<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Parth Shah</title>
    <description>The latest articles on DEV Community by Parth Shah (@parths2019).</description>
    <link>https://dev.to/parths2019</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4248667%2F8a748042-fb73-4d8e-afac-931506e5a271.png</url>
      <title>DEV Community: Parth Shah</title>
      <link>https://dev.to/parths2019</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/parths2019"/>
    <language>en</language>
    <item>
      <title>Cluster Autoscaler vs. Karpenter on Amazon EKS: Seven Claims Tested, and What Agent Workloads Change</title>
      <dc:creator>Parth Shah</dc:creator>
      <pubDate>Wed, 18 Nov 2026 21:57:00 +0000</pubDate>
      <link>https://dev.to/parths2019/cluster-autoscaler-vs-karpenter-on-amazon-eks-seven-claims-tested-and-what-agent-workloads-change-1mfh</link>
      <guid>https://dev.to/parths2019/cluster-autoscaler-vs-karpenter-on-amazon-eks-seven-claims-tested-and-what-agent-workloads-change-1mfh</guid>
      <description>&lt;p&gt;Ask five platform engineers whether to run Cluster Autoscaler or Karpenter and you will get the same short answer: "Karpenter, obviously." The reasons they give are usually a mix of true, half-true and outdated.&lt;/p&gt;

&lt;p&gt;I run EKS clusters that started on node groups and Cluster Autoscaler, and I am moving them to Karpenter, so I have had to separate the claims that hold up from the ones that do not. This post goes through the seven I hear most, with a verdict on each, and then looks at what changes when the workloads are AI agents.&lt;/p&gt;

&lt;h2&gt;
  
  
  First, the mental model
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Cluster Autoscaler scales node groups.&lt;/strong&gt; On AWS those are Auto Scaling groups. When pods are pending, it picks a group that could fit them and raises its desired count. Everything it can do is bounded by the groups you defined ahead of time.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Karpenter scales nodes.&lt;/strong&gt; It reads the requirements of pending pods (requests, affinities, topology spread, volumes) and asks EC2 for a node that fits, with no Auto Scaling group in between. A NodePool says what nodes are allowed and how they may be disrupted. An EC2NodeClass says how they are built.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;EKS Auto Mode is Karpenter run for you.&lt;/strong&gt; AWS operates the controller off-cluster and uses the same NodePool API. Keep that in mind for claim five.&lt;/p&gt;

&lt;h2&gt;
  
  
  Seven claims, tested
&lt;/h2&gt;

&lt;h3&gt;
  
  
  "Karpenter is faster." Mostly true
&lt;/h3&gt;

&lt;p&gt;Karpenter calls EC2 directly. Cluster Autoscaler raises an Auto Scaling group's desired count and waits for the group to deliver. Removing that step helps, and most reports are of nodes appearing sooner.&lt;/p&gt;

&lt;p&gt;But a pending pod waits for more than the launch call: node boot, add-on startup and image pulls all count, and neither tool controls those. Benchmark on your own cluster before you promise anyone a number.&lt;/p&gt;

&lt;h3&gt;
  
  
  "Karpenter is always cheaper." Only sometimes
&lt;/h3&gt;

&lt;p&gt;Karpenter saves money in three ways: it sizes nodes to pending pods, it keeps repacking, and it makes Spot practical by spreading across many instance types. If your pods are uniform and your node groups are already well tuned, there is less to win.&lt;/p&gt;

&lt;p&gt;It also sizes nodes from pod requests, so it cannot fix pods that ask for more than they use. Inflated requests become inflated nodes. And on Auto Mode you pay a management fee on top of EC2 for the nodes it manages.&lt;/p&gt;

&lt;p&gt;AWS advertises Spot discounts of up to 90% below On-Demand, which is the ceiling, not the expectation. Compare Cost Explorer before and after, grouped by purchase option, instead of trusting anyone's percentage, including mine.&lt;/p&gt;

&lt;h3&gt;
  
  
  "Cluster Autoscaler is legacy." No
&lt;/h3&gt;

&lt;p&gt;It is maintained by Kubernetes SIG Autoscaling, has been the default since 2016, and still runs most production clusters that never switched. It works across clouds and across node group types. For a small, steady cluster with similar pods and little Spot, it is a sound choice.&lt;/p&gt;

