Amazon EKS (Elastic Kubernetes Service) is AWS's managed Kubernetes. AWS runs and scales the Kubernetes control plane - the API server and etcd - across availability zones, patches it, and keeps it highly available, while you run your workloads on worker nodes. You get standard, conformant Kubernetes (so your manifests and tooling are portable) without operating the cluster's brain.

What EKS manages, and what you still do

AWS owns the control plane; you own the data plane and the application. Worker capacity comes as managed node groups (EC2 instances AWS provisions and rolls), Fargate (serverless pods, no nodes to manage), or Karpenter (fast, flexible node autoscaling). You still design workloads, set requests/limits, configure autoscaling, and manage upgrades of your add-ons.

AWS integration is the differentiator

  • IAM Roles for Service Accounts (IRSA) - grant pods fine-grained AWS permissions (e.g., read one S3 bucket) without static keys.
  • VPC CNI - pods get real VPC IPs, integrating with AWS networking and security groups.
  • Load balancers, EBS/EFS storage, CloudWatch - native drivers wire Kubernetes to AWS infrastructure.

Why data engineers meet it

EKS is a common home for Spark-on-Kubernetes, Airflow (the official Helm chart or MWAA's underpinnings), and other data workloads on AWS. IRSA is the clean way to let those pods reach S3, Glue, and other services with scoped permissions instead of embedded credentials.

EKS vs the alternatives

EKS is the AWS counterpart to Azure's AKS and Google's GKE - all managed Kubernetes, differing mainly in cloud integration and defaults (GKE is often considered the most hands-off, EKS the most AWS-integrated). Against running your own Kubernetes on EC2, EKS removes the hardest, riskiest operational work (the control plane) for a per-cluster fee. Against ECS (AWS's simpler proprietary orchestrator), EKS gives you portable, standard Kubernetes at the cost of more complexity. Choose EKS when you want Kubernetes specifically and you're on AWS; choose ECS or Fargate-only if you want less Kubernetes surface area.