Percentiles over averages, again, with a graph

An average response time of 180 ms describing a distribution where a fifth of requests take over a second.

the same endpoint, one hour:

  mean     180ms
  p50       94ms
  p75      120ms
  p90      410ms
  p95      880ms
  p99    2,140ms

the mean sits between p75 and p90, describes nothing,
and moves when either tail moves.

what the p99 was: requests from customers with more
than 200 orders. 3% of customers, and the ones who
pay the most.

The correlation between the slow tail and the valuable customers is the argument that ends the discussion, and it is available in any tracing system by grouping the slowest traces. An average optimises for the median user by construction; the p99 is where the people are who will leave.