<?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 Mohan Sidda on Medium]]></title>
        <description><![CDATA[Stories by Mohan Sidda on Medium]]></description>
        <link>https://medium.com/@mohan.sidda?source=rss-50e94397d0c6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vWPcW3GrVxDy3bXvYY4N1w.png</url>
            <title>Stories by Mohan Sidda on Medium</title>
            <link>https://medium.com/@mohan.sidda?source=rss-50e94397d0c6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 07:37:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mohan.sidda/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[Why do we need vLLM and llm-d in AI Infrastructure?]]></title>
            <link>https://medium.com/@mohan.sidda/why-do-we-need-vllm-and-llm-d-in-ai-infrastructure-9d5ba81976ab?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d5ba81976ab</guid>
            <category><![CDATA[dllm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vllm]]></category>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 08:35:22 GMT</pubDate>
            <atom:updated>2026-09-15T00:54:29.109Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RLfJg8-KKldsrGZ0Nbp1uA.png" /></figure><h3>We spent decades making compute faster. Now we need to make AI infrastructure smarter.</h3><p>The AI infrastructure conversation often starts with GPUs.</p><p>How many GPUs do I need?</p><p>Which GPU should I buy?</p><p>How much GPU memory do I need?</p><p>But there is another question that is becoming equally important:</p><p>How intelligently are we using those GPUs?</p><p>That question takes us on an interesting journey from CPUs and GPUs, to LLM inference, to the limitations of traditional load balancing, and finally to technologies such as <a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> and <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> that are changing how we think about AI infrastructure</p><h3>1. The GPU changed the economics of AI</h3><p>For decades, the CPU was the workhorse of computing. But AI workloads introduced a different kind of problem. Modern neural networks perform enormous numbers of mathematical operations, particularly matrix and vector operations, that can be executed in parallel. And this is where GPUs changed the game.</p><p>GPUs were originally designed to accelerate graphics. But their ability to perform many mathematical operations in parallel made them extremely well suited to machine learning.</p><p>Instead of asking one powerful processor to execute a relatively small number of operations sequentially, GPUs can execute huge numbers of similar operations concurrently. That parallelism became fundamental to modern AI. Training large models required enormous amounts of compute.</p><p>And eventually, inference actually serving those models to users became a major infrastructure challenge of its own.</p><p>But there is an important distinction:</p><p>Training a model and serving a model are not the same problem.</p><p>Training is typically a massive computational job.</p><p>Inference is a real-time serving problem.</p><p>And inference introduces a new dimension:</p><p>Every user request is different.</p><h3>2. LLM inference is not just “send a request to a GPU”</h3><p>Consider two users.</p><p>User A asks:</p><p>“Summarize this paragraph.”</p><p>User B asks an AI agent to analyse a large collection of documents, reason through several steps and generate a long answer. From the perspective of a traditional load balancer, these may both look like:</p><p>1 request</p><p>But from the perspective of an AI inference system, they can be radically different workloads.</p><p>One might require relatively little computation.</p><p>The other might consume substantial GPU compute, memory bandwidth and KV-cache capacity for much longer.</p><p>This is where traditional load-balancing approaches start to show their limitations.</p><h3>3. The problem with traditional load balancing</h3><p>Traditional load balancers were designed around applications where servers can often be treated as relatively interchangeable.</p><p>Requests arrive. The load balancer selects an available backend. Algorithms such as round-robin or least-connections distribute the requests.</p><p>That model makes sense for many conventional applications. But LLM inference has something traditional applications generally don’t have:</p><p>state that matters to the cost of processing the next request.</p><p>One of the most important examples is the KV cache.</p><p>During LLM inference, the model generates and stores key/value information associated with the tokens it has processed. Reusing this cached information can avoid repeating expensive work.</p><p>Now imagine that a conversation has already been processed by GPU A. The next request arrives. A conventional load balancer may simply send it to GPU B because GPU B appears available.</p><p>GPU B may not have the useful cached context. The system may therefore have to perform work that GPU A could have avoided.</p><p>The problem isn’t that the load balancer is “wrong.”</p><p>The problem is that the load balancer doesn’t have enough knowledge about the workload to make an inference-aware decision.</p><p>And this distinction is important.</p><p>AI traffic isn’t just about:</p><p>“Which server has capacity?”</p><p>It is increasingly about:</p><p>“Which accelerator is the best place to process this particular request?”</p><p>That is a much harder problem.</p><h3>4. There is another complication: Prefill and Decode</h3><p>LLM inference itself has two fundamentally different phases:</p><h3>Prefill</h3><p>The model processes the input prompt.</p><p>This phase processes many input tokens and is generally compute-intensive.</p><h3>Decode</h3><p>The model generates the response, typically one token at a time.</p><p>This phase is generally much more memory-bandwidth sensitive.</p><p>So we have another interesting problem.</p><p>The same GPU infrastructure is being asked to handle workloads with very different characteristics.</p><p>Imagine running a data center where some GPUs are primarily dealing with compute-heavy workloads while others are spending their time serving long-running, memory-intensive token generation.</p><p>If we treat everything as simply:</p><p>Request → GPU</p><p>we are throwing away a lot of information.</p><p>The infrastructure needs to understand what is actually happening inside the inference workload.</p><h3>5. Significance of vLLM</h3><p>This is where <a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> becomes important.