Why do we measure latency using percentiles instead of averages?

Question

Accepted Answer

An average hides the worst experiences. If 99 requests take 10 ms and one takes 2 seconds, the average is about 30 ms - which describes nobody.Percentiles describe the distribution:p50 (median) - the typical request.p95 - the slowest 1 in 20.p99 - the slowest 1 in 100, where timeouts and retries live.p99.9 - matters most for your highest value users, who make the most requests and so hit the tail most often.A page that makes 100 backend calls will, on average, hit its p99 backend latency at least once - so tail latency becomes typical latency at the page level. Set SLOs on p95 or p99, never on the mean.

Why do we measure latency using percentiles instead of averages?

Answer with Explanation