A city needs instant permit-risk scores while clerks enter applications at the counter. Which deployment option should MLOps choose?

Deploying the model to a real-time managed online endpoint for interactive scoring is correct when clerks need instant permit-risk scores at the counter. Online endpoints provide low-latency, request/response inference suitable for interactive municipal workflows, with Azure ML managing hosting, scaling, and invocation APIs. A managed online endpoint scaling to zero for interactive scoring cannot meet counter-service timing because cold starts add latency when no instances are warm regardless of model quality.

Deploying the model to a managed online endpoint scaling to zero for interactive scoring fails because permit counseling happens in the moment; cold starts after scale-to-zero leave clerks waiting instead of returning instant scores. Deploying the model to an online endpoint in shadow mirror mode for interactive scoring fails conceptually: mirrored traffic is for silent testing and comparison, not for returning live scores to the counter application. Deploying the model to a container instance dev-test endpoint for interactive scoring fails because container-instance endpoints are for low-scale development and testing, not production counter volume; they lack the performance, scale, and SLA to return reliable instant scores while clerks wait.

Exam caveat: real-time endpoints need capacity planning, auth, and network controls appropriate for city systems; “online” does not mean publicly anonymous. Monitor latency and error rates during peak permit hours. Operational check: deploy to a managed online endpoint, authenticate a test client, measure p95 latency on representative permit payloads, and only then point the counter application at the scoring URL.