DE MLO: Experiment Tracking and Model Lifecycle

What the exam tests

This domain covers MLflow Tracking and the path from an experiment run to a registered model version and deployment alias. Questions test whether you know what belongs in a run, how runs relate to experiments and registered models, and how Unity Catalog changes model naming, permissions, and lifecycle promotion.

Core rules

• An MLflow experiment groups runs. Each run records parameters, metrics, tags, artifacts, source information, and start/end times. Use nested runs when a parent tuning job launches child trials.
• Log models with the framework flavor or mlflow.log_model, and include a signature plus an input example. The signature defines input/output schema and prevents many serving failures.
• In Unity Catalog, register models with a three-level name such as catalog.schema.churn_model. Each registration creates an immutable model version.
• Use aliases such as champion and challenger to point at versions. Aliases are the current promotion mechanism; legacy stages are deprecated and should not be the basis for new lifecycle designs.
• Lifecycle states are conceptual: develop, validate, deploy, monitor, retire. Promotion changes an alias, not the version number.

Worked example

A candidate trains two scikit-learn classifiers in one experiment. For each run they log C, max_depth, validation F1, the Git commit, and the Delta table version used for training. They log the better model with a signature and input example, register it as main.default.churn_model, and set alias champion. A serving endpoint references champion. If the model regresses, rollback means moving champion to the previous version. The run ID links metrics and artifacts to the version, the version is immutable, and the alias gives deployment a stable pointer.

Traps that cost marks

• Registering every run creates clutter; register only validated candidates.
• Omitting the signature or input example causes schema and serving errors.
• Assuming the registry stores training data. It stores model artifacts and metadata; data lineage comes from Unity Catalog and logged source information.
• Deleting experiments or runs to clean up breaks lineage. Archive instead.
• Treating a model version number as a deployment pointer. Consumers should resolve an alias or explicit version intentionally.

Key points

• Experiments contain runs; runs contain params, metrics, tags, artifacts, and source metadata.
• Log signatures and input examples with models intended for serving.
• Unity Catalog model names are three-level and versions are immutable.
• Use aliases for promotion and rollback; avoid relying on deprecated stages.
• Preserve run lineage instead of deleting runs.