</p><p><a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> is an open-source inference engine designed specifically to make LLM serving more efficient.</p><p>One of its foundational innovations was <a href="https://www.redhat.com/en/topics/ai/what-is-vllm#vllm-and-pagedattention">PagedAttention</a>, which applies ideas inspired by virtual memory to the management of <a href="https://www.redhat.com/en/topics/ai/what-is-vllm#vllm-and-pagedattention">KV cache</a>.</p><p>Instead of treating GPU memory as one large, rigid allocation, <a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> can manage KV-cache blocks much more efficiently. It also supports techniques such as continuous batching, allowing new requests to join ongoing inference work rather than waiting for a traditional static batch to finish.</p><p>The result?</p><p>Better utilization of expensive accelerator resources. More efficient memory management.</p><p>Higher concurrency. And a much more sophisticated inference runtime.</p><p><a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> solves an important part of the problem the inference engine itself.</p><p>What happens when we have hundreds or thousands of GPUs?</p><p>That’s where the next layer becomes interesting.</p><h3>6. From inference engine to distributed inference</h3><p>Imagine a data center with hundreds of GPUs.</p><p>You have multiple <a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> instances. Some are processing prompts.</p><p>Some are generating tokens. Some have valuable cached context.</p><p>Some are heavily loaded. Some have just become available.</p><p>And requests keep arriving with wildly different characteristics.</p><p>Now the problem isn’t simply:</p><p>“How do I run an LLM efficiently on this GPU?”</p><p>The problem becomes:</p><p>“How do I intelligently coordinate the entire fleet of GPUs?”</p><p>This is the distributed inference problem.</p><p>And this is where <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> plays a significant role.</p><h3>7. llm-d: Making the infrastructure AI-aware</h3><p>A useful way to think about <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> is:</p><p><a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> is the inference engine. <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> is the distributed intelligence around the inference fleet.</p><p><a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> is a Kubernetes-native distributed inference stack that works with model servers such as <a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> and adds inference-aware orchestration and routing.</p><p>Its router can make decisions based on information that conventional load balancing doesn’t normally understand, including workload characteristics, cached prefixes and the state of inference servers.</p><p>Instead of blindly asking:</p><p>“Which backend should receive the next request?”</p><p>the system can ask a much more useful question:</p><p>“Which inference instance is the best place to process this request?”</p><p>That is a fundamental change.</p><h3>8. Think of llm-d as an air-traffic controller</h3><p>This is the analogy I find most useful.</p><p>Imagine an airport. A traditional traffic controller might know:</p><ul><li>which runway is available</li><li>which aircraft is arriving</li><li>which gate is free</li></ul><p>But AI infrastructure needs much more context.</p><p>One “aircraft” might be a tiny request. Another might be a massive reasoning workload that takes considerably longer.</p><p>Some GPUs may already have the context needed for a request.</p><p>Some may be processing prefill. Others may be busy with decode.</p><p>Some may have available capacity but the wrong cache locality.</p><p>So the AI infrastructure needs to understand the state of the entire inference environment.</p><p>That’s where <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a>’s intelligent routing becomes powerful.</p><p>It can work with inference-aware gateways and routing mechanisms to make decisions across model-server replicas rather than treating every backend as identical.</p><h3>9. From “request count” to “workload awareness”</h3><p>This is perhaps the biggest conceptual shift.</p><p>Traditional load balancing might see:</p><p>GPU A → 10 requests GPU B → 10 requests GPU C → 10 requests</p><p>Looks load is balanced.</p><p>But what if:</p><ul><li>GPU A has ten tiny requests</li><li>GPU B has ten huge reasoning requests</li><li>GPU C already has the KV cache for the next incoming conversation</li></ul><p>Those three GPUs are absolutely not equivalent.</p><p>AI infrastructure therefore needs to move beyond counting requests.</p><p>It needs to understand the shape of the workload.</p><p>That means considering things such as:</p><ul><li>prompt length</li><li>expected output length</li><li>KV-cache locality</li><li>current inference load</li><li>prefill versus decode characteristics</li><li>latency objectives</li><li>available accelerator resources</li></ul><p>This is the difference between load balancing and intelligent inference routing.</p><h3>10. Disaggregating Prefill and Decode</h3><p>Now take the idea one step further.</p><p>If prefill and decode have different performance characteristics, why should we necessarily run them on the same resources?</p><p>With disaggregated inference, the infrastructure can separate these phases.</p><p>Think of it as two specialized teams.</p><h3>Prefill workers</h3><p>Optimized for processing incoming prompts and generating the initial KV cache.</p><h3>Decode workers</h3><p>Optimized for the repeated token-generation phase.</p><p>The two can then be scaled and managed according to their respective workloads.</p><p><a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> supports orchestration patterns for disaggregated serving, including coordination between specialized prefill and decode workers.</p><p>This creates another important benefit:</p><p>Infrastructure can be matched more closely to workload characteristics.</p><p>Instead of building one giant pool of identical resources and hoping a generic load balancer distributes the work efficiently, we can start designing the infrastructure around the actual behaviour of AI workloads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DGTIkL9hrmF0QnSM" /><figcaption>image courtesy: Red Hat (from blog: <a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it)">https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it)</a></figcaption></figure><p>Important aspect here is an AI-aware gateway doesn’t mean traditional networking disappears.</p><p>Instead, these technologies work together.