CI feedback latency measures how long a submitted change waits for its evaluation result. Track it by percentile, not average: a good P50 can conceal a damaging P99 caused by burst contention. section 10.7.1
Use the metric to expose throughput limits:
- Compare P50, P95, and P99 to distinguish normal feedback from queueing outliers. section 10.7.1
- Investigate rising tail latency when multiple agents submit changes together; shared CI runners can create burst contention. section 10.7.1
- Add runner capacity when queueing dominates and verify that the tail improves rather than optimizing only the median. section 10.7.1
A hidden throughput wall
A system with a 2.3-minute P50 still had an 18-minute P99 because five agents submitted pull requests within one minute. Adding two CI runners reduced P99 to 4.1 minutes. section 10.7.1
