Maintenance windows, done right

Every team has a deploy ritual, and most of them involve someone saying "ignore the alerts for the next twenty minutes." That works until the deploy overlaps a real failure — then the ignored alert was the important one, and nobody can tell afterward which downtime was planned and which wasn't.
The deploy problem
Monitors can't distinguish a restart from an outage. A rolling deploy that drops a check for ninety seconds looks exactly like a ninety-second outage, so it pages whoever's on call, opens an incident, and paints a red bar on the status page. Multiply by weekly deploys and your incident history becomes noise: customers learn to ignore the status page, and the team learns to ignore the alerts. Both lessons are expensive the one time they shouldn't apply.
Pause lies, windows tell the truth
The instinct is to pause the monitor during the deploy. Don't — pausing stops the checks, so the gap in history is indistinguishable from missing data. A maintenance window keeps the checks running and marks the span as planned instead:
- Checks continue, so you still see response times through the deploy.
- Incidents and notifications stay quiet for the window's duration.
- The status page shows the window as the reason, not a mysterious gap.
- When the window ends, everything resumes automatically — nothing to unpause.
Honest history compounds. Six months later, "planned window for the Postgres 16 upgrade" next to a yellow bar is an explanation; a gray gap is a question nobody remembers the answer to.
Setting one up
Pick the monitors the deploy touches, set a start and end with a little buffer on both sides, and write the reason the way you'd explain it to a customer — because the status page will show exactly that. If the deploy runs long, extend the window rather than letting it lapse; a window that expires mid-deploy pages someone for doing their job.