&lt;p&gt;Its limits are real but specific. AWS's own guidance says instances in a mixed-instances policy should share the same CPU, memory and GPU shape, and that the first listed type is used to simulate scheduling, which makes it a poor fit for diversified Spot. Its scale-down only removes nodes, by default after they sit under 50% utilization for ten minutes. If neither of those hurts you, you may not need to move.&lt;/p&gt;

&lt;h3&gt;
  
  
  "Karpenter is AWS-only." Not any more, but check your platform
&lt;/h3&gt;

&lt;p&gt;The vendor-neutral core now lives with Kubernetes SIG Autoscaling, and the AWS-specific code sits in a separate provider. Other providers exist, including Azure's Node Auto Provisioning on AKS. As of mid-2026 there is still no official GKE provider.&lt;/p&gt;

&lt;p&gt;So "Karpenter everywhere" is not yet true. If you need one autoscaler across several clouds, confirm a provider exists for each before you plan around it.&lt;/p&gt;

&lt;h3&gt;
  
  
  "Auto Mode makes self-managed Karpenter pointless." No, it makes it optional
&lt;/h3&gt;

&lt;p&gt;Auto Mode removes the controller from your list of things to run, and you still define NodePools. The trade-offs are:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Bottlerocket only, with no custom AMIs and no SSH or SSM access to nodes&lt;/li&gt;
&lt;li&gt;A management fee on top of the EC2 price&lt;/li&gt;
&lt;li&gt;Managed node groups can run alongside Auto Mode nodes in the same cluster, so adoption can be gradual&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you need AMI control, node access or a different cost model, self-managed Karpenter is still the answer. If you do not, Auto Mode is the lower-effort default for new clusters.&lt;/p&gt;

&lt;h3&gt;
  
  
  "Spot with Karpenter is set and forget." No
&lt;/h3&gt;

&lt;p&gt;Karpenter handles Spot interruption through an SQS queue and EventBridge rules, cordoning and draining nodes ahead of a reclaim. That is a start, and it is not the whole job:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Spot-to-Spot consolidation needs at least 15 candidate instance types, so narrow NodePool requirements quietly disable it.&lt;/li&gt;
&lt;li&gt;Services need at least two replicas, topology spread and a PodDisruptionBudget.&lt;/li&gt;
&lt;li&gt;Pods must shut down cleanly inside the two-minute warning.&lt;/li&gt;
&lt;li&gt;The &lt;code&gt;karpenter.sh/do-not-disrupt: "true"&lt;/code&gt; annotation stops voluntary disruption, not a Spot reclaim.&lt;/li&gt;
&lt;li&gt;Databases and singletons belong on On-Demand.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  "Consolidation is free savings." It has a price
&lt;/h3&gt;

&lt;p&gt;Consolidation moves pods. At small scale that is invisible. At larger scale unconstrained consolidation can trigger waves of rescheduling that hurt pod startup latency. Karpenter gives you the controls (&lt;code&gt;consolidateAfter&lt;/code&gt;, disruption budgets and schedules), and the defaults are not tuned for your workloads. Raise &lt;code&gt;consolidateAfter&lt;/code&gt; for production and block disruption during peak hours.&lt;/p&gt;

&lt;h2&gt;
  
  
  What other teams have reported
&lt;/h2&gt;

