Decide why you need Kubernetes
Many public services run perfectly well on AWS Lambda, Amazon ECS or a small number of virtual machines, and each of those carries less operational work than a Kubernetes cluster. Amazon EKS earns its place when several teams deploy a growing number of services and need a common way to release, observe and govern them, or when the organisation already has Kubernetes skills and a reason to keep workloads portable. Record that reasoning as an architecture decision, because it will be questioned at the first spend review and again when the people who made it move on.
Know what AWS runs for you
With standard EKS, AWS operates the Kubernetes API server and its etcd data store. EKS Auto Mode extends that to the worker nodes: AWS selects and patches the node images, scales capacity for pods that cannot be scheduled, provisions load balancers for services and ingress, and manages ephemeral storage and pod networking.
Your team still owns the application containers, including their availability, security and monitoring, together with the VPC and the cluster configuration. Auto Mode also replaces each node after at most 21 days, so every workload has to tolerate being rescheduled. A pod disruption budget that blocks eviction can hold up those replacements, which makes disruption budgets an engineering decision for each service rather than a default left untouched.
Plan for the version calendar
Each Kubernetes version receives 14 months of standard support on EKS, followed by up to 12 months of extended support charged at an additional rate per cluster hour. New and existing clusters default to the extended policy. When extended support ends, AWS upgrades the control plane automatically, at any time and without notice, and that upgrade cannot be rolled back. Managed and self-managed node groups stay on the earlier version until you upgrade them.
For a public-sector service this belongs in the budget and the change calendar. Treat upgrades as routine work scheduled several times a year, rehearse each one in a non-production cluster, check for removed APIs and add-on compatibility, and agree in advance whether the organisation will pay for extended support or keep to the standard window.
Decide who can reach the cluster
EKS access entries are the recommended way to grant IAM roles access to the Kubernetes API. They are managed through the EKS API and replace the older aws-auth ConfigMap. Enabling access entries on an existing cluster only creates an entry for the identity that created the cluster, so every other mapping has to be recreated deliberately. That is a useful moment to remove access nobody can explain.
Workloads that call AWS services should use EKS Pod Identity, which maps an IAM role to a Kubernetes service account and gives pods temporary credentials. Keep both the human and the workload mappings in Terraform, so that an access review becomes a review of code with a history behind it.
Write the responsibility split down
A short table usually settles it. The platform team owns cluster configuration, Kubernetes versions, add-ons, node pools, ingress and the baseline for logs and metrics. Each application team owns its manifests, resource requests, health checks, disruption budget and the alerts for its own service. AWS owns the control plane and, under Auto Mode, the node lifecycle. Include the table in the operational acceptance pack and update it whenever the arrangement changes.
Read about platform engineering or discuss your path to production.
Further reading
Amazon EKS: Kubernetes version lifecycle
