1 min
Paavo Renk
Why we stopped quoting average latency
Averages hide the slow runs your users remember. We publish p50, p95 and p99, and here is why.
An average of 300 ms can mean every run takes 300 ms. It can also mean most runs take 150 ms and one in twenty takes three seconds.
The slow run is the one users remember
Nobody writes in about a fast run. They write in about the one that hung. So we stopped publishing averages and show three numbers instead: p50, p95 and p99.
If the p99 is close to the p50, the system is calm. If it is far away, something is waiting in line.