Amazon Elastic Kubernetes Service (EKS)
Amazon Elastic Kubernetes Service (EKS) is a managed Amazon Web Services service that runs Kubernetes and lets users manage Kubernetes applications in the AWS cloud or on premises.
AI Practitioner focus
- EKS runs Kubernetes-based AI/ML platforms and custom model-serving stacks where portability, ecosystem tooling, GPU scheduling, or cluster-level control justifies the operational complexity.
- It is more customizable but higher-operations than Amazon Bedrock or a managed SageMaker endpoint; select it only when Kubernetes is an actual requirement.
Key points
- An EKS cluster consists of two components: the Amazon EKS control plane and Amazon EKS nodes.
- Kubernetes clusters and applications can be managed across hybrid environments without altering the code.
- The control plane consists of nodes running the Kubernetes software, such as etcd and the Kubernetes API server.
- Runs Kubernetes control plane instances across multiple Availability Zones for high availability.
- Automatically replaces unhealthy control plane instances and provides automated upgrades and patches for new control planes.
- Batch workloads can be executed on the EKS cluster using the Kubernetes Jobs API across Amazon EC2, AWS Fargate, and Spot Instances.
- Two methods for creating a new Kubernetes cluster with nodes:
- eksctl — a command-line utility (built on kubectl) for creating/managing Kubernetes clusters on EKS.
- AWS Management Console and AWS CLI.
- Integrations for unique capabilities:
- Images — Amazon Elastic Container Registry (ECR) for container images.
- Load distribution — Elastic Load Balancing (ELB).
- Authentication — AWS Identity and Access Management (IAM).
- Isolation — Amazon VPC.
- EKS Auto Mode extends AWS management beyond the control plane to the data plane: it provisions and scales EC2 nodes (via Karpenter, on locked-down Bottlerocket AMIs with a 21-day max node lifetime), and manages pod/service networking, load balancer provisioning, cluster DNS (Domain Name System), block storage (EBS CSI), and GPU drivers as built-in cluster capabilities rather than self-managed add-ons.
- Amazon EKS Anywhere runs the same EKS cluster software on customer-owned infrastructure (on-premises or edge), for a consistent Kubernetes experience outside AWS Regions.
- Kubernetes version support in EKS has two tiers: standard support for 14 months after a version’s release, billed at 0.60 per cluster per hour for clusters that stay on an older version.