MLflow, registry, and promotion | Databricks Machine Learning Associate: training and operations

Concept and mechanism

An MLflow experiment groups related runs. Each run can record parameters, metrics, and artifacts, such as tree depth, validation error, and an analysis report. Autologging helps with supported integrations, but a custom operational metric may require explicit logging. Compare runs evaluated on a compatible basis; ordering scores from different populations does not demonstrate improvement. Also preserve data and code identifiers. In MLflow 3, the default registry is databricks-uc; check the registry URI and use three-part names when working with Unity Catalog. A model signature describes inputs and outputs and is required when registering new model versions in UC.

Guided application

An alias such as champion points to a version and can change. A batch consumer that resolves the alias while loading may use a different version on its next execution. An endpoint already configured with a version does not automatically change just because the alias changed. Treat promotion as a decision with evidence, identity, actual version, acceptance, and recovery. Tags help describe validation but do not replace access control or a traffic change. Decide whether to promote code for training in production or an already trained artifact; document data, cost, and reproduction conditions.

IN PRACTICE

Champion points to v8, but the endpoint remains configured for v7 until updated.

Common pitfalls

Tag treated as executable approval; alias treated as automatic serving update.

Related topics: Environment, AutoML, and reproducibility · Temporal features and consistency · Data, preprocessing, and validation

Take this idea with you

Record evidence and confirm the version actually serving consumers.

Create account

Reference: Model lifecycle in Unity Catalog · 2025-03-01