Consider a synthetic workload: 95 requests take 40 ms each, and 5 take 800 ms. The mean weights every request equally; a percentile identifies a position in the sorted distribution.
A service that calls several dependencies can inherit the tail of each one. Deadlines, bounded retries, and queue limits should form one coherent budget.
This calculation uses the nearest-rank percentile: sort the observations and select rank ceil(p × n). With 100 requests, p99 is the 99th observation. Track percentiles alongside request volume and the measurement window; changing either can change how much evidence supports the reported tail.