Model Rollback & Recovery¶

This runbook describes how to recover from model-related incidents.

When to rollback a model¶

• performance regression detected,
• data or prediction drift alerts,
• increased error rates or latency,
• unexpected business behavior.

Rollback strategy¶

Model rollback is performed via MLflow Model Registry.

Steps: 1. Identify last known good model version. 2. Update model alias or stage (e.g., Production). 3. Redeploy serving service if required.

No retraining is required.

Verification¶

After rollback: - verify active model version in logs and dashboards, - monitor latency and error rate, - confirm alert resolution.

Escalation¶

If rollback does not resolve the issue: - disable affected inference endpoints, - fall back to degraded mode if available, - escalate to investigation.