An API's average latency is stable, but premium users report very slow requests. Which signal best tests whether a small fraction of requests is slow?

1. A request-duration distribution with upper quantiles
2. The median request duration alone
3. The total request counter
4. The number of active application instances