p99 latency

p99 latency is the 99th percentile of response times: the value that 99% of requests beat and 1% exceed, used to describe how slow a system is in its worst common cases.

Averages hide the slow tail. Google's site reliability engineering book shows a service whose typical request takes about 50 ms while 5% of requests are 20 times slower, a gap an average would mask. It recommends percentiles: the 50th for the typical case and a high percentile such as the 99th or 99.9th for a plausible worst case (Google, Site Reliability Engineering, chapter "Service Level Objectives").

In real-time bidding the tail decides revenue. An exchange drops any bid that misses its tmax deadline, so a bidder with a fast average but a slow p99 loses every auction in that slow tail. A bidder also calls several internal services per request, such as user lookup, budget check and model scoring, and the slowest of those sets its own tail. Teams therefore set targets on p99 and p99.9, alert on them and trace the requests that cause them.

Example: A bidder measures p50 at 18 ms, p95 at 45 ms and p99 at 130 ms on its own servers. An exchange sends "tmax": 120, and the network adds 25 ms each way. Every request slower than about 70 ms of compute now arrives late, which is far more than 1% of traffic.

Related terms: tmax, timeout, QPS, real-time bidding, bid request.

Authoritative source: Google, "Site Reliability Engineering", chapter 4 "Service Level Objectives" (accessed 9 November 2026).

Read the full guide: QPS, latency and bid performance.