</p><p>Current llm-d architecture explicitly builds on existing Gateway API and proxy technologies while adding inference-aware routing through its router and endpoint-picker components.</p><h3>11. The bigger lesson: AI infrastructure is becoming workload-aware</h3><p>For decades, infrastructure was largely designed around the idea that:</p><p>compute is compute, and requests are requests.</p><p>AI is challenging that assumption. An AI request has context.</p><p>It has tokens. It has cache locality.</p><p>It has different compute and memory characteristics.</p><p>It may generate a handful of tokens or thousands.</p><p>And two requests that look identical from a network perspective can have completely different consequences for the underlying GPUs.</p><p>That means the infrastructure needs to understand more than networking.</p><p>It needs to understand inference.</p><h3>12. The future isn’t just more GPUs</h3><p>This is perhaps the most important conclusion. The answer to AI infrastructure isn’t simply:</p><p><em>“Buy more GPUs.”</em></p><p>The bigger opportunity is:</p><p><em>“Get more useful work from the GPUs you already have.”</em></p><p>That requires innovation at multiple layers:</p><p>Better accelerators → more compute</p><p>Better inference engines → more efficient execution</p><p>Better memory management → better KV-cache utilization</p><p>Better distributed inference → better utilization across the cluster</p><p>Better routing → send the right workload to the right resource</p><p>Better orchestration → coordinate everything as one system</p><p>This is why the evolution from GPU → vLLM → llm-d is so interesting.</p><p>We are moving from simply accelerating computation to intelligently orchestrating computation.</p><p>And that may ultimately be one of the biggest factors determining the economics of enterprise AI.</p><h3>13. The takeaway</h3><p>The AI infrastructure journey is no longer just about CPUs versus GPUs.</p><p>It is becoming a journey from:</p><p>General-purpose compute</p><p>→ Accelerated compute</p><p>→ Efficient inference</p><p>→ Distributed inference</p><p>→ Intelligent inference orchestration</p><p><a href="https://www.redhat.com/en/topics/ai/what-is-vllm">vLLM</a> makes the individual inference engine highly efficient.</p><p><a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">llm-d</a> extends that thinking across the distributed environment, bringing inference-aware routing, cache awareness, workload-aware scheduling and disaggregated serving into the picture.</p><p>And Kubernetes provides the foundation on which this distributed infrastructure can operate.</p><p>The real question for the next generation of AI infrastructure may therefore not be:</p><p>“How many GPUs do you have?”</p><p>It may be:</p><h3>“How intelligently can you use every GPU you have?”</h3><p><strong>Reference:</strong></p><ol><li><a href="https://llm-d.ai/docs/well-lit-paths">https://llm-d.ai/docs/well-lit-paths</a></li><li><a href="https://llm-d.ai/docs/well-lit-paths">https://llm-d.ai/</a> &amp; <a href="https://docs.vllm.ai/en/latest/">https://docs.vllm.ai/en/latest/</a></li><li><a href="https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it">https://www.redhat.com/en/blog/what-llm-d-and-why-do-we-need-it</a></li><li><a href="https://www.redhat.com/en/topics/ai/what-is-vllm">https://www.redhat.com/en/topics/ai/what-is-vllm</a></li><li><a href="https://www.redhat.com/en/topics/ai">https://www.redhat.com/en/topics/ai</a></li><li><a href="https://developers.redhat.com/articles/2026/06/11/intelligent-inference-scheduling-llm-d-red-hat-ai#">https://developers.redhat.com/articles/2026/06/11/intelligent-inference-scheduling-llm-d-red-hat-ai#</a></li><li><a href="https://github.com/llm-d/llm-d/tree/main/guides/optimized-baseline">https://github.com/llm-d/llm-d/tree/main/guides/optimized-baseline</a></li><li><a href="https://developers.redhat.com/articles/2025/10/07/master-kv-cache-aware-routing-llm-d-efficient-ai-inference#">https://developers.redhat.com/articles/2025/10/07/master-kv-cache-aware-routing-llm-d-efficient-ai-inference#</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d5ba81976ab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is KVM, and What is KubeVirt?]]></title>
            <link>https://medium.com/@mohan.sidda/what-is-kvm-and-what-is-kubevirt-10ed7c547690?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/10ed7c547690</guid>
            <category><![CDATA[red-hat-openshift]]></category>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[modernization]]></category>
            <category><![CDATA[virtualization]]></category>
            <category><![CDATA[migration]]></category>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Sun, 30 Mar 2025 23:20:03 GMT</pubDate>
            <atom:updated>2025-03-30T23:20:03.807Z</atom:updated>
            <content:encoded><![CDATA[<h3>What is KVM, and What is KubeVirt? The significance of both in the era of AI &amp; cloud-native development!</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/447/0*0slE79gsfLBcyGPA" /><figcaption>image courtesy: <a href="https://www.redhat.com/en">Red Hat</a></figcaption></figure><h3>Bridging the Old and the New: A Journey Through Infrastructure Modernization</h3><p>Today’s IT leaders face the dual challenge of embracing cloud-native technology and hybrid cloud strategies while optimizing cost, ensuring security, and integrating AI-driven workloads. A key hurdle is that many enterprise-critical workloads still run on virtual machines (VMs). While leadership is eager to modernize, adopt AI, strengthen security, and reduce costs, they cannot afford the disruptions of an abrupt shift from traditional <a href="https://www.redhat.com/en/topics/virtualization/what-is-virtualization"><strong>virtualization</strong></a> to containerized environments. This is where <a href="https://www.redhat.com/en/topics/virtualization/what-is-KVM"><strong>KVM</strong></a><strong> and </strong><a href="https://www.redhat.com/en/topics/virtualization/what-is-kubevirt#what-is-kubevirt"><strong>KubeVirt</strong></a> come into play.</p><h3>A Familiar Foundation: KVM</h3><p><strong>KVM (Kernel-based Virtual Machine) is a Linux-based, open-source hypervisor</strong> that has been powering enterprise-grade virtualization since 2007. It transforms the Linux kernel into a robust hypervisor, enabling <strong>near-native performance for VMs</strong>. With most enterprises relying on KVM-based virtualization technologies, it’s no wonder that KVM has become a trusted foundation for countless organizations worldwide.