Amazon EKS
Managed Kubernetes control plane — AWS runs the masters, you run the workloads.
Verified against AWS documentation · 2026-08-28
Key facts
- 01AWS operates the Kubernetes control plane across multiple AZs; you bring worker compute — managed node groups (EC2), self-managed EC2, or Fargate
- 02Runs upstream, certified Kubernetes — existing manifests, kubectl, and Helm charts work unchanged, and workloads stay portable off AWS
- 03Fargate runs pods serverless — no nodes to patch, scale, or right-size; each pod gets its own isolated compute
- 04Hybrid flavors: EKS Anywhere (run clusters on your own hardware), EKS Distro (the same Kubernetes distribution, self-run anywhere)
Exam triggers
Phrases and scenarios that point here.
- "Kubernetes", "K8s", "open-source orchestration", "avoid vendor lock-in" → EKS, not ECS
- "Migrate existing on-prem Kubernetes workloads to AWS with minimal changes" → EKS
- Kubernetes + "no server management" → EKS on Fargate
- Kubernetes on your own on-prem hardware → EKS Anywhere
Not this service
- Containers with no Kubernetes requirement, simplest AWS-native option → ECS
- Just deploy a container from a repo, zero orchestration knowledge → App Runner
Use cases
- 01Lift-and-shift an on-prem Kubernetes estate to AWS — the team keeps its manifests, Helm charts, and muscle memory instead of relearning a proprietary orchestrator
- 02Multi-cloud or exit-strategy mandates, where the same workload definitions must run on AWS today and somewhere else tomorrow
- 03Teams already invested in the Kubernetes ecosystem (Istio, ArgoCD, Prometheus operators) that ECS can't host
- 04Batch/ML jobs on spare cluster capacity — Kubernetes scheduling packs bursty jobs alongside steady services on the same nodes
- 05Platform teams offering an internal 'paved road': one shared cluster, many product teams deploying via namespaces and RBAC
How it's charged
- 01Flat per-cluster hourly fee for the control plane, billed whether or not any nodes are running — per cluster, not per node
- 02The fee is tiered by Kubernetes version: fall past the ~14-month standard support window and the cluster silently moves to extended support at 6x the hourly rate
- 03Worker compute is billed separately as normal EC2 or Fargate usage — this, not the cluster fee, is where most of the bill lives
- 04Auto Mode adds a per-instance management surcharge on top of the EC2 price; Savings Plans and RIs discount the EC2 portion only, never the surcharge
- 05Usual adjacent costs apply: EBS volumes, load balancers, public IPv4, and cross-AZ data transfer between nodes
