Computer Science › Math for Programmers · also in Observability, Performance & Scalability

Percentiles (p50, p95, p99)

The value below which a given share of measurements fall; the right way to read latency.

Also known as: percentiles, p50, p95, p99, quantile, median latency

A percentile answers: what value do X% of the measurements fall below?

• p50 (the median): half of the requests were faster than this.
• p95: 95% were faster; the slowest 5% were slower.
• p99: 99% were faster; the slowest 1% were slower.

Say a service’s response times are: p50 = 80 ms, p95 = 300 ms, p99 = 1.8 s. Half the users wait up to 80 ms, but one in a hundred waits nearly two seconds.

Why not just use the average?

An average hides the slow cases. Take 100 requests where 99 take 50 ms and one takes 5 seconds:

times = [50] * 99 + [5000]
sum(times) / len(times)      # 99.5 ms: looks fine
# p99 shows the 5-second request; the average does not

Averages mix fast and slow into a number nobody actually experienced. Percentiles show what real users see, and especially the tail, the slowest few percent (tail latency). Those are often your most active users, since users who make more requests are more likely to hit a slow one.

Computing them

import numpy as np
np.percentile(times, [50, 95, 99])

or statistics.quantiles(times, n=100) in the standard library.

Rules of thumb

• Report p50 and a high percentile (p95 or p99) for anything users wait on.
• You can’t average percentiles. The average of two servers’ p99 values isn’t the combined p99. Aggregate the raw data, or histogram buckets, then compute it.
• Need enough samples. A p99 from 50 data points is mostly noise.
• Service goals (SLOs) are typically written in percentiles: “99% of requests under 300 ms”.
• The same idea applies beyond latency: pipeline run times, file sizes, query durations.

See also latency and statistics for engineers.