</p><p>While KVM looks solid, how do I bridge it with Kubernetes, which IT leaders are pushing for today?</p><h3>KubeVirt: The Missing Link</h3><p>As mentioned earlier, the missing link is clear, while IT leadership is eager to modernize by shifting away from legacy virtualization, enhancing security, reducing costs, and integrating AI-driven workloads, <a href="https://www.redhat.com/en/topics/virtualization/what-is-kubevirt#what-is-kubevirt"><strong>KubeVirt</strong></a> provides the perfect answer. <a href="https://www.redhat.com/en/topics/virtualization/what-is-kubevirt#what-is-kubevirt"><strong>KubeVirt</strong></a> extends Kubernetes, enabling VMs and containers to run together on a unified platform. It allows organizations to retain their existing VM-based applications while progressively adopting cloud-native architecture, all without the need to rewrite workloads from scratch</p><p>This is exactly what IT leaders need today! A way to modernize without throwing away our existing infrastructure.” According to <a href="https://www.forrester.com/blogs/predictions-2024-technology-infrastructure/"><strong>Forrester predictions</strong></a>, in 2024 20% of customers would have started their escape from legacy virtualization solutions“</p><p>Both KVM (Kernel-based Virtual Machine) and KubeVirt are open-source projects, with KVM originating from the<a href="https://linux-kvm.org/page/Main_Page"> Linux kernel</a> and <a href="https://kubevirt.io/">KubeVirt</a> being a project under the Cloud Native Computing Foundation (<a href="https://kubevirt.io/">CNCF</a>).</p><h3>In a way, it is safe to say that KubeVirt = Kubernetes + KVM?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/521/0*H0DTyI5YM5ACn2pS" /></figure><p>Yes, in a simplified way, you can think of <strong>KubeVirt as Kubernetes + KVM</strong>.</p><p>KVM provides the virtualization layer that allows Linux to act as a hypervisor for running VMs. Kubernetes is the container orchestration platform that manages workloads in a cloud-native way.</p><p>KubeVirt bridges the gap by enabling Kubernetes to orchestrate both VMs and containers on the same platform. So, KubeVirt = Kubernetes + KVM, but with additional capabilities like VM lifecycle management, networking, and storage integration within Kubernetes. It essentially brings <strong>traditional virtualization into the cloud-native world</strong>, allowing organizations to modernize at their own pace without disrupting existing VM-based applications.</p><h3>The Power of Open-Source and Red Hat’s Role</h3><p><a href="https://linux-kvm.org/page/Main_Page"><strong>KVM</strong></a><strong> and </strong><a href="https://kubevirt.io/"><strong>KubeVirt</strong></a><strong> are open-source projects</strong>, with KVM originating from the <a href="https://linux-kvm.org/page/Main_Page"><strong>Linux kernel</strong></a> and KubeVirt being a part of the <strong>Cloud Native Computing Foundation (</strong><a href="https://kubevirt.io/"><strong>CNCF</strong></a><strong>)</strong>. But what stands out is <strong>Red Hat’s significant role</strong> in their development. Check Red Hat’s position on the contributions to the KubeVirt <a href="https://kubevirt.devstats.cncf.io/d/5/companies-table?orgId=1&amp;var-period_name=Last%20decade&amp;var-metric=contributions">project here</a>.</p><p>Red Hat, a <strong>leader in open-source innovation</strong>, has been a <strong>major contributor</strong> to both projects, ensuring their stability, enterprise-readiness, and seamless integration. Their OpenShift platform, powered by <strong>OpenShift Virtualization (built on KubeVirt and KVM)</strong>, enables today’s businesses to run and manage <strong>VMs and containers in a single, flexible hybrid environment</strong>.</p><h3>How Red Hat KubeVirt is a game changer in the context of AI workloads</h3><p>Red Hat <strong>KubeVirt</strong> is redefining how enterprises manage AI workloads by bridging the gap between <strong>virtual machines (VMs) and containers</strong> within a Kubernetes-native environment. Here’s why it’s a game-changer for AI workloads:</p><ol><li><strong>Seamless AI/ML Model Deployment</strong></li></ol><p>AI workloads often require <strong>GPU acceleration, high-performance storage, and custom environments</strong>, which VMs traditionally provide. KubeVirt enables running AI/ML models in <strong>hybrid environments</strong>, where VMs and containers coexist, ensuring a <strong>smooth transition to cloud-native AI workloads</strong>.</p><p><strong>2. Hardware Acceleration for AI</strong></p><p>AI models demand <strong>GPU and TPU acceleration</strong>. KubeVirt leverages Kubernetes’ <strong>device plugins</strong> and <strong>Red Hat OpenShift’s GPU support</strong>, allowing AI workloads to use hardware acceleration efficiently, whether they run in <strong>VMs or containers</strong>.</p><p><strong>3. AI Infrastructure Modernization</strong></p><p>Many AI workloads still rely on <strong>legacy applications and traditional VM-based data pipelines</strong>. With KubeVirt, organizations can <strong>modernize AI infrastructure step by step</strong> by migrating workloads to Kubernetes <strong>without disrupting existing AI models and workflows</strong>.</p><p><strong>4. Unified AI Development &amp; Operations</strong></p><p>KubeVirt integrates AI workloads into a <strong>single, cloud-native platform</strong>, allowing <strong>data scientists, ML engineers, and IT teams</strong> to manage AI pipelines, model training, and inference in a <strong>Kubernetes-driven DevOps workflow</strong>.</p><p><strong>5. Flexibility for Hybrid &amp; Multi-Cloud AI</strong></p><p>AI workloads require <strong>scalability</strong> across different environments. KubeVirt ensures AI applications can run in <strong>on-premise, hybrid, and multi-cloud</strong> setups <strong>without vendor lock-in</strong>, making it easier to scale AI models dynamically.</p><p><strong>6. Enhanced Security &amp; Compliance</strong></p><p>AI workloads often involve <strong>sensitive data</strong>. KubeVirt inherits <strong>Red Hat OpenShift’s enterprise-grade security</strong>, ensuring AI models comply with regulatory requirements like <strong>HIPAA, GDPR, and ISO standards</strong>.</p><p><strong>7. Cost Optimization for AI Training &amp; Inference</strong></p><p>By consolidating AI workloads into a <strong>Kubernetes-native environment</strong>, Red Hat OpenShift helps organizations <strong>reduce licensing costs</strong>, optimize compute resource usage, and <strong>maximize infrastructure efficiency</strong>.