Question 1044

Averages Hide the Fire

Your dashboard graphs per shard CPU and per shard request rate, and both look normal, yet clients keep reporting p99 timeouts against one shard. What instrumentation would actually name the hot key?

A

Per shard p99 latency alongside the mean, which correctly identifies the shard whose tail latency has degraded

B

Sampled per key request counters fed into a top-K heavy hitter sketch, so the busiest individual keys are reported by name

C

A much finer scrape interval on the same per shard counters, since second-level resolution will surface the spike that minute averages hide

D

Distributed tracing across the whole request path, so you can see which upstream service is slow before the shard even responds