Kubernetes cost rarely explodes from one bad decision - it drifts up through generous requests, nodes that never scale down and preview environments that outlived their pull request. FinOps is the practice of making that visible and reversible.
Attribute cost with labels
You cannot bill back what you cannot attribute. Enforce team/owner labels so tools like OpenCost or Kubecost can split the bill.
metadata:
labels:
team: payments
cost-center: "4021"
env: prodTurn on the cheap wins
- Right-size requests from real usage (p95), not day-one guesses
- Enable cluster autoscaling and verify nodes actually scale down
- Use spot/preemptible nodes for stateless and batch workloads
- Garbage-collect preview namespaces when their PR closes
# idle cost: requested but unused CPU across the cluster
sum(kube_pod_container_resource_requests{resource="cpu"}) - sum(rate(container_cpu_usage_seconds_total[1h]))You cannot optimize what you cannot see - and you will not sustain what only one person watches.
Put a cost panel next to latency and error rate on each team's dashboard. Visibility changes behavior faster than any top-down policy.