</p><h3>Conclusion</h3><p>KVM, a robust, open-source hypervisor integrated into the Linux kernel, has been a cornerstone of virtualization since 2007. According to <a href="https://theirstack.com/en/technology/kvm"><strong>TheirStack</strong></a>, over <strong>7,500 Fortune 500 companies</strong> rely on KVM technology for their infrastructure.</p><p><strong>KubeVirt</strong>, introduced in 2016, has been<strong> bundled with Red Hat OpenShift</strong>, enabling enterprises to run VMs alongside containers in a Kubernetes-native environment. Since 2020, OpenShift Virtualization has been generally available, leveraging both <strong>KVM and KubeVirt</strong> to unify traditional and modern workloads.</p><p>With <strong>Red Hat OpenShift AI</strong>, enterprises can efficiently manage <strong>AI workloads</strong>, combining the strengths of <strong>virtualization and Kubernetes</strong> to drive <strong>next-generation AI innovation</strong> at scale</p><p>Is it time to consider migrating or modernizing your legacy VM estate with cloud-native virtualization? To enhance scalability, efficiency, and future readiness. Embrace the power of Kubernetes and KVM to streamline operations and drive innovation.</p><h3>Appendix<strong>:</strong></h3><ol><li><a href="https://www.redhat.com/en/technologies/cloud-computing/openshift/virtualization">https://www.redhat.com/en/technologies/cloud-computing/openshift/virtualization</a></li><li><a href="https://kubevirt.io/">https://kubevirt.io/</a></li><li><a href="https://kubevirt.devstats.cncf.io/d/5/companies-table?orgId=1&amp;var-period_name=Last%20decade&amp;var-metric=contributions">https://kubevirt.devstats.cncf.io/d/5/companies-table?orgId=1&amp;var-period_name=Last%20decade&amp;var-metric=contributions</a></li><li><a href="https://github.com/kubevirt/kubevirt">https://github.com/kubevirt/kubevirt</a></li><li><a href="https://linux-kvm.org/page/Main_Page">https://linux-kvm.org/page/Main_Page</a></li><li><a href="https://www.redhat.com/en/resources/15-reasons-adopt-openshift-virtualization-ebook">https://www.redhat.com/en/resources/15-reasons-adopt-openshift-virtualization-ebook</a></li><li><a href="https://www.redhat.com/en/blog/12-factor-app-containers">https://www.redhat.com/en/blog/12-factor-app-containers</a></li></ol><h3>Disclaimer</h3><p><em>The views and opinions expressed in this blog are solely those of the author and do not necessarily reflect the official policy or position of any organization, including Red Hat or its affiliates. The information provided is for </em><strong><em>informational and educational purposes only</em></strong><em> and should not be considered as professional or legal advice.</em></p><p><em>All product names, trademarks, and registered trademarks mentioned are the property of their respective owners. While every effort has been made to ensure accuracy, the author and publisher make no warranties or representations regarding the completeness, reliability, or timeliness of the information. Readers are encouraged to conduct their own research and consult with experts before making any business or technical decisions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10ed7c547690" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Significance of Hybrid Cloud in Digital Transformation Journey]]></title>
            <link>https://medium.com/@mohan.sidda/significance-of-hybrid-cloud-in-digital-transformation-journey-3060517c426b?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/3060517c426b</guid>
            <category><![CDATA[hybrid-cloud-computing]]></category>
            <category><![CDATA[hybrid-cloud]]></category>
            <category><![CDATA[hybrid-cloud-solutions]]></category>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Mon, 28 Nov 2022 03:31:06 GMT</pubDate>
            <atom:updated>2022-11-28T03:31:06.357Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/943/1*JOC4qRXr1sQ-j_mQAbXJ7w.png" /><figcaption>image credits to Red Hat(<a href="https://www.redhat.com/en/topics/cloud/open-hybrid-cloud-approach">https://www.redhat.com/en/topics/cloud/open-hybrid-cloud-approach</a>)</figcaption></figure><p>“Hybrid cloud” is a buzzword that has been making circles in the last two years. I want to put some thought into why it is gaining popularity and the importance of hybrid cloud in “digital transformation success.”</p><p>Let us begin with what a “Hybrid Cloud” means.</p><p>A hybrid cloud is a combination of private and public clouds. The private cloud can be one or more existing datacenters built on bare-metal (physical hardware) or a virtual environment. A public cloud is an IT model where computing services can be provided over the public internet by hyperscaler’s on-demand/pay-as-you-go model.</p><p>Let us take a look at how major IT vendors define the hybrid cloud.</p><p><a href="https://www.ibm.com/sg-en/cloud/learn/hybrid-cloud">IBM</a>: Hybrid cloud combines and unifies public cloud and private cloud services from multiple cloud vendors to create a single, flexible, cost-optimal IT infrastructure.</p><p><a href="https://azure.microsoft.com/en-us/resources/cloud-computing-dictionary/what-is-hybrid-cloud-computing/">Microsoft</a>: A hybrid cloud — sometimes called a cloud hybrid — is a computing environment that combines an on-premises datacenter (also called a private cloud) with a public cloud, allowing data and applications to be shared between them. Some people define hybrid cloud to include “multicloud” configurations where an organization uses more than one public cloud in addition to their on-premises datacenter.</p><p><a href="https://www.redhat.com/en/topics/cloud-computing/what-is-hybrid-cloud">Red Hat</a>: A hybrid cloud refers to a combination of at least two computing environments that share information with one another and run a uniform series of applications for a business or enterprise.</p><p><a href="https://cloud.google.com/learn/what-is-hybrid-cloud#:~:text=A%20hybrid%20cloud%20is%20one,entirely%20on%20the%20public%20cloud.">Google</a>: A hybrid cloud is one in which applications are running in a combination of different environments. Hybrid cloud computing approaches are widespread because almost no one today relies entirely on the public cloud.</p><p><a href="https://www.vmware.com/topics/glossary/content/hybrid-cloud.html">VMware</a>: Hybrid Cloud refers to a cloud computing model that uses a combination of at least one private cloud and at least one public cloud, which works together to provide a flexible mix of cloud computing services.