&lt;p&gt;Published numbers vary a lot, and that spread is the point. All of these are self-reported, from different starting points, so read them as a range, not a forecast.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Salesforce&lt;/strong&gt; moved more than 1,000 EKS clusters from Cluster Autoscaler to Karpenter. It reported scaling latency falling from minutes to seconds and cost savings of about 5% in the first fiscal year, with another 5–10% expected the year after.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Slack&lt;/strong&gt; adopted Karpenter on its internal EKS platform and reported about 12% cost savings, plus better bin packing and faster node provisioning.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audi&lt;/strong&gt; reported saving more than 60% on compute for its car-configurator backend after adopting Karpenter, according to an AWS case study.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anthropic&lt;/strong&gt; engineers were reported to have said at re:Invent that Karpenter helped cut their AWS bill by around 40%. The same coverage quotes a team describing Cluster Autoscaler getting stuck when Spot capacity ran out, because it could not fall back to On-Demand.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Speed:&lt;/strong&gt; vendor benchmarks commonly put Karpenter at roughly 45 to 90 seconds to a ready node against 3 to 5 minutes for Cluster Autoscaler, and the same vendor notes that AMI caching, region and instance type change the result.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Performance can also go the wrong way.&lt;/strong&gt; One vendor warns that on clusters around 500 nodes, unconstrained consolidation can create waves of rescheduling that slow pod startup.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;My reading: teams whose clusters were already reasonably tuned saw single-digit to low-double-digit savings, and teams with a lot of idle capacity or unused Spot saw far more. Scale-up speed improved in nearly every account, and stability depended on how carefully disruption was configured. Your own before-and-after numbers are worth more than any of these.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where agent workloads change the picture
&lt;/h2&gt;

&lt;p&gt;For ordinary web services, a slow or clumsy autoscaler is an annoyance. Agent platforms turn it into a product problem.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Bursts have a person waiting.&lt;/strong&gt; One request can fan out into many tool calls and sandbox pods at once, so node provisioning time lands directly in someone's wait.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pod shapes are wildly uneven.&lt;/strong&gt; A small tool server, a code sandbox and a GPU inference pod have nothing in common. Under Cluster Autoscaler that becomes a growing list of node groups. Under Karpenter it becomes a few NodePools with requirements.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;GPU capacity is a purchasing problem.&lt;/strong&gt; EKS documentation lists On-Demand, Spot, On-Demand Capacity Reservations and Capacity Blocks for ML as supported options for Auto Mode and Karpenter, using reserved capacity first. For accelerated nodes, the question is often which kind of capacity you hold, not how fast you can scale.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Idle time is expensive, and so is interruption.&lt;/strong&gt; Agent sessions spend most of their time waiting on model responses, which makes consolidation attractive. But an evicted pod can mean a lost session. Put long-lived session pods in their own NodePool with a conservative disruption policy, mark runs that must not move with &lt;code&gt;do-not-disrupt&lt;/code&gt;, keep fragile agents on On-Demand, and store session state outside the pod.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Untrusted code needs its own nodes.&lt;/strong&gt; If an agent runs generated code, that pod should not share a node with your platform services. A dedicated NodePool with a taint, a short node lifetime and a CPU limit gives you a boundary and an easy cleanup:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;karpenter.sh/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;NodePool&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;agent-sandbox&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;template&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;labels&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="na"&gt;workload&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;agent-sandbox&lt;/span&gt;
    &lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;nodeClassRef&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="na"&gt;group&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;karpenter.k8s.aws&lt;/span&gt;
        &lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;EC2NodeClass&lt;/span&gt;
        &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;default&lt;/span&gt;
      &lt;span class="na"&gt;taints&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;workload&lt;/span&gt;
          &lt;span class="na"&gt;value&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;agent-sandbox&lt;/span&gt;
          &lt;span class="na"&gt;effect&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;NoSchedule&lt;/span&gt;
      &lt;span class="na"&gt;expireAfter&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;24h&lt;/span&gt;
      &lt;span class="na"&gt;requirements&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;karpenter.sh/capacity-type&lt;/span&gt;
          &lt;span class="na"&gt;operator&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;In&lt;/span&gt;
          &lt;span class="na"&gt;values&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"spot"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"on-demand"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
  &lt;span class="na"&gt;limits&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;cpu&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"100"&lt;/span&gt;
  &lt;span class="na"&gt;disruption&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;consolidationPolicy&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;WhenEmpty&lt;/span&gt;
    &lt;span class="na"&gt;consolidateAfter&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;2m&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Sandbox pods then carry a matching toleration and node selector. &lt;code&gt;WhenEmpty&lt;/code&gt; removes nodes only when no pods remain, so nothing is repacked underneath a running sandbox. On Auto Mode, &lt;code&gt;nodeClassRef&lt;/code&gt; points at its &lt;code&gt;NodeClass&lt;/code&gt; instead.&lt;/p&gt;

