# Watching for an outage

A watch is not a load test. It samples a request on a fixed interval with a short timeout, and answers one question: how long was the service unreachable?

It is built for tuning high availability. Start a watch, then pull a node, fail over the database or roll out a release, and the outage log tells you what each step actually cost your clients. Change the setup and run it again to see whether the gap got shorter.

![An outage watch with one outage of 7.3 seconds](watch.png)

## Starting a watch

1. Open **Performance** and pick the requests to watch, for example a health endpoint and a login.
2. Choose the **Watch** profile.
3. Set the interval and the timeout. The timeout is required: with a long one, a failover arrives as one slow request instead of an outage.
4. Leave the duration empty to watch until you stop it.

Each request can have its own interval, so a login checked every few seconds and a health endpoint checked twice a second can share one run.

## The outage log

Every outage is listed from its first failed sample to its last, with its duration, the number of failed samples and the reason. The log is exact: a two second outage inside a twenty minute watch is never rounded away.

An outage is known only to within one interval, and gurl says so beside the numbers.