Runbook: DecisionModelRolloutFailures¶
What it means¶
One or more rollouts ended in rolled_back, rolled_back_after_promotion, or failed
in the last 15 minutes (decisionmodel_rollouts_total).
How to confirm¶
kubectl get dm -A
kubectl describe dm <name> -n <ns> # failedRevision, Ready/Degraded conditions, Events
rolled_back means a candidate failed but a stable revision kept serving;
rolled_back_after_promotion means a just-promoted revision was reverted;
failed means a candidate failed with no stable revision to fall back to (user-facing outage risk).
Common causes¶
- Bad candidate image/model (digest/device mismatch) — see DecisionModelModelMismatch.
- Eval-gated promotion rejected the candidate (accuracy below
minAccuracyor a drop beyondmaxAccuracyDrop) — see DecisionModelEvaluationPoor. - Prefetch failing permanently (missing tag, digest mismatch, or an upstream tag move).
Fix¶
Inspect status.failedRevision and the Evaluated/ModelReady conditions to see which
gate rejected the candidate. Fix the model/tag/dataset and re-apply. For a failed
(no stable) case, prioritise: there is no healthy revision serving.