Percentile

A percentile marks the point in a sorted dataset below which a given share of observations falls. The 90th percentile of page load times, for example, is the value that 90% of measured loads were faster than or equal to; the 50th percentile is the median.

Percentiles are computed by sorting the data and finding the value at the corresponding rank; different software packages use slightly different interpolation rules for values that fall between two observations, so the same dataset can report marginally different percentile values depending on the method used. Percentiles are closely related to quartiles and the interquartile range, which are just the 25th, 50th and 75th percentiles used as a group to describe spread.

Percentiles matter because they describe the shape of a distribution without being distorted by extreme values the way an average can be, which is why performance engineering reports latency as p50, p95 and p99 rather than a mean. The common pitfall is treating a high percentile like p99 as a rare edge case: at scale, a p99 latency affects one in a hundred requests, which can be a large absolute number of users when traffic is high.