deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

EKS Auto Mode preview resizes pod resource requests based on observed usage

Amazon EKS Auto Mode's resource optimization preview adjusts pod CPU and memory requests from measured usage, resizing running pods in place without editing the templates GitOps tools track.

EKS Auto Mode preview resizes pod resource requests based on observed usage

Amazon EKS Auto Mode, the variant of AWS's managed Kubernetes service where AWS owns the worker nodes and provisions instances to fit declared pod requests, has gained a preview feature that right-sizes those requests from what containers actually consume. According to dev.to, which summarized the AWS announcement, resource optimization was published on 17 November 2026 for clusters running Kubernetes 1.36 or later, in seven AWS regions, at no additional charge.

What the feature changes

Auto Mode's cost machinery has always treated the resource requests written in pod templates as fixed inputs. Bin-packing picks instance types to match the sum of those requests, consolidation removes nodes when the same pods fit on fewer of them, and right-sizing shifts pods onto smaller nodes as workloads scale down. None of it looked at real consumption, and requests were usually written once at deploy time and never revisited.

The new feature observes each container's CPU and memory usage and turns the measurement into a recommended request. The scope is deliberately narrow: only Deployments and StatefulSets, and only CPU and memory. GPUs, ephemeral storage, and replica counts keep their declared values, the metrics that horizontal autoscalers such as HPA and KEDA rely on are untouched, and workloads managed by the Vertical Pod Autoscaler can be excluded. Recommendations appear as a ResourceOptimization object, short name ro, in the workload's namespace, created and deleted by the cluster rather than edited by hand.

Two modes, one cautious default

A ClusterResourceOptimizationPolicy selects the namespaces and workloads to manage and picks one of two modes. Recommend, the default, collects usage and writes recommendations without altering any pod, which makes it usable on a production cluster from the day the policy is created. Apply goes further and resizes running pods in place through the Kubernetes pod resize subresource, changing the pod without recreating it.

That mechanism matters for GitOps users. Because the resize never edits the Deployment or StatefulSet pod template, controllers such as Argo CD and Flux keep comparing the cluster against the manifest in the repository and find no drift to revert. Pods created afterwards, whether by scale-out, node consolidation, or a rollout, start at the recommended size.

How the sizing works

The preview ships a single profile, Balanced. For CPU it sets the request so that the 95th percentile of observed usage lands at roughly 87 percent of the request, leaving about 13 percent as buffer before throttling or node pressure. Timing is equally concrete: a container needs around 12 hours of usage history before its first recommendation, and until then it keeps the request from its template. A workload deployed on Friday afternoon therefore waits until Saturday morning, a detail dev.to flags as decisive for rollout planning.

Operational gaps

Two limitations deserve attention. There is no automatic rollback, so a recommendation that degrades a workload stays applied until someone opts the workload out or switches its policy back to Recommend. And there is no dedicated console page: policies and recommendation objects surface read-only in the EKS console's Kubernetes resource viewer, which leaves the command line and API as the working interface.

The feature also changes the input to Auto Mode's cost decisions rather than the decisions themselves. It adds no budget control, and a NodePool limits setting remains the only hard ceiling on provisioned capacity, still stopping node launches when the limit is reached even with pods pending.

Why it matters

Over-provisioned requests translate directly into over-provisioned clusters: Auto Mode sizes instances to the sum of declared requests, so requests set at double the real need produce a cluster roughly twice as large as the work requires, and consolidation cannot correct that because the pods still claim the capacity. Feeding measured usage back into the request is one of the most direct cost levers a Kubernetes operator has, and doing it in place without conflicting with Argo CD or Flux removes a common objection to automated rightsizing. With no rollback and no console tooling in the preview, the rollout burden falls on policy scoping, and the cautious sequence the documentation implies, Recommend first, compare against dashboards, then Apply one namespace at a time, is the path platform teams will likely take.

  • #aws
  • #kubernetes
  • #eks
  • #cost-optimization
  • #devops

Related posts