A city fraud team is registering grant-screening model v3 while grant season still may require a fast return to v2. How should versions be handled?

Registering grant-screening model v3 while keeping v2 available in the registry so staged rollback remains possible is correct because grant season often demands fast reversion if v3 mis-scores or breaks integrations. Azure ML model versioning exists so municipal fraud/grant teams can promote forward without destroying the prior known-good artifact. Side-by-side versions enable controlled cutover, comparison, and rollback without scavenger hunts outside the registry.

Registering v3 as a new version while keeping v2 assigned to the staging endpoint for evaluation fails because staging needs to validate the v3 candidate, not continue evaluating the old v2, so v3 would advance without proper testing. Registering v3 under a separate model name while keeping v2 available in the registry for rollback fails because it fragments lineage and comparison; versions must live under the same registered model to enable controlled promotion and a clean, staged rollback to v2. Registering v3 while keeping v2 available in the registry and routing production scoring to v3 fails because it pushes unvalidated v3 straight to production instead of staging it first, removing the controlled cutover that grant season requires before shifting live traffic.

Exam caveat: keeping v2 available does not mean both should serve production traffic indefinitely—use stage labels and traffic splitting deliberately. Ensure rollback runbooks name the exact prior version and validation checks. Operational check: register v3 as a new version, leave v2 registered, run staging tests on v3, define a rollback step that re-points the endpoint to v2, and only then advance grant-season traffic.