Metric and resource experiments
From August to October 2026, metric work relied on written measurement contracts rather than dashboards or favorable snapshots. A contract made it possible to tell a real change from workload mix, scrape gaps, restarts, or instrumentation limits.
What a contract named
| Part | Fields |
|---|---|
| Purpose | Beneficiary, scope, editable levers, authority |
| Measurement | Numerator and denominator, start and stop events, window, query, sample count |
| Judgment | Baseline, target, noise tolerance, independent guards |
| Safety | Rollback and a result record |
What we learned
Early measurements exposed metric semantics that needed correcting:
- Restart estimates can be confounded by OOM kills.
- Counters reset.
- Aggregate snapshots cannot establish cause.
Record corrections next to the result instead of carrying the superseded reading forward.
Resource requests were right-sized from a bounded window of CPU and working-set data, leaving limits and replica counts unchanged. Lower requests change what the scheduler reserves. They do not change actual consumption or power draw.
Using these results
The records are dated and say nothing about current utilization. To compare, repeat the measurement under its original contract or version the contract explicitly.