&lt;p&gt;A taint and a separate pool do not make untrusted code safe on their own. They stop a mistake from spreading to your platform nodes, and you still want stronger isolation such as a sandboxed runtime.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would do
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Small, steady cluster with uniform pods and little Spot:&lt;/strong&gt; stay on Cluster Autoscaler and spend the effort on pod requests.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Varied pods, meaningful Spot use or bursty demand:&lt;/strong&gt; move to Karpenter. Start with Auto Mode if Bottlerocket and no node access are acceptable, and self-manage when you need AMI or node control.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;An agent platform:&lt;/strong&gt; Karpenter with separate NodePools for sandboxes, session pods and GPU work, with disruption settings treated as part of the design.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you do move, the two tools can run side by side. Add Karpenter next to your node groups, remove Cluster Autoscaler before you drain anything because two autoscalers on one cluster fight each other, keep a small managed node group for Karpenter and core add-ons, shrink the old groups gradually behind PodDisruptionBudgets, and put a CPU limit on every NodePool.&lt;/p&gt;

&lt;p&gt;Karpenter is the better default for new AWS clusters in 2026, but "better default" is not "mandatory."&lt;/p&gt;

</description>
      <category>karpenter</category>
      <category>eks</category>
      <category>autoscaling</category>
      <category>costsavings</category>
    </item>
    <item>
      <title>Amazon ECS vs. Amazon EKS : Architecture, Cost, Skills, and Running Agentic AI Workloads</title>
      <dc:creator>Parth Shah</dc:creator>
      <pubDate>Tue, 17 Nov 2026 01:09:19 +0000</pubDate>
      <link>https://dev.to/parths2019/amazon-ecs-vs-amazon-eks-architecture-cost-skills-and-running-agentic-ai-workloads-3ajk</link>
      <guid>https://dev.to/parths2019/amazon-ecs-vs-amazon-eks-architecture-cost-skills-and-running-agentic-ai-workloads-3ajk</guid>
      <description>&lt;h2&gt;
  
  
  The question that keeps coming back
&lt;/h2&gt;

&lt;p&gt;I have spent seven years building AWS infrastructure, and I currently run workloads on both Amazon ECS and Amazon EKS. Almost every new project starts with the same question: which one should this go on?&lt;/p&gt;

&lt;p&gt;On ECS, I run long-lived web applications and an identity service backed by a managed relational database. These need a load balancer, secrets and little else. On EKS, I run a GitOps-managed platform of microservices and AI workloads, where node autoscaling and Spot capacity do most of the cost work.&lt;/p&gt;

&lt;p&gt;The old rule of thumb was "ECS when you are small, EKS when you scale." I no longer think that holds, for two reasons:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;AWS removed most of the operational gap.&lt;/strong&gt; EKS Auto Mode, EKS Capabilities, ECS Managed Instances and ECS Express Mode all shipped in the last two years. A lot of what made EKS heavy and ECS limited is now optional.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agentic AI changed the shape of the workloads.&lt;/strong&gt; Agents run untrusted code, hold long sessions, sit idle waiting on model calls, and sometimes need GPUs. Those requirements pull in different directions.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;This guide covers what changed, five signals to score your team, three scenarios, the real cost drivers, and where agent workloads fit.&lt;/p&gt;




&lt;h2&gt;
  
  
  1. Why most enterprises still default to EKS
&lt;/h2&gt;

&lt;p&gt;In the CNCF Annual Cloud Native Survey published in January 2026, &lt;strong&gt;82% of container users reported running Kubernetes in production&lt;/strong&gt;, up from 66% in 2023 (&lt;a href="https://www.cncf.io/reports/the-cncf-annual-cloud-native-survey/" rel="noopener noreferrer"&gt;CNCF&lt;/a&gt;). The survey polls the cloud native community, so it leans toward Kubernetes users, but the direction is clear.&lt;/p&gt;

