Karpenter replaces Kubernetes Cluster Autoscaler with intelligent bin-packing and just-in-time node provisioning. This guide covers setup, consolidation, cost optimization, and production patterns for EKS clusters.
Autoscaling was supposed to make costs predictable by matching capacity to demand. Instead, it introduced feedback loops, burst amplification, and — with AI workloads — a new class of non-deterministic spend that no scaling policy anticipates.
AWS surprise bills from autoscaling follow a small set of repeatable failure patterns: feedback loops, scale-out without scale-in, burst amplification from misconfigured metrics, and commitment mismatches after scaling events. Each pattern has a specific fix.