Metric Guide

What Is p95 Latency?

p95 latency shows the response time that 95% of requests met or beat during a test.

The short definition

p95 latency is a percentile response time. If a load test report says p95 latency is 420ms, it means 95% of measured requests finished in 420ms or less. The remaining 5% were slower.

This is useful because real users do not experience averages. Some users land on fast responses, while others hit cache misses, queueing, cold starts, database waits, or network delays. p95 helps show the slower end of normal experience without letting one extreme request dominate the whole story.

A simple example

Imagine 100 API requests sorted from fastest to slowest. The 95th request in that sorted list is the p95 value. If requests 1 through 95 are 300ms or faster and requests 96 through 100 take longer, p95 is about 300ms.

Example: reading p95 in a report

Suppose a small test sends 50 requests to your product search API. The average latency is 160ms, p50 is 120ms, p95 is 480ms, and p99 is 900ms. The average looks fine, but p95 shows that several users may still wait noticeably longer. That is a signal to inspect slow queries, cache misses, or upstream calls before launch.

Why average latency is not enough

Average latency can look healthy while a noticeable group of users still waits too long. Ten very slow requests can be hidden inside hundreds of quick ones, especially when a dashboard only shows the mean. p95 makes that tail behavior visible.

How to read p50, p95, and p99 together

p50: the middle request. This reflects a typical response when the system is behaving normally.
p95: the slower edge of common user experience. This is often a better launch-readiness signal.
p99: the slowest 1% of requests. This can reveal rare stalls, retries, and busy services.

What is a good p95 latency?

There is no universal number. A health check endpoint should usually be very fast. A complex report export may reasonably take longer. The better question is whether p95 is acceptable for the action users are taking and whether it stays consistent as traffic rises.

How to improve p95 latency

Find slow database queries and add the right indexes or query limits.
Reduce work done inside the request path, especially calls to slow external services.
Cache expensive reads where freshness rules allow it.
Check whether concurrency is causing queueing at the app server, database, or upstream API.
Compare reports before and after each change instead of relying on one isolated run.

p95 Latency FAQ

Is p95 better than average latency?

It is better for spotting slower user experiences. Average latency is still useful, but p95 shows whether most requests stay within an acceptable response time.

Can p95 be high when the site feels fast?

Yes. The typical request may be fast while a smaller group waits on slow database queries, cold starts, or external services.

How do I compare p95 between tests?

Use the same URL, traffic settings, request body, and headers. Then compare p95, error rate, and throughput across the saved reports.

Use p95 in context

p95 latency is most helpful when paired with error rate, timeout count, throughput, and status code breakdowns. For a fuller walkthrough, see how to read a load test report, then run a small controlled test to compare your own numbers.