Cloud bills rarely explode because of one bad decision. They drift upward through a hundred reasonable-at-the-time defaults: generous requests, nodes that never scale down, forgotten environments, and logs nobody reads but everyone pays to store. The good news: most of it is recoverable without a migration.
Right-size requests and limits first
The single biggest source of Kubernetes waste is CPU and memory requests set far above real usage, which forces the scheduler to reserve capacity nobody uses. Measure actual usage over a representative window, then set requests to reality plus headroom - not to a guess made on day one.
The checklist
- Right-size requests/limits from real usage data, not defaults
- Enable cluster autoscaling and make sure nodes actually scale down
- Use spot/preemptible nodes for stateless and batch workloads
- Kill zombie environments - preview namespaces that outlived their PR
- Move cold logs and metrics to cheaper retention tiers
- Consolidate under-utilized clusters; every control plane has a cost
- Put a cost dashboard in front of the teams who create the spend
Make cost visible to the people who cause it
FinOps is not a spreadsheet exercise done once a quarter by finance. It works when the team deploying a service can see what that service costs, in a dashboard next to its latency and error rate. Visibility changes behavior faster than any policy.
You cannot optimize what you cannot see - and you will not sustain what only one person watches.
A Cloud Cost & Security Review typically finds the first double-digit percentage of savings in the first few days, with a prioritized list your team can execute. The point is control, not a one-time cut that drifts back up.