The mean means nothing: data visualization to debug a latency problem | Hacker News

A Hacker News discussion highlights a data-visualization approach for diagnosing latency issues and warns that averages and medians can hide serious outliers.

In distributed systems, parallel requests and fan-out across many servers can make a few slow requests dominate service-level latency.

Citing Jeff Dean, the post notes that tail latency can determine user experience even when individual servers look fast.

Engineers should rely on percentiles and latency distributions, not just mean or median, to catch real performance problems.