Why measure p95 and p99 instead of the average?
Averages hide the slow requests that users actually remember. A system with a 50 ms average can have a 2 second 99th percentile. At scale one user request often fans out to many backend calls, so rare slow calls become common slow pages. Track percentiles and set targets on them.
Answered in
Reducing Latency: Find the Layer First, Then Pull the Right LeverNetwork, processing, database and rendering delay need different fixes. Caching, CDNs, indexes, async work, compression, connection reuse and tail latency.
Read the full analysisOther questions this article answers
More web performance questions
- What problem does HTTP/2 solve compared with HTTP/1.1?
- What is head-of-line blocking and does HTTP/2 remove it?
- Why does HTTP/3 use UDP?
- How can I check which HTTP version a site is using?
- What is the difference between latency, bandwidth and throughput?
- What is the quickest win for network latency?
- How do I decide whether to make something asynchronous?
- What is a cold start?
Every answer on Crashtech is written by the editor of the article it comes from — never auto-summarised. Browse all answers or the Web Performance beat.