Score how noisy an alerting setup is from false-positive and repeat-alert ratios.
Noise score = (false_positives + auto_resolved) / total_alerts × 100. Both categories represent alerts that didn't require the human response an alert implies: false positives never reflected a real problem, and auto-resolved alerts fixed themselves before anyone acted. A high noise score is the leading driver of alert fatigue — responders start ignoring or delaying response to a channel that's frequently wrong or self-resolving.
noise_score = (non_actionable_alerts / total_alerts) × 100
It's rarely achievable in practice — some legitimate transient issues will always auto-resolve before action is needed. Most well-tuned alerting systems land in the 5-15% range rather than exactly zero.
A false positive never corresponded to a real problem at all (a monitoring bug, a bad threshold, a flapping check), while an auto-resolved alert reflected a real but transient condition that fixed itself — both are noise from a responder's perspective, but they call for different fixes.
Increase the `for:` duration so transient conditions don't cross the firing threshold, or add a secondary condition requiring sustained severity, so only conditions that would matter to a human actually notify one.
Audit alert rules against actual incident postmortems — if an alert rule has fired many times with zero corresponding incidents, either its threshold is miscalibrated or it's not actually measuring something that matters.