&lt;p&gt;The reasons I see in practice:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hiring.&lt;/strong&gt; An engineer from GKE or AKS is productive on EKS in days. ECS concepts are AWS-specific and have to be learned.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ecosystem.&lt;/strong&gt; Flux, Argo CD, Kyverno, Cilium, Prometheus, Crossplane and KEDA all assume the Kubernetes API.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Portability.&lt;/strong&gt; Manifests and Helm charts move between EKS, on-premises clusters and other clouds. ECS task definitions do not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Operators.&lt;/strong&gt; Stateful systems and GPU stacks are usually packaged as Kubernetes operators. ECS has no equivalent extension model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local parity.&lt;/strong&gt; &lt;code&gt;kind&lt;/code&gt; or &lt;code&gt;k3d&lt;/code&gt; gives developers a real control plane on a laptop. ECS local testing is closer to Docker Compose.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of these is about scale. They are about people and tooling.&lt;/p&gt;




&lt;h2&gt;
  
  
  2. What changed: the 2026 landscape
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Launch&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;th&gt;Why it matters to this decision&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;EKS Auto Mode&lt;/strong&gt; (Dec 2024)&lt;/td&gt;
&lt;td&gt;AWS manages nodes, autoscaling, load balancing, DNS and block storage&lt;/td&gt;
&lt;td&gt;Removes most add-on and node upgrade work; adds a per-instance management fee&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;EKS Capabilities&lt;/strong&gt; (Nov 2025)&lt;/td&gt;
&lt;td&gt;AWS-hosted Argo CD, ACK and kro, running outside your cluster&lt;/td&gt;
&lt;td&gt;GitOps without operating the GitOps tool&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;ECS Managed Instances&lt;/strong&gt; (Sep 2025)&lt;/td&gt;
&lt;td&gt;AWS provisions and patches EC2 for ECS, including GPU instances&lt;/td&gt;
&lt;td&gt;EC2 flexibility on ECS without owning AMIs or Auto Scaling groups; adds a management fee&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;ECS Express Mode&lt;/strong&gt; (Nov 2025)&lt;/td&gt;
&lt;td&gt;Image plus two IAM roles gives you a load-balanced, autoscaled HTTPS service&lt;/td&gt;
&lt;td&gt;Shortest path from container to URL on AWS&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Managed MCP servers for EKS and ECS&lt;/strong&gt; (preview, Nov 2025)&lt;/td&gt;
&lt;td&gt;AI assistants can inspect and troubleshoot clusters through IAM-governed tools&lt;/td&gt;
&lt;td&gt;Both platforms are now operable by agents&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;What this means:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On EKS, the "upgrade tax" is now a choice. With Auto Mode you still plan Kubernetes version upgrades and test deprecated APIs, but AWS handles nodes and core add-ons. Run it self-managed and you own the VPC CNI, CoreDNS, kube-proxy, the Load Balancer Controller, Karpenter and the rest.&lt;/li&gt;
&lt;li&gt;On ECS, "no GPUs, no EC2 flexibility without toil" is no longer true.&lt;/li&gt;
&lt;li&gt;What has &lt;strong&gt;not&lt;/strong&gt; changed: ECS has no control plane fee and no version lifecycle. EKS still has both.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  3. Five signals: which way does your team lean?
&lt;/h2&gt;

&lt;p&gt;For each signal, pick the statement that sounds more like your team, then count how many times you picked EKS.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Who operates the platform?&lt;/strong&gt;&lt;br&gt;
ECS if it is one or two DevOps engineers, or developers who own their own deployments. EKS if a platform team of three or more runs shared cluster infrastructure as its main job.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Do you need the Kubernetes ecosystem?&lt;/strong&gt;&lt;br&gt;
ECS if ALB, IAM, CloudWatch and CodeDeploy cover your needs. EKS if you depend on operators, CRDs, policy engines, a service mesh or existing Helm charts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. How portable must workloads be?&lt;/strong&gt;&lt;br&gt;
ECS if you are committed to AWS. EKS if you need the same deployment model on-premises or on another cloud.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. How many services, and how varied?&lt;/strong&gt;&lt;br&gt;
ECS if you run fewer than about 25, mostly stateless APIs, workers and scheduled jobs. EKS if you run dozens to hundreds, with stateful systems or mixed scheduling needs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. What do your AI workloads look like?&lt;/strong&gt;&lt;br&gt;
ECS if you call hosted models such as Amazon Bedrock and run agents as ordinary services. EKS if you self-host models on GPUs and need queue-based scaling, GPU sharing or custom schedulers.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Reading your count&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;0 or 1 EKS answers: choose ECS.&lt;/strong&gt; Start with Fargate or Express Mode. EKS would add work without a payoff.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;2 or 3: either works.&lt;/strong&gt; Choose ECS if delivery speed matters most. Choose EKS with Auto Mode if you already have Kubernetes skills or expect to need the ecosystem within a year.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;4 or 5: choose EKS.&lt;/strong&gt; The control plane fee and upgrade planning are justified. Pick Auto Mode or self-managed nodes based on how much control you need.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Signals 2 and 3 carry the most weight. If either one is a firm EKS answer, treat the result as EKS even when the count is low.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. Three scenarios
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;A. Preview environments for every pull request.&lt;/strong&gt; 18 environments, 6 small containers each.&lt;/p&gt;

