A water utility is registering a quality-prediction MLflow model and needs reliable scoring contracts for downstream systems. What should registration include?

Including schema and signature metadata when registering the MLflow model is correct because water-utility quality-prediction scoring must declare input and output contracts for downstream SCADA-adjacent or analytics systems. MLflow model signatures (and related schema metadata) let Azure ML and consumers validate payloads, catch breaking changes early, and document types/shapes expected at inference. Weights without a contract produce fragile integrations and silent mis-scores when field names or dtypes shift.

Attaching Unity Catalog lineage and a training dataset reference to the version fails because lineage and dataset pointers do not declare an enforceable input-output scoring contract for downstream callers. Registering the pyfunc flavor and production alias for model serving fails conceptually because those settings control serving deployment and routing, not the machine-validated input and output schema that scoring contracts require. Logging evaluation metrics and parameter history with the registered version fails because metrics and parameter records describe training results, not the field types, shapes, and payload structure that downstream systems need to validate at scoring time.

Exam caveat: a signature must match the real feature pipeline—including types, optional fields, and tensor shapes—or validation will reject legitimate traffic or accept bad traffic. Update signature and model version together when features change. Operational check: infer or define the MLflow signature from training examples, register it with the model, send a schema-valid and a schema-invalid test request, and confirm downstream systems read the registered contract before cutover.