Why Average Latency Misleads Engineers and What p90, p95, p99 Actually Mean

Average latency metrics can mask serious performance problems because slow outlier requests get diluted by the majority of fast ones, making dashboards look healthy while some users experience significant delays. Percentile metrics like p90, p95, and p99 offer a clearer picture by describing the experience of specific portions of traffic rather than a single blended figure. For example, a p99 of 3,100ms means one in every hundred requests takes over three seconds, a fact that an average would never reveal. Calculating these percentiles requires no advanced statistics — response times are simply sorted and the value at the corresponding position in the list is read off. Engineers are advised to track all three percentiles together, since p90 reflects typical performance while p99 exposes worst-case conditions on an ordinary day.