Playbook · 8 chapters · Free

Kubernetes Cost Optimization Playbook

The same playbook we run in client cost audits — the steps, tools, and decision frameworks that typically recover 30-50% of Kubernetes compute spend without touching application code or accepting worse reliability.

What's inside: real utilization measurement, VPA rollout without breaking things, Karpenter migration, spot/reserved mix sizing, non-prod scale-down, orphan cleanup, namespace cost attribution, and governance policies that make savings stick.

Contents

• Step 01

  Measure before you cut

  How to extract real utilization data — requested vs. actual CPU and memory, p95 not average — from Prometheus, kubectl top, or your cloud provider's cost explorer. The single metric that tells you how big your leak is before you spend a dollar on tooling.

• Step 02

  Rightsize with VPA (without breaking things)

  Deploy the Vertical Pod Autoscaler in recommendation mode first — not auto. Two weeks of observation, then batch rollouts starting with highest-spend, lowest-risk services. Most teams safely cut requests by 30-50% without a single incident.

• Step 03

  Replace Cluster Autoscaler with Karpenter

  Why default autoscaler behavior spreads pods across half-empty nodes. Karpenter provisions right-sized nodes, consolidates aggressively, and supports mixed instance types natively — teams typically see node count drop 20-35% with the same workload.

• Step 04

  Spot, Reserved, and Savings Plans — the right mix

  A decision matrix by workload class: spot for stateless services and CI runners, on-demand for stateful and latency-sensitive workloads, Reserved or Savings Plans for baseline capacity. Concrete formulas for sizing the commitment.

• Step 05

  Non-production environments

  Dev and staging clusters running 24/7 at production size are one of the most fixable leaks. Scheduled scale-down patterns (CronJob or Kube-downscaler) that recover 60-70% of non-prod compute cost — roughly 5-8% of total spend.

• Step 06

  Storage and load balancer cleanup

  PersistentVolumeClaims and LoadBalancer-type Services that outlive the workloads that created them. A monthly automated sweep script and the tagging strategy that prevents orphans from accumulating.

• Step 07

  Namespace-level cost attribution

  How to allocate cloud spend to teams, products, or customers using Kubernetes labels, Prometheus metrics, and tools like Kubecost or OpenCost. The tagging taxonomy that makes FinOps reporting work without manual allocation.

• Step 08

  Governance — making it stick

  Admission controllers (Kyverno or OPA Gatekeeper) that enforce resource requests and limits, prevent unbounded deployments, and require cost labels. The policies that keep spend from creeping back after the initial optimization.