</p><h3><strong>Why hybrid cloud makes sense in today’s digital economy</strong></h3><p>A hybrid cloud delivers a consistent platform across bare metal, virtual, private, public, and edge environments. A hybrid cloud unlocks a set of capabilities essential in today’s world, including speed, performance, agility, and portability. Let us have a quick look at some of the hybrid cloud advantages:</p><ol><li>Make use of the public cloud ecosystem, build rapid cloud-native applications, and meet the everchanging infrastructure demands to gain speed and agility.</li><li>Define the hyperscaler’s affinity based on the workload(ex: AI/ML workloads will be deployed in the ‘X’ public cloud).</li><li>Avoid vendor lock-in.</li><li>Create a healthy competitive pricing advantage.</li><li>Build your own disaster recovery strategy where there are no multiple hyperscalers regions in the same country.</li></ol><p>How does the “Hybrid Cloud” make sense in today’s era? When it comes to unique customer requirements, I see three approaches here:</p><p>1. The first type of customers are starting the cloud adoption in Dev and Testing environments in the cloud. However, they continue to leverage their on-prem datacenter for production environments. The driving factors for this approach are the customers who want to thrive in cloud-native development and faster innovation to meet the ever-changing business demands. On the other hand, the customers have very stringent security requirements to comply with their internal compliance. Thus, they are choosing Uat &amp; Prd in their on-premises.</p><p>2. The second type of customer wants the opposite of this approach. They want to continue the Dev and Testing on their on-premises Uat and Prd in the public cloud. This approach minimizes the CapEX (heavily invested in on-prim DC), and simultaneously drives the auto-scaling, enriched day two operations, performance, applications availability, and stability. This approach will help to gain more value from your IT investments (CapEx).</p><p>3. The third type of customers with legacy applications that can’t be migrated to the public cloud due to incompatibility challenges will use hybrid cloud as an alternative approach to continue these legacy workloads on-premises until these applications are migrated to cloud-native or repurchased as SaaS services in the future. These organizations will adopt public cloud services for building modern cloud-native applications or migrate compatible workloads.</p><h3><strong>Conclusion</strong>:</h3><p>Hybrid cloud architecture is a strategy that brings speed, performance, agility, portability, and variable expense(avoids huge CapEx) to the table. However, you must also choose the right turnkey application platform to complement a hybrid cloud strategy. The turnkey application platform will add the following advantages and pave the path to success in a hybrid cloud strategy.</p><ol><li>Unified experience across private and public clouds.</li><li>Workload portability between private &amp; public clouds.</li><li>Cloud Adoption — Inconsistency Challenges.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3060517c426b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Essence of Event-Driven Architecture (EDA)]]></title>
            <link>https://medium.com/@mohan.sidda/the-essence-of-event-driven-architecture-eda-6bd87f86813d?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/6bd87f86813d</guid>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Sun, 29 Aug 2021 02:52:17 GMT</pubDate>
            <atom:updated>2021-08-29T02:52:17.114Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cNmECa-_RIRXQRfofXILuQ.jpeg" /><figcaption>(image credit: Ian Dooley. <a href="https://unsplash.com/images)">https://unsplash.com/images)</a></figcaption></figure><h3>What is EDA?</h3><p>Event-driven Architecture (EDA) is a software architecture pattern to implement the publishing and consuming of events. In other words, the EDA design paradigm helps to decouple the applications to communicate asynchronously.</p><p><strong>EDA is not software or a product. It is an architectural style to implement loosely coupled, event-based, asynchronous application integration in an enterprise.</strong></p><p>An event represents an action of significant interest. Often, events correspond to a creation or a change of state of some entity. For example, “Order Raised” is an event in an e-commerce application, as is “Product Dispatched”. “Review Submitted” is an event by a customer submitting a review for a received product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1013/0*yLBozpS-OX71NsQf" /></figure><h3>EDA vs CEP</h3><p>Now that we’ve seen what EDA is, let’s consider CEP.</p><p>CEP (Complex Event Processing) correlates multiple events within a given time frame and makes meaningful action as necessary.</p><p>An example can be “Fraud Detection.” CEP can be implemented for fraud detection in ATM withdrawals. When there are more than three withdrawals within a min, there can be automated SMS/call initiated to the account holder, or the account can be disabled for an hour ..etc. Actions can be implemented correlating multiple withdrawals(messages) within a given time frame and act based on the requirement/use case.</p><h3>Why do we need EDA?</h3><p>You may be wondering: “Does each microservice need to communicate synchronously? Or can this mindset can be changed and implemented in asynchronously adopting the EDA? As long as the enterprise architect knows which applications &amp; microservices produce and consume what events, can we just choreograph using a publish/subscribe event broker to decouple these microservices?”</p><p>EDA is not a new architecture style. EDA has been implemented by most of the capital markets trading platforms for more than a decade. Today, EDA is being adopted by more enterprises irrespective of the domain(Banking, Telco, Retail…etc.).</p><p>In the current era, enterprises starting the digital transformation and application modernization are trying to adopt the event-first mindset. The reason is their existing service-oriented architecture (SOA); extract, transform and load (ETL); and batch-based approaches need to evolve to meet real-time needs and business demands. There are limitations in hub-spoke or Enterprise Service Bus (ESB) architectures when it comes to scalability.</p><p>These architecture styles and lack of application compatibility restricted our mindset to having tightly coupled request/reply(synchronous) interactions, which blocks scalability, agility, and cloud-native application development.</p><p>In most cases, microservice need not communicate synchronously! Our mindset needs to be changed!! We should be open to decoupling these microservices using EDA to gain scalability, agility, and less error-prone in microservices adoption.