Databricks ML Associate: the platform section
Databricks Machine Learning is worth 38% of the Machine Learning Associate exam — around 18 of the 48 scored questions and the largest section by a clear margin. It is platform knowledge, not modelling theory, and it is where candidates with strong ML backgrounds most often lose marks.
If you learn one section properly, learn this one.
The ML runtime and workspace
- The Databricks ML runtime, which ships the common ML libraries preinstalled, and why you would choose it over a standard runtime.
- Cluster configuration for ML work, including when a GPU is justified.
- Notebooks and repos as the working surface.
MLflow: tracking
The component that appears most often on this exam.
Tracking records what happened during experimentation:
| Concept | What it is |
|---|---|
| Run | One execution of your training code |
| Parameters | Inputs you chose — hyperparameters, settings |
| Metrics | Outputs you measured — accuracy, loss, RMSE |
| Artefacts | Files produced — the model, plots, data samples |
| Experiment | A collection of related runs |
The exam’s framing: whenever a scenario involves comparing attempts, reproducing a result, or recovering which settings produced the best model, the answer is MLflow tracking.
MLflow: model registry
Where a model becomes a managed asset rather than a file:
- Registering a model from a run.
- Versions — each registration creates a new one.
- Stages and promotion, moving a version toward production.
- Knowing exactly which version is serving, and rolling back.
The distinction to hold: tracking is about experiments, the registry is about lifecycle.
AutoML
- Runs a set of experiments automatically to produce a baseline.
- Crucially, it generates editable notebooks showing exactly how each model was built. It is not a black box, and questions test that you know this.
- Best used as a starting point and a benchmark — if your hand-built model cannot beat AutoML, that is useful information.
Unity Catalog
Governance across data and models:
- Access control over tables, volumes, models and features.
- Lineage — what a model was trained on, and what uses it.
- Consistent permissions across workspaces.
The clean split the exam relies on: MLflow manages the model’s lifecycle. Unity Catalog governs who can access what and where it came from. Questions routinely offer both.
Feature management
- Storing engineered features so they can be reused.
- Ensuring training and inference compute features identically — the fix for training/serving skew.
- Discoverability, so teams do not rebuild the same features repeatedly.
Sample questions
Question 1. A model in production is producing unexpected results. The team must determine which training run produced the deployed version and what hyperparameters it used. Where do they look?
- A. Unity Catalog lineage for the underlying tables
- B. The MLflow model registry, linked to the originating tracked run
- C. The AutoML experiment summary
- D. The cluster event logs
Show answer
Answer: B
The MLflow registry identifies which version is deployed and links back to the originating run, whose parameters are recorded in MLflow tracking. Unity Catalog governs access and lineage over data rather than run parameters, AutoML generates baselines, and cluster logs record compute events.
Question 2. A team runs AutoML on a new dataset. What should they expect as output?
- A. A single accuracy score with no model artefact
- B. Deep learning models only
- C. A baseline model plus editable notebooks showing how the trials were built
- D. A model automatically deployed to a production endpoint
Show answer
Answer: C
AutoML produces a baseline model together with generated notebooks showing how each trial was built, which the team can inspect and modify. It does not return only a score, it is not restricted to one algorithm family, and it does not deploy to production automatically.
Question 3. Features are computed one way in the training notebook and reimplemented in the inference job. Which platform capability addresses the resulting risk?
- A. Centralised feature management so both paths use the same definitions
- B. Logging both jobs to MLflow tracking
- C. Running the inference job on a larger cluster
- D. Applying a cluster policy to both workloads
Show answer
Answer: A
Centralised feature management lets training and inference use the same feature definitions, preventing training/serving skew from divergent reimplementation. MLflow tracking records runs, a larger cluster changes compute, and a cluster policy constrains configuration.
What to practise
Run one AutoML experiment and open the notebooks it generates. Log a manual training run to MLflow, register it, create a second version, and promote one. Then set permissions on that model in Unity Catalog.
That sequence takes an afternoon and covers the largest section of the exam. There is no reading substitute — the questions assume you have seen these screens.