Cutting EKS Costs ~40% with Karpenter and Spot

TL;DR — Replaced the Cluster Autoscaler with Karpenter, enabled consolidation, and shifted stateless workloads to Spot. Result: roughly 40% lower monthly compute spend with no measurable impact on availability.

The Problem

The cluster ran on a handful of fixed managed node groups. Bin-packing was poor: nodes sat at 35–45% utilisation because the autoscaler could only add/remove whole nodes of pre-defined types, and it never re-packed existing workloads.

The Approach

Karpenter provisions right-sized nodes just-in-time based on actual pending-pod requirements, and — crucially — consolidates: it actively moves pods onto fewer nodes and terminates the emptied ones.

apiVersion: karpenter.sh/v1
kind: NodePool
metadata: { name: default }
spec:
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 1m
  template:
    spec:
      requirements:
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot", "on-demand"]
        - key: kubernetes.io/arch
          operator: In
          values: ["amd64", "arm64"]

Stateless, retry-tolerant workloads were allowed onto Spot; stateful and singleton workloads stayed on-demand via nodeAffinity.

Validation

• PodDisruptionBudgets on every important deployment so consolidation and Spot reclaims never take down too many replicas at once.
• Watched the Karpenter metrics and node-count graphs for a week before trusting it in production.

The Result

Utilisation climbed past 70%, on-demand spend fell as Spot absorbed the bulk of stateless capacity, and the total compute bill dropped by about 40%. Availability SLOs held because disruption budgets and on-demand fallbacks kept the critical path safe.

---

This is example content — swap in your own figures and screenshots. The folder src/content/blog/ accepts any new Markdown file and it appears automatically.