P99 latency.
also: P99 · p99 · tail latency · 99th percentile · percentile latencyThe latency 99% of requests beat: percentiles describe the slow tail that averages hide.
the two-minute refresher
- 950 requests at 50 ms and one at 5 s: the average looks fine, the tail doesn't.
- P50 is the median; P95 and P99 show what your slowest users feel.
- The tail comes from GC pauses, cold caches, network congestion and slow dependencies.
- Set timeouts and SLOs from percentiles, never from the average.