# Latency, p50, p95 and p99

Latency is how long a call takes. This guide explains percentiles, why an average can hide slow calls and how to interpret the latency figures in Service Health.

## What is a percentile?

Imagine sorting 100 call durations from fastest to slowest:

| Measure | What it describes |
| --- | --- |
| p50, the median | The duration around the middle: half the calls finish within it |
| p95 | The duration within which about 95 of the 100 calls finish |
| p99 | The duration within which about 99 of the 100 calls finish |

If most calls take 100 ms but a few take several seconds, the average can look reasonable while those slow calls are painful. Higher percentiles describe this slow end of the distribution, often called the **tail**.

For example, p50 of 100 ms and p99 of 2 seconds suggests a very different experience for the slowest callers. A customer making many calls may encounter that tail regularly.

## What the app can and cannot infer

Service Health reads source percentiles. When several request names become one API, or several APIs become a service, it combines their percentiles using a call-weighted mean. Busier groups contribute more.

That combined p95 is **not an exact p95 of all the underlying requests**. Percentiles cannot generally be merged by averaging. Use these figures to locate changes and compare like with like; do not infer an exact slow-call count from an aggregated percentile.

The speed objective avoids that problem: it uses the counted calls exceeding the chosen slow line directly. [SLOs](/docs/service-health/slos#speed) explains the distinction.

## How to investigate a slowdown

1. Open the affected service, then its API.
2. Check whether speed budget started falling at the same time as latency increased.
3. Open the day to inspect five-minute p95 and p99 points, if they were retained.
4. Look at **What it waits on** to separate changes in the API from changes in its dependencies.
5. Check your source telemetry for the requests behind that time before deciding on the cause.

The app's sustained-slowdown marker looks for p95 over the slow line for 30 minutes. It is a diagnostic clue, separate from the objective's count of slow calls.

Five-minute detail covers at most the last 14 days and only the APIs selected by your connector setting. A chart with daily totals cannot recreate detail that was not kept.

## What good looks like

Compare a service with its own history and the response time its users need. A batch job and an interactive checkout do not need identical latency. Long-lived streams and recognised background work are judged on success only, rather than against an interactive speed target.

## Learn more

- [Dependencies and waits](/docs/service-health/dependencies)
- [Google SRE book: Monitoring distributed systems](https://sre.google/sre-book/monitoring-distributed-systems/)