A production-focused guide to deploying Karpenter v1.13.0 on Amazon EKS, covering the full setup from IAM configuration via CloudFormation, Pod Identity binding, Helm installation, EC2NodeClass and NodePool configuration, to smoke testing. Key topics include Karpenter's direct EC2 API provisioning (45–60s vs. 3–4 min for Cluster Autoscaler), the v1 API changes (NodePool replacing Provisioner, EC2NodeClass replacing AWSNodeTemplate), Spot interruption handling via SQS/EventBridge, disruption budgets, Prometheus metrics to monitor, and common gotchas with PodDisruptionBudgets and StatefulSets with EBS volumes. The guide also covers safe rollback procedures and acknowledges Karpenter's limitation: it cannot fix overprovisioned pod requests, which the author's company (Cast AI) addresses with a complementary product.
