><p>With Ingress annotations, failover policies, and multi-region patterns, you can build highly available global architectures powered entirely by Kubernetes manifests.</p><p>And before you go, don’t forget to <strong>clap</strong> and <strong>follow</strong> the writer️!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4768f8330cda" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/externaldns-on-amazon-eks-automating-route-53-records-from-ingress-to-multi-region-failover-4768f8330cda">ExternalDNS on Amazon EKS: Automating Route 53 Records from Ingress to Multi-Region Failover</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What’s New in Kubernetes 1.34: A Hands-On Guide for DevOps]]></title>
            <link>https://aws.plainenglish.io/whats-new-in-kubernetes-1-34-a-hands-on-guide-for-devops-98f659e322aa?source=rss-e08a82d83e2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/98f659e322aa</guid>
            <category><![CDATA[container-orchestration]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[microservices]]></category>
            <dc:creator><![CDATA[The Agentic Builder]]></dc:creator>
            <pubDate>Mon, 08 Sep 2025 00:24:37 GMT</pubDate>
            <atom:updated>2025-09-27T18:27:11.918Z</atom:updated>
            <content:encoded><![CDATA[<p>Kubernetes 1.34 “Of Wind &amp; Will (O’ WaW)” was released on August 27, 2025. This release includes 58 enhancements: 23 have reached stable, 22 moved to beta, and 13 entered alpha. Key highlights are the General Availability of Dynamic Resource Allocation (DRA), the introduction of per-container restart rules in alpha, and the stabilization of the Job pod replacement policy.</p><p>This article provides a hands-on overview of the most important features for Kubernetes Administrators.</p><ol><li><strong>Dynamic Resource Allocation (DRA) — GA</strong></li></ol><p>The core of DRA is now generally available. It allows workloads to dynamically request specialized hardware such as GPUs, NPUs, or FPGAs.</p><p>Instead of relying on static device plugins or node affinities, applications use a DeviceClass and ResourceClaim, then attach the claim to a Pod.</p><pre># 1) Cluster-scoped DeviceClass<br>apiVersion: resource.k8s.io/v1<br>kind: DeviceClass<br>metadata:<br>  name: nvidia-gpu-a100<br>spec:<br>  selectors:<br>    - cel:<br>        expression: has(device.vendor) &amp;&amp; device.vendor == &quot;nvidia&quot; &amp;&amp; device.model == &quot;a100&quot;<br><br>---<br># 2) Namespaced ResourceClaim<br>apiVersion: resource.k8s.io/v1<br>kind: ResourceClaim<br>metadata:<br>  name: my-a100-claim<br>spec:<br>  devices:<br>    requests:<br>      - name: gpu0<br>        deviceClassName: nvidia-gpu-a100<br><br>---<br># 3) Pod consumes the claim<br>apiVersion: v1<br>kind: Pod<br>metadata:<br>  name: ml-trainer<br>spec:<br>  resourceClaims:<br>    - name: gpu<br>      resourceClaimName: my-a100-claim<br>  containers:<br>    - name: trainer<br>      image: ghcr.io/example/ml:latest<br>      resources:<br>        requests:<br>          cpu: &quot;2&quot;<br>          memory: &quot;8Gi&quot;</pre><p>DRA is enabled by default using the resource.k8s.io/v1 API.</p><p><strong>2. Per-Container Restart Rules — Alpha</strong></p><p>In previous releases, the restartPolicy was defined at the Pod level and applied to all containers. Kubernetes 1.34 introduces per-container restart rules, which also support exit-code — based behavior.</p><pre>apiVersion: v1<br>kind: Pod<br>metadata:<br>  name: per-container-restart-demo<br>spec:<br>  restartPolicy: OnFailure<br>  containers:<br>    - name: app<br>      image: ghcr.io/example/api:1.0<br>    - name: log-shipper<br>      image: ghcr.io/example/shipper:2.1<br>      restartPolicy: Always<br>      restartPolicyRules:<br>        - exitCodes: [137, 143]<br>          restartPolicy: OnFailure</pre><p>This feature is behind the ContainerRestartRules feature gate and is currently alpha.