Definition
Tail latency is the latency experienced at the slow end of a system's distribution — the 95th or 99th percentile (p95/p99) rather than the average. A service can average 40ms while its p99 is 2 seconds, and the average hides it completely.
The tail matters because of fan-out amplification: modern pages touch many backend services, and if one user action makes 100 backend calls, the probability of avoiding every 1% tail is 0.99^100 — roughly 37% — so about 63% of user actions experience at least one p99-slow call. This is why latency objectives are stated in percentiles, and why tail-specific techniques exist: request hedging (send a duplicate after a delay, take the first response), aggressive timeouts, and load shedding.
How it comes up in interviews
Quoting average latency in a design review invites the correction; quote percentiles unprompted. Interviewers probe whether you can explain the fan-out math and name a mitigation — hedged requests and per-call timeouts are the expected vocabulary.