Percentiles lie under load

A percentile is computed over completed requests. Requests that time out never enter the histogram, so a system shedding its slowest work will show an improving p99 as it degrades.

This is why saturation needs a direct signal — queue depth, utilisation, rejections per second — rather than an inferred one. See Little's Law for the relationship that makes queue depth predictive rather than merely descriptive.

Worked example in the article on connection pool saturation.