</p><p><strong>3. Job Pod Replacement Policy — GA</strong></p><p>Kubernetes 1.34 makes the pod replacement policy for Jobs generally available. This allows you to control when new Pods are created while old ones are terminating.</p><pre>apiVersion: batch/v1<br>kind: Job<br>metadata:<br>  name: job-replacement-policy-demo<br>spec:<br>  completions: 3<br>  parallelism: 1<br>  podReplacementPolicy: TerminatingElapsed<br>  template:<br>    spec:<br>      restartPolicy: Never<br>      containers:<br>        - name: worker<br>          image: ghcr.io/example/worker:1.0</pre><p>•	TerminatingStart: A replacement Pod is created as soon as the old one starts terminating.</p><p>•	TerminatingElapsed: A replacement Pod is created only after the old one has completely terminated.</p><p><strong>4. ServiceAccount Tokens for Image Pulls — Beta</strong></p><p>Kubelet can now use ServiceAccount tokens for pulling images from private registries, reducing the need for long-lived imagePullSecrets.</p><pre>apiVersion: v1<br>kind: ServiceAccount<br>metadata:<br>  name: puller<br>---<br>apiVersion: v1<br>kind: Pod<br>metadata:<br>  name: sa-pull-demo<br>spec:<br>  serviceAccountName: puller<br>  containers:<br>    - name: app<br>      image: private.registry.example.com/team/app:1.0</pre><p>Configure the kubelet credential provider to exchange tokens with your registry.</p><p><strong>5. CPU Manager Optimization — Beta</strong></p><p>For CPUs with split uncore caches, Kubernetes can now align CPU allocations to improve performance and reduce cross-cache traffic.</p><pre>apiVersion: kubelet.config.k8s.io/v1beta1<br>kind: KubeletConfiguration<br>cpuManagerPolicy: static<br>cpuManagerPolicyOptions:<br>  prefer-align-cpus-by-uncorecache: &quot;true&quot;</pre><p><strong>6. Resource Resize Policies — Refresher</strong></p><p>Resize policies enable fine-grained control over whether container resource adjustments require a restart.</p><pre>apiVersion: v1<br>kind: Pod<br>metadata:<br>  name: resize-policy-demo<br>spec:<br>  containers:<br>    - name: web<br>      image: nginx<br>      resources:<br>        requests: { cpu: &quot;250m&quot;, memory: &quot;256Mi&quot; }<br>        limits:   { cpu: &quot;1&quot;,    memory: &quot;512Mi&quot; }<br>      resizePolicy:<br>        - resourceName: cpu<br>          restartPolicy: NotRequired<br>        - resourceName: memory<br>          restartPolicy: RestartContainer</pre><p><strong>Key Takeaways for Cluster Admin Teams</strong></p><p>•	<strong>GPU and accelerator clusters: </strong>Move to Dynamic Resource Allocation for cleaner scheduling.</p><p>•	<strong>Workloads with sidecars: </strong>Experiment with per-container restart rules (alpha).</p><p>•	<strong>Batch processing pipelines:</strong> Adopt Job pod replacement policies for predictable scheduling.</p><p>•	<strong>Private registries: </strong>Transition to ServiceAccount token — based pulls.</p><p>•	<strong>Latency-sensitive workloads:</strong> Enable CPU manager cache alignment options.</p><p><strong>Conclusion</strong></p><p>Kubernetes 1.34 focuses on giving operators more control, flexibility, and efficiency. With DRA now generally available, per-container restarts in alpha, and Job replacement policies reaching stability, DevOps teams gain practical tools to reduce downtime and streamline operations.</p><p>Testing these features in staging environments will help organizations prepare for production adoption as they stabilize in future releases.</p><p>And before you go, don’t forget to clap and follow the writer️! ❤️</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98f659e322aa" width="1" height="1" alt=""><hr><p><a href="https://aws.plainenglish.io/whats-new-in-kubernetes-1-34-a-hands-on-guide-for-devops-98f659e322aa">What’s New in Kubernetes 1.34: A Hands-On Guide for DevOps</a> was originally published in <a href="https://aws.plainenglish.io/">AWS in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>