16.4p50 vs p95 vs p99

Suppose response times:

most requests:
300 ms

occasional requests:
8 sec

Average may look:

700 ms

which hides the problem.

Instead:

p50 = 300 ms
p95 = 2 sec
p99 = 8 sec

This tells you:

1% of users are having a terrible experience.

For integrations, tail latency often matters a lot.