</p><h3>When to use EDA:</h3><ol><li>Opaque consumers. Cases where producers are generally unaware of consumers. Typically trading platforms..etc.</li><li>High fan-out. Scenarios where multiple, diverse consumers might process one event.</li><li>Outbox Pattern: often microservices, updating their own data store and need to notify other services about data changes that happened. The <a href="https://debezium.io/blog/2019/02/19/reliable-microservices-data-exchange-with-the-outbox-pattern/">outbox pattern </a>describes an approach for letting services execute these two tasks safely and consistently.</li><li>Stream Processing and Complex pattern matching: Enables continuous, real-time applications built to react to, process, or transform streams. Streams of events processed (map, join, aggregate) match the patterns and execute the action for designed business needs.</li><li>Event-Sourcing: <a href="https://martinfowler.com/eaaDev/EventSourcing.html">Event Sourcing</a> ensures that all changes to the application state are stored as a sequence of events. Not only can we query these events, but we can also use the event log to reconstruct past states and as a foundation to adjust the state to cope with retroactive changes automatically.</li><li>Command-query responsibility segregation(CQRS). <a href="https://martinfowler.com/bliki/CQRS.html">CQRS</a> is a pattern that separates read and update operations for a data store. CQRS plays a significant role in today’s microservices adoption when business demands high throughput, ensuring that the database is not the bottleneck where multiple reads and writes happen simultaneously. Implementing CQRS can improve the scalability, modularity, and resilience of applications, with some consistency trade-offs. This pattern is commonly associated with EDA.</li></ol><h3>Merits and demerits of the EDA:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/639/1*RkeMRObveTUGPq76YAx7Ag.png" /><figcaption>Merits and Demerits. Ref: <a href="https://www.redhat.com/en/blog/container-tidbits-does-pets-vs-cattle-analogy-still-apply">cattle and not pets</a></figcaption></figure><h3>What problems we are trying to solve with EDA:</h3><p>By now, you pretty much know what problems EDA is trying to solve.</p><p>Tightly-coupled, synchronous (request/reply) integration, hub-spoke, ESB integrations, ETL and batch processing, has the limitation in scalability, agility, and real-time demands.</p><p>In the microservices adoption journey, <a href="https://12factor.net/">The Twelve-Factor</a> app methodology is essential to succeed. Similarly, we need to adopt the EDA to gain more advantages in microservices adoption. Enterprise architect’s mindset needs to be changed from tightly coupled integration(Synchronous) to Asynchronous/loosely coupled integration and use events, CEP, and streaming.</p><p>EDA is an effective tool for reducing coupling between the components of a system by modeling interactions using the concepts of producers, consumers, events, and streams. EDA may not be the silver bullet to solve all the problems. But it is worth considering loose coupling and asynchronous communications in today’s microservices architecture. EDA is an essential part of any successful microservices adoption.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6bd87f86813d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[RH SSO federating two data sources — LDAP and Informix Database]]></title>
            <link>https://medium.com/@mohan.sidda/rh-sso-federating-two-data-sources-ldap-and-informix-database-69036ea99778?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/69036ea99778</guid>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Fri, 03 May 2019 13:16:07 GMT</pubDate>
            <atom:updated>2019-05-03T13:16:07.212Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>RH SSO federating two data sources — LDAP and Informix Database</strong></h3><p>I am delighted to share the experience of implementing RHSSO federating two data sources at one of our insurance customer place.</p><p>The customer is storing the external agent’s profile, and credentials in the Informix database and all internal users profile in active directory. i.e. Identity Access Management has two data sources.</p><p>Apparently, the web services exposed outside the enterprise are authenticated using the credentials stored in the database. The customer has a complex legacy — Policy Administration Services (PAS). Since there is no standardized ESB to integrate their northbound and southbound systems, the customer initiated the API centric integration implementation program.</p><p>As part of the API centric integration and container adaptation initiative, the customer is exposing their newly adopting insurance domain driven microservices in 3Scale API Gateway. The API’s exposed in 3Scale Gateway has to be secured by RH-SSO.</p><p>RH-SSO will be federating these two (Informix and Active Directory) data sources. The reference architecture looks like :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R82KWtM8AkLt354t" /></figure><p>To Integrate the RH-SSO with Informix Database, we used upstream (KeyCloak) JPA code. Ref: <a href="https://github.com/keycloak/keycloak/tree/openshift-integration/examples/providers/user-storage-jpa">https://github.com/keycloak/keycloak/tree/openshift-integration/examples/providers/user-storage-jpa</a></p><p>The upstream uses JPA and connects to the h2 database. However, we need to enhance this code to fulfill the following requirement:</p><ol><li>Getting the key from the database to decrypt the password.</li><li>Many filter conditions that need to be applied when we are validating the user credentials in the database</li><li>The database only returns boolean when the user authenticated in the database.</li></ol><p>To fulfill the above requirements, we used the stored procedure instead of JPA.</p><p>With updated JPA code and along with RH-SSO 7.1 image (registry.access.redhat.com/redhat-sso-7/sso71-openshift:1.1) we build the custom docker image (Ref: <a href="https://github.com/mohansidda/SSO-SPI">https://github.com/mohansidda/SSO-SPI</a>) and deployed the custom image in OpenShift.</p><p>Basically, the Docker file is used to build the custom image. It does the following tasks:</p><ol><li>Docker file contains the RH SSO (7.1) image,</li><li>SPI jars, modules, updated (added Informix data source) standalone-openshift.xml</li><li>Changing the permissions for few directories for JBoss user.</li></ol><p>Now the API’s are exposed in 3Scale and secured through RH-SSO. The SSO is federating two data sources as shown in the above diagram.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69036ea99778" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Red Hat — AMQ Cluster Deployment in OpenShift(OCP)]]></title>
