Engineering, production, governance, cost, communication · insight-0271

Keep rollback one alias change away and test it before you need it

Situation
A new model version misbehaves, and rollback requires rebuilding the old container and feature pipeline.
Action
Keep the previous model artifact, its feature version and its serving config deployable; point production at a registry alias; and rehearse switching the alias back as part of every release checklist.
Reason
A mutable alias that names the serving version turns rollback into a pointer change, provided the old version's dependencies still exist.
Boundary
Any model served through a registry or versioned endpoint.

Symptom or risk

Rollback takes hours, or fails because the previous model expects features that were removed.

Changes that migrate stored state (embeddings in an index, database schema) need a separate reversible migration plan.

Keeping old feature pipelines alive costs maintenance.

Deleting deprecated features in the same release that changes the model.

MLflow model aliases are mutable named references to a model version, so reassigning an alias changes which version a deployment uses.
ML Model Registry (MLflow documentation)