Blog

p50, p95, p99: why percentile latency matters for APIs

Averages hide tail latency; percentiles show user-visible slowness and noisy neighbors.

1 min read
observabilitylatencysre

Means vs tails

The average can look healthy while a fraction of requests time out—percentiles expose that tail.

Sliding windows or histograms in your gateway make regressions obvious after deploys.

Actionable alerts

Alert on p95 crossing SLO for sustained periods, not single spikes, to reduce noise.

Correlate with rate-limit denials and upstream 5xx to find capacity issues.

Back to blog