Skip to content
BENIBIX Learn

Tail Latency

Infrastructure📈 Monitoring & ObservabilityTelemetry Signals

Also known as p99 Latency, p95 Latency, 99th Percentile Latency, Latency Percentile, p99

How slow the slowest few requests are, often read as p99.

Think of it like

Think of it like the last kids out of school. Most leave fast, but someone still waits at the gate for the slow ones.

Example

Lucy's API has a median of 80 milliseconds, but its p99 is 2 seconds. One request in a hundred takes 25 times longer than usual.

Why it matters

Your busiest users make the most requests, so they meet the slow tail most often. The average makes things look better than they are.

Definition

Tail latency is how slow the slowest requests to a service are. It is read as a percentile: p99 latency is the time that 99 out of 100 requests beat. The average hides these slow requests, because a few long waits barely move it. When one page calls many services, the slowest of them sets the speed. So a rare slow request in each service becomes a common one for the page.
In the bottom bar, most requests are near the median. The p99 mark far to the right is the tail.

Common mistake

Averaging the p99 from different servers. That is not the real p99, so work it out from all the timings together.

Key takeaways

  1. 1
    The slow end. Not the typical request.
  2. 2
    Read as a percentile. 99 in 100 requests beat the p99.
  3. 3
    Grows with fan-out. Many calls make slow ones common.