Cutting at the Median

A dependency measured over the last month has p50 latency of 20ms, p99 of 120ms, and p99.9 of 900ms. A new client ships with a per-call timeout of 25ms plus two retries. What is the most likely effect in production?

The two retries make it safe, since three independent chances at 25ms give roughly the success rate of one 120ms timeout

Roughly half of all calls abort and retry even though the dependency is healthy, multiplying its load while errors still reach users

Only the slowest tenth of a percent is affected, since 25ms sits between the measured p99 and p99.9, so the tail is trimmed cheaply

The dependency gets faster overall, because each aborted call frees its server-side worker at once and returns the connection to the pool