Existing clusters¶

A cluster created before 12 August 2026 does not change behaviour when the feature launches in your Region. Amazon EKS never alters a running cluster just because a new parameter becomes available.

CLI commands on this page need AWS CLI 2.36.21 or later. See Requirements. The console does not.

What changes immediately (and what doesn't)¶

Two things happen on launch day. Only one is visible, and neither changes how your cluster schedules pods.

The API response gains new fields. describe-cluster returns kubeSchedulerConfig, kubeControllerManagerConfig, and kubeApiServerConfig even if you never set them. On a cluster that has never been configured, these fields are null — not populated with defaults.

The console shows a new panel. The cluster detail page includes a "Control plane configuration" section. That the panel exists does not mean anything was applied.

Timeline¶

flowchart TD
  A["Cluster exists<br/>Created before 12 Aug 2026 · K8s 1.31+"] --> B["Feature launches in the Region"]
  B --> C["DescribeCluster returns new fields<br/>null until you Save"]
  C --> D{"Do you call<br/>UpdateClusterConfig?"}
  D -- "No" --> E["Nothing changes<br/>LeastAllocated · 60 min TTL · 30000–32767"]
  D -- "Yes" --> F["Rolling control plane update<br/>Status: UPDATING → ACTIVE"]
  F --> G["Only the fields you set change<br/>New events / new binds only<br/>Running pods stay bound"]

Defaults on an unchanged cluster¶

Requires Kubernetes 1.31 or later. If your cluster is on an older version, upgrade before attempting a configuration update.

Migrating off a custom scheduler¶

Skip this section unless the cluster already runs a second scheduler. That is the workaround from before this feature: a customer-managed scheduler in the data plane, because the EKS default could not pack.

This lab cluster (lab-cluster) did not take that path. Save set event TTL to 45m and left scoring on LeastAllocated. The steps below apply only if you are replacing a custom packer with MostAllocated on the managed scheduler.

Amazon EKS does not choose a scheduler for a pod. Kubernetes does, from a field on the pod:

spec:
  schedulerName: custom-scheduler   # omit this (or set default-scheduler) to use the EKS-managed process

UpdateClusterConfig only changes how default-scheduler scores nodes. Pods that still name custom-scheduler never reach that process.

The three columns use two example Deployments: web never set the field; pack did. The trap is the middle column — both schedulers pack, and pack still ignores the managed one.

1. Call UpdateClusterConfig with scoringStrategy: MostAllocated. Wait until the control plane is ACTIVE.
2. Remove spec.schedulerName from the packing Deployments (or set it to default-scheduler). Roll them. Only new pods switch; running pods stay bound until they are recreated.
3. Confirm new pods land on more-allocated nodes (kubectl get pods -o wide). If they do not, check affinity, taints, topology spread, and that the field is actually gone on the new pods.
4. Delete the custom-scheduler Deployment, its RBAC, and any admission webhook that injected schedulerName.

Summary for cluster owners¶

On launch day an existing cluster keeps LeastAllocated, 60 minute event TTL, NodePorts 30000–32767, and HPA 15s. None of that changes until you Save.

This lab then set event TTL to 45 minutes. Scoring stayed LeastAllocated. NodePorts and HPA stayed at the defaults. Overview Enabled: Yes because TTL is no longer 60m.

1. Call UpdateClusterConfig (or click Manage → Save in the console).
2. Wait for the rolling control plane update to complete (typically a few minutes).
3. Accept that only the fields you set change, and only for new events and new scheduling decisions.

Running pods are never moved by a scoring strategy change.

lab-cluster · Details · created 8 Jul 2026

lab-cluster · Control plane configuration · Enabled No

lab-cluster · Manage · Enable unchecked

lab-cluster · Manage · Enable checked · defaults

lab-cluster · configuration update in progress

lab-cluster · Cluster info · Updating

lab-cluster · Control plane configuration · Enabled Yes

$ aws eks describe-cluster --name lab-cluster \
    --query 'cluster.{
      version: version,
      createdAt: createdAt,
      scheduler: kubeSchedulerConfig,
      controller: kubeControllerManagerConfig,
      apiserver: kubeApiServerConfig
    }' --region ap-southeast-2 --output json
{
    "version": "1.36",
    "createdAt": "2026-07-08T08:28:40.494000+12:00",
    "scheduler": null,
    "controller": null,
    "apiserver": null
}

$ aws eks describe-cluster --name lab-cluster \
    --query 'cluster.{
      version: version,
      createdAt: createdAt,
      scheduler: kubeSchedulerConfig,
      controller: kubeControllerManagerConfig,
      apiserver: kubeApiServerConfig
    }' --region ap-southeast-2 --output json
{
    "version": "1.36",
    "createdAt": "2026-07-08T08:28:40.494000+12:00",
    "scheduler": {
        "nodeResourcesFit": {
            "scoringStrategy": {
                "type": "LeastAllocated",
                "resources": [
                    {"name": "cpu", "weight": 1},
                    {"name": "memory", "weight": 1}
                ]
            }
        }
    },
    "controller": {
        "horizontalPodAutoscalerControllerConfig": {
            "horizontalPodAutoscalerSyncPeriod": "15s"
        }
    },
    "apiserver": {
        "eventTtl": "45m",
        "serviceNodePortRange": {
            "minPort": 30000,
            "maxPort": 32767
        }
    }
}

null is before Save. After a non-default (eventTtl: 45m) the three objects are populated. Scoring is still LeastAllocated. HPA is still 15s on Standard. Overview Enabled: Yes matches this JSON, not a wait.