Decoder. plain-English AI glossary

Tail latency

● Core

The latency of the slowest requests (e.g., p99 or p95) — much more important to user experience than average latency.

Think of it like

One slow checkout line ruins the shopping trip even if 99 others are fast.

Example

Average latency is 50ms, but p99 is 500ms — some users wait 10x longer, which feels broken.

How it actually works

Average latency can be misleading. If 99% of requests are 50ms and 1% are 500ms, the average looks good but users in that tail suffer. Tail latency is caused by GC pauses, OS scheduling, cache misses, or congestion. Reducing tail latency often means: better hardware isolation, fast autoscaling, or shedding load.

For product teams

P99 latency determines user experience — focus on it, not average.

For engineers

Profile and eliminate sources of variance; prioritize speed over average efficiency.

Related

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open DecoderAdd to Chrome