Skip to lesson
Exit
GPU Economics & Capacity Planning1 / 2

1 min lesson

Latency is a distribution, not an average

Give a plain answer to "Why track latency as a distribution rather than an average?" Then ground it in one lesson detail.

Step 1 of 2

Latency is a distribution, not an averagethe tail is what users feel

An average hides the requests that actually hurt. If p50 is 40ms but p99 is 600ms, one in a hundred completions feels broken and a developer hits that tail many times an hour. Manage latency as percentiles and write SLAs against the tail.

Reading a latency distribution
p50 (median)
The typical experience; good for capacity math, useless as an SLA
p95
Where most SLAs live; the bad-but-common case
p99 / p99.9
The tail - rare per request, frequent across a day of use; this is what users remember

Batching that lifts throughput often fattens the tail - watch p99, not just the mean.