Product & Analytics / Performance & Reliability

Percentile Latency p50 / p95 / p99 (also "tail latency", "response-time percentiles")

The response time below which a given share of requests fall -- a tail-aware read on speed, unlike an average.

The fork, why two teams get different numbers

WHICH percentile and OF WHAT population.

In the Metric Library

The full fork, both definitions worked all the way through
The trap that makes the number lie, on a real export
Every formula variant, spelled out
The reconciliation anchor, what to tie it to and when to refuse

Reference: Site-reliability / performance-engineering convention (no governing body); t-digest and HDR-histogram percentile-merging methods cited as reference