25 — Governance & Trust

Incident Response for ML#

Operations✓ Mathematical
◆ The PatternWhen models fail in production — rollback, fallback, postmortem

ML incidents are different from software bugs: the code runs fine, but predictions are wrong. Rollback reverts to the previous model version. Fallbacks (rule-based defaults, cached predictions) serve something when the model is down. Circuit breakers automatically switch to fallback when error rates spike.

// Interactive — incident response decision tree
SeverityResponseTimeline
P0 — Model serving errorsRollback immediatelyMinutes
P1 — Accuracy degradationSwitch to fallback, investigateHours
P2 — Slight drift detectedSchedule retrainingDays
P3 — Feature quality warningMonitor and logNext sprint
Blameless postmortems: After every incident, document what happened, why detection was delayed, and what systemic fix prevents recurrence. Blame the system, not the person.
Pattern bridge: ML incident response mirrors stop-loss discipline in trading — predefined rules that limit damage when things go wrong. The power analysis framework helps design monitoring that catches problems early enough to act.
← Previous
Reproducibility
Open in the full reader, with the topic sidebar →