Percentiles per consumer, not per endpoint

An endpoint with a healthy p95 and one consumer for whom it was consistently slow.

GET /api/orders, one week:

  all consumers    p50 40ms   p95 210ms

  consumer A       p50 38ms   p95 180ms
  consumer B       p50 41ms   p95 190ms
  consumer C       p50 44ms   p95 220ms
  consumer D       p50 890ms  p95 4,100ms

D is 3% of traffic and requests include=lines,
refunds,shipments.tracking on a 200-item page.

Aggregating across consumers averages away exactly the case that generates support tickets, and the fix here was not performance work — it was a conversation with one integrator about pagination. Labelling the metric by API key costs one label and makes this visible without any investigation.