            <link>https://medium.com/@mohan.sidda/red-hat-amq-cluster-deployment-in-openshift-ocp-5af7dc0f4c50?source=rss-50e94397d0c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/5af7dc0f4c50</guid>
            <dc:creator><![CDATA[Mohan Sidda]]></dc:creator>
            <pubDate>Fri, 03 May 2019 07:39:03 GMT</pubDate>
            <atom:updated>2019-05-03T07:39:03.650Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Red Hat — AMQ Cluster Deployment in OpenShift(OCP)</strong></p><p>In this article, I’ve shown a comparison between traditional broker topologies with the broker cluster provisioned in OCP. In general, the customers used to provisioning a broker cluster in HA / HA-LB topology in VM’s or bare-metal. I am sharing my experience with both the traditional approach as well as deploying in Open Shift (OCP).</p><p><strong>Traditional broker topologies:</strong></p><p><strong>HA Cluster — Scenario 1:</strong> <strong>Two brokers deployed in two different VM’s using the same shared storage to support the high availability (HA).</strong> Always the active/master broker will acquire the file lock. Whenever the active/master broker goes down, the slave/passive broker will become the active broker and acquire the lock.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZWpD252OkECMyr4P" /></figure><p><strong>HA &amp; LB Cluster- Scenario 2: </strong>Provisioning two HA(High Availability) broker groups together as one LB(Load Balance) broker group in two different VM’s.</p><p><strong>HA group A</strong>: Broker-A1 (master) deployed in VM 1 and Broker -A2(slave) deployed in VM 2 to support HA.</p><p><strong>HA group B</strong>: Broker-B1 (master) deployed in VM 2 and Broker -B2(slave) deployed in VM 1 to support HA.</p><p>Always the active/master broker will acquire the file lock. Whenever the active/master brokers go down, the slave/passive broker will become the active broker and acquire the lock.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x6mYZ5wANP-GTsUY" /></figure><p>In these topologies, adding a new broker to the cluster is a very tedious process. With some of the message broker vendors, the shared topology is not even supported. Some products you can add the third broker in only in share-nothing topology.</p><p><strong>Deploying the AMQ Cluster in Openshift/Kubernetes</strong> :</p><p>I am delighted to share my experience provisioning AMQ Cluster in Openshift using “Share Nothing (or) Network Replication” topology.</p><p><strong>Why AMQ cluster in Openshift/Kubernetes cluster?</strong></p><p>Though the customer is looking for the traditional above (HA &amp; LB) cluster in VM”s, we listed the pros and cons of AMQ provisioning inside OCP and Outside OCP.</p><p><strong>The influential pros</strong> :</p><ol><li>Since all the services which are going to use AMQ cluster will be deployed in the OCP cluster. It makes sense to provisioning the AMQ cluster inside OCP to avoid an extra hop.</li><li><strong>Auto Scaling </strong>— One of the NFR(Non-Functional Requirement) is when the CPU utilization reaches to 60% the customer would like to auto scale one more brokers. This requirement can be implemented using the OpenShift/Kubernetes out of the box.</li><li>The AMQ Operator will make sure that minimum broker replicas are always maintained.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*er2OS6ak2vNm1AzP" /></figure><p><strong>Scaledown Scenario — what happens if the message is undelivered due to the offline/slow consumer?</strong></p><p>When the CPU usage goes above 60% utilization, auto-scaling is helping to scale up the new brokers. What happens when the CPU utilization goes down? OCP automatically scales down and keep the AMQ cluster state as per the minimum replicas definition.</p><p>Everything looks awesome !! Think of a corner case, where the particular queue has slow/offline consumer, and there will be &lt;N&gt; messages pile-up all the time in this specific queue. Adding fuel to the fire, In this scenario, the OpenShift controller has trigged the scale-down trigger where the particular queue is residing, and &lt;N&gt; messages are undelivered.</p><p>Well, that’s the absolute time the Drainer(scaledown controller) comes into action !!</p><p>When the scale down process gets triggered if the message is undelivered/stuck in the persistent volume (PV) the “Drainer/scaledown controller” will move the undelivered messages to another live broker in the cluster. That’s how the message guarantee is handled in Openshift — AMQ cluster.</p><p><strong>What is Drainer POD (or) Scaledown Controller?</strong></p><p>Drainer POD will come in action only when the AMQ broker killed/deleted or whenever if there is a broker cluster scale down scenario occurs. In these, cases if there is an undelivered/orphaned message, the Drainer POD will take care of redistribution of the message.</p><p>POD draining is the method used to manage the integrity of messaging data in a Kubernetes StatefulSet. POD draining is used for message migration, which refers to the removal and redistribution of “orphaned” messages from a persistent volume, due to broker POD failure or intentional scale down.</p><p>Find more about Drainer/scaledown controller: <a href="https://access.redhat.com/documentation/en-us/red_hat_amq/7.2/html-single/deploying_amq_broker_on_openshift_container_platform/index#what_is_pod_draining-broker-ocp">https://access.redhat.com/documentation/en-us/red_hat_amq/7.2/html-single/deploying_amq_broker_on_openshift_container_platform/index#what_is_pod_draining-broker-ocp</a></p><p><strong>AMQ Cluster Deployment :</strong></p><p>isn’t it awesome to bring-up the AMQ cluster(#CloudNative, #ContainerAddaption) in Kubernetes/OpenShift cluster by applying few yaml files?</p><p>Excited to deploy AMQ in OpenShift/Kubernetes cluster? Follow the Red Hat documentation :</p><p><a href="https://access.redhat.com/documentation/en-us/red_hat_amq/7.2/html-single/deploying_amq_broker_on_openshift_container_platform/index#install-deploy-ocp-broker-ocp">https://access.redhat.com/documentation/en-us/red_hat_amq/7.2/html-single/deploying_amq_broker_on_openshift_container_platform/index#install-deploy-ocp-broker-ocp</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5af7dc0f4c50" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>