2 / 6 Counters, Top-K & Sketches

Wrong Kind Of Average

Forty API servers each compute a local p99 latency every minute, and the dashboard shows the arithmetic mean of those 40 numbers as the fleet p99. Users report far worse tail latency than the dashboard admits. What is wrong and what fixes it?

A

The servers take uneven traffic, so weighting each local p99 by that server's request count makes it exact

B

Quantiles do not average, so merge a quantile sketch from each server and query that

C

The window is too short for a stable p99, so computing each local p99 over an hour makes the mean correct

D

The mean is wrong for a tail metric, so the median of the 40 local p99 values is the real fleet p99