Traditional alerting on static thresholds fails in a reliability programme because it treats every threshold crossing identically regardless of severity or speed of error budget consumption. An alert that fires when error rate exceeds 1% catches both a catastrophic outage consuming the entire monthly budget in two hours and a minor sustained degradation that would take three months to exhaust it — yet these two situations require fundamentally different responses.
Burn rate is the concept that resolves this problem: it measures not just whether the error rate is above the SLO threshold, but how fast the error budget is being consumed relative to the rate that would exhaust it exactly over the compliance window. The multi-window multi-burn-rate model, developed by Google and published in the SRE Workbook, uses combinations of observation windows and burn rate thresholds to catch both rapid severe failures and slow subtle ones while minimising false positives.
Understanding and implementing this model is the difference between an alerting programme that on-call engineers trust and respond to and one they habituate to ignore because it fires on noise as readily as on genuine signal.