Monitoring tools are judged by their false alarms. This is how we avoid them.
One check
A check is one request from one of our nodes, with the full timing breakdown: DNS lookup, TCP connect, TLS handshake, time to first byte, total. We follow redirects, record the final address and, for HTTP monitors, run your content checks.
Retry on the same node
When a check fails the monitor becomes suspected down – visible in the app, no alert. The same node retries after 10 seconds and again after 20.
Confirm from another country
If the retries also fail, a node in a different country runs the check. Only when it fails too is the outage confirmed: an incident opens and alerts go out. If the second node succeeds, it was a local problem – we note it and go back to up.
Recovery
Two consecutive successful checks close the incident and send the recovery message with the duration. If the site fails again within 30 minutes we reopen the same incident instead of alerting again (flapping).
Degraded
A slow but successful response is degraded: visible, recorded, never paged. On Pro we also compare to the monitor's own rolling p95.