Study Databricks ML-ASSOC MLflow Runs and Logging: key concepts, common traps, and exam decision cues.
MLflow is easy to reduce to “log some stuff.” The exam is testing whether you know what should be logged, how runs differ from models and registry entries, and what information the UI actually helps you inspect.
| Item | Why it matters |
|---|---|
| params | explain what configuration produced the result |
| metrics | compare model quality across runs |
| artifacts | preserve outputs such as plots, files, and model objects |
| run | keeps one experiment attempt coherent |
| UI | helps inspect runs, artifacts, and comparisons visually |
| Layer | What belongs there |
|---|---|
| run | one attempt with its params, metrics, and artifacts |
| experiment | collection of related runs |
| registered model | managed named model object for lifecycle control |
| alias or version | promotion pointer inside the registry layer |
If you mix these layers, the answer choices start to look similar. The exam often hides the miss there.
| If the stem says… | Strong reading |
|---|---|
| “identify the best run” | compare logged runs consistently |
| “manually log metrics, artifacts, and models” | know what belongs in each logging bucket |
| “information available in the MLflow UI” | understand the UI as an inspection surface for tracked work |
Good MLflow logging is not administrative overhead. It gives you the evidence needed to answer three exam-style questions:
If an answer only preserves the final score and throws away the configuration or outputs, it is usually weaker than Databricks wants.
| Trap | Better rule |
|---|---|
| treating params and metrics as interchangeable | params describe the attempt; metrics judge the result |
| assuming a run is the same thing as a registered model | a run is experiment history, not registry lifecycle |
| logging only the score | reproducibility also needs config and outputs |
| Scenario clue | Stronger answer shape |
|---|---|
| “why did this run score higher?” | inspect params and metrics together |
| “where is the confusion-matrix image or saved model file?” | artifacts |
| “which candidate should move forward?” | compare runs, then connect the chosen run to registry actions later |
| “what can the UI help the team inspect quickly?” | tracked run details and comparisons, not every deployment-control operation |
MLflow questions usually hinge on whether you are talking about one experiment attempt or a managed lifecycle object. Runs capture one tracked training attempt with metrics, params, and artifacts. Registered models are lifecycle objects built from those experiments. The exam usually rewards logging enough context for reproducibility instead of only the final score.