Tail Latency
also called p99, p999
The latency experienced by the slowest small percentage of requests, which is what users and dependent services actually feel.
Averages are close to useless for latency because the distribution is not normal — it has a long right tail produced by garbage collection, cache misses, lock contention, retries and noisy neighbours.
Tail latency compounds in distributed systems: if a page requires ten backend calls and each has a p99 of one second, roughly one in ten page loads waits a second. This is tail amplification, and it is why p99 matters far more in a service architecture than in a monolith.
The mitigations are specific: hedged requests, where a second copy is sent after a short delay and the first response wins; reducing fan-out; and setting timeouts from the tail rather than the mean, so a slow call fails fast enough to retry within the deadline.