Amazon EKS · Lesson 2 of 11 · 3 min read
Compute options
Choices
| Option | What it is | Good for | Trade-offs |
|---|---|---|---|
| Managed node groups | EKS-managed EC2 Auto Scaling groups with EKS-optimised AMIs; managed rolling updates, drain on scale-in | Default for most clusters; system node pools | Fixed instance types per group; scaling via CA |
| Self-managed nodes | Your own ASGs/launch templates | Custom AMIs/bootstrapping needs | More work |
| Karpenter | Open-source node provisioner launching EC2 directly per pending pod needs | Diverse workloads, Spot, cost efficiency | You operate it (or use Auto Mode) |
| Fargate | Serverless pods — each pod in its own micro-VM; no nodes | Spiky/small workloads, strong isolation, no node ops | Higher per-pod cost, no DaemonSets, no privileged pods, no GPUs, limited storage (EFS only), slower start |
| EKS Auto Mode (Dec 2024) | AWS manages node provisioning (Karpenter-based), core add-ons (CNI, CoreDNS, kube-proxy, LB controller, EBS CSI), OS patching | Teams wanting less undifferentiated ops | Less control; additional management fee |
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
A Fargate profile selects pods by namespace and labels; matching pods run on Fargate.
Node OS and AMIs
- Amazon Linux 2023 (EKS-optimised AL2023 AMIs; bootstrapping via
nodeadm) — the default today. Kubernetes 1.32 was the last version with Amazon Linux 2 EKS AMIs, and AL2 AMI support ended in November 2025. - Bottlerocket — minimal, container-optimised, immutable OS with atomic updates and a small attack surface.
- Graviton (arm64) instances (m7g, c7g, r7g) are typically ~20% cheaper for similar performance — requires multi-arch images.
- GPU AMIs for ML workloads.
Node IAM role
Every EC2 node needs an instance role with at least: AmazonEKSWorkerNodePolicy, AmazonEKS_CNI_Policy (better: give the CNI its own role via IRSA/Pod Identity and remove it from the node), AmazonEC2ContainerRegistryReadOnly (pull from ECR), and usually AmazonSSMManagedInstanceCore for Session Manager access. Keep it minimal — anything on the node role is reachable by any pod that can reach instance metadata (see security lesson).