# Alert tuning

Symptom: alert fatigue (ignored pages) or silent incidents (no alert fired).

## Diagnose

List current alert configs (Atlas UI, Alerts, Alert Settings). For each firing alert ask: did a human need to act? For each recent incident ask: did any alert fire?

## Confirm

Check the metric baseline first. A connections alert at a fixed number is meaningless without knowing your normal peak. Pull a week of the metric and set thresholds relative to it.

## Fix

- Thresholds from baseline: e.g. connections above 80% of tier cap sustained 5 minutes, disk above 75%, replication lag above your SLA.
- Route by severity: page for data-loss risks (replication, backup failures), notify chat for early warnings (disk 75%, slow query uptick).
- Maintenance windows during deploys and index builds so planned work does not page.
- Every alert needs a runbook link: what to check first. An alert nobody knows how to answer is noise.

## Verify

Run a controlled trigger (a staging load test that crosses a threshold) and confirm the right channel fires with the runbook attached. Review quarterly.