&lt;p&gt;The fair comparison is 18 sets of ECS services against &lt;strong&gt;one&lt;/strong&gt; EKS cluster with 18 namespaces, which costs about $73 a month for the control plane. Nobody should run 18 clusters (that would be about $1,314 a month before any compute). &lt;strong&gt;Call: ECS on Fargate&lt;/strong&gt; if the team has no Kubernetes skills, because the environments cost nothing when idle and need no cluster. If you already run EKS, namespaces are just as good.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;B. 80 microservices, steady traffic, pressure to cut compute cost.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;At this service count, bin-packing and Spot handling decide the bill. Karpenter consolidates nodes continuously and handles Spot interruptions well. &lt;strong&gt;Call: EKS with Karpenter (or Auto Mode).&lt;/strong&gt; ECS Managed Instances narrows this gap, so benchmark it if you are already on ECS.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;C. An agent that executes generated code for a few seconds.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Fargate gives each task its own microVM, which is the right isolation for untrusted code. The catch is startup: a new Fargate task normally takes 30 seconds or more to start, and billing has a one-minute minimum. For a 15-second job, that is the wrong shape. &lt;strong&gt;Call: a purpose-built sandbox&lt;/strong&gt; such as Amazon Bedrock AgentCore Code Interpreter or AWS Lambda for short runs; &lt;strong&gt;Fargate &lt;code&gt;RunTask&lt;/code&gt;&lt;/strong&gt; when jobs run for minutes and need a full container image.&lt;/p&gt;




&lt;h2&gt;
  
  
  5. Cost: what actually drives the bill
&lt;/h2&gt;

&lt;p&gt;Compute rates are the same on both platforms. The differences are these:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Control plane.&lt;/strong&gt; ECS charges nothing. EKS charges $0.10 an hour per cluster, about $73 a month.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Falling behind on versions.&lt;/strong&gt; An EKS cluster in extended support costs $0.60 an hour, about $438 a month. ECS has no version lifecycle.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Managed compute fees.&lt;/strong&gt; ECS Managed Instances and EKS Auto Mode each add a per-instance management fee on top of EC2.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bin-packing.&lt;/strong&gt; Karpenter consolidates nodes continuously and diversifies Spot well. ECS capacity providers and Managed Instances have narrowed the gap.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;People.&lt;/strong&gt; This is usually the largest line. EKS has more moving parts: version upgrades, add-ons, policy and more to know.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A small application on a self-managed EKS cluster can cost more in engineering time than in AWS charges. Kubernetes standard support lasts 14 months per version, with three releases a year, so upgrade planning never stops.&lt;/p&gt;




&lt;h2&gt;
  
  
  6. Why this matters more in the agentic era
&lt;/h2&gt;

&lt;p&gt;Until recently, most container workloads looked alike: long-running stateless services behind a load balancer. Either orchestrator handled them well, so the choice came down to team preference. Agent workloads break that symmetry in five ways.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. The code is untrusted.&lt;/strong&gt; An agent that writes and runs code is executing something no human reviewed. Isolation becomes a design requirement. On a standard Kubernetes node, pods share a kernel, so you need extra layers (dedicated nodes, sandboxed runtimes, strict network policy). On Fargate, every task gets its own microVM by default.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Sessions are long and mostly idle.&lt;/strong&gt; An agent may hold state for hours while spending most of that time waiting on model responses. Per-task pricing for idle time gets expensive, which favors dense packing on shared nodes. That is where EKS with Karpenter, or ECS on EC2, does well.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Bursts are sharp.&lt;/strong&gt; One user request can fan out into dozens of tool calls and sub-agents. Startup time now matters as much as steady-state cost. Warm capacity on EC2-backed clusters beats cold Fargate starts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Identity has to be per agent.&lt;/strong&gt; Each agent or tool server should have its own narrowly scoped IAM role. ECS task roles and EKS Pod Identity both do this. The more agents you run, the more this design work dominates.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. There is now a third option.&lt;/strong&gt; Amazon Bedrock AgentCore Runtime gives each agent session a dedicated microVM and manages scaling for you. For many teams the real question is no longer "ECS or EKS" but "do I need to run the agent's compute at all?"&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where each workload fits&lt;/strong&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Workload&lt;/th&gt;
&lt;th&gt;Best fit&lt;/th&gt;
&lt;th&gt;Why&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Short code-execution sandboxes&lt;/td&gt;
&lt;td&gt;AgentCore Code Interpreter or Lambda&lt;/td&gt;
&lt;td&gt;Fast start, built for untrusted code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Long-running agent sessions with isolation&lt;/td&gt;
&lt;td&gt;AgentCore Runtime, or ECS on Fargate&lt;/td&gt;
&lt;td&gt;MicroVM per session or task&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MCP tool servers and agent APIs&lt;/td&gt;
&lt;td&gt;ECS (Fargate or Express Mode)&lt;/td&gt;
&lt;td&gt;Ordinary HTTP services; simplest to run&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Agents calling Bedrock, coordinated by SQS or Step Functions&lt;/td&gt;
&lt;td&gt;ECS&lt;/td&gt;
&lt;td&gt;Native AWS integration, little to operate&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Self-hosted model inference on GPUs&lt;/td&gt;
&lt;td&gt;EKS; ECS Managed Instances for simpler cases&lt;/td&gt;
&lt;td&gt;GPU operators, KEDA and Karpenter on EKS; managed GPU nodes on ECS&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A platform hosting many teams' agents&lt;/td&gt;
&lt;td&gt;EKS&lt;/td&gt;
&lt;td&gt;Namespaces, policy engines, quotas and operators&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The practical consequence: &lt;strong&gt;many organizations will run more than one of these.&lt;/strong&gt; A common shape is hosted models through Bedrock, tool servers on ECS, and a GPU inference pool on EKS. Choosing per workload is better than forcing everything onto one platform.&lt;/p&gt;




&lt;h2&gt;
  
  
  7. What your team needs
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;For ECS:&lt;/strong&gt; one or two DevOps engineers, or developers who own their services. They need solid VPC networking, a clear grasp of task roles versus execution roles, ALB target groups, and Terraform or CDK. Day-2 tooling is Container Insights, Secrets Manager, and CodeDeploy or native blue/green.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;For EKS:&lt;/strong&gt; a platform team, or Auto Mode plus at least one strong Kubernetes engineer. They need the Kubernetes API, Helm or Kustomize, Pod Identity and RBAC, Karpenter and GitOps. Day-2 tooling is Prometheus or ADOT, a policy engine, image scanning and a GitOps controller.&lt;/p&gt;




&lt;h2&gt;
  
  
  Verdict
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Choose ECS&lt;/strong&gt; when you are committed to AWS, the team is small, and the workloads are APIs, workers and tool servers. Start with Fargate or Express Mode.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Choose EKS&lt;/strong&gt; when you need the Kubernetes ecosystem, portability, GPU scheduling, or a shared platform for many teams. Start with Auto Mode unless you have a reason to manage nodes yourself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;For agents&lt;/strong&gt;, decide per workload, and check whether a managed runtime removes the question.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The best orchestrator is the one your team can operate well at 2 a.m.&lt;/p&gt;

</description>
      <category>ecs</category>
      <category>eks</category>
      <category>aws</category>
      <category>containers</category>
    </item>
  </channel>
</rss>
