Databricks ML Professional study plan

Six weeks at 8–10 hours a week, assuming roughly the one year of hands-on experience Databricks recommends. You need a workspace and, ideally, a model you have run long enough to have seen behave badly.

The plan follows the 44/44/12 weighting: half the time on model development, half on ML Ops, and a short block on deployment.

Weeks 1 and 2: model development (44%), part one and two

Week 1 — rigorous development:

• Feature engineering at depth, and managing features so training and inference agree.
• Training approaches and when each algorithm family suits a problem.
• Hyperparameter tuning properly: search strategies, tuning on a validation set, and knowing when further tuning stops paying.
• Distributed training considerations.

Week 2 — validation and reproducibility:

• Validation strategy: holdout, cross-validation, and time-based splits for temporal data.
• Data leakage — how it arises and what it looks like. Excellent validation scores and poor production performance is the signature.
• Evaluation beyond a single metric: precision and recall on imbalanced classes, calibration, and error analysis by segment.
• Reproducibility: recording data version, code version, parameters and environment so a model can be rebuilt.
• MLflow tracking used properly rather than incidentally.

Weeks 3 and 4: ML Ops (44%)

The section with no Associate equivalent. Two weeks.

Week 3 — monitoring and drift:

• What to monitor: input distributions, prediction distributions, and outcome quality where ground truth arrives.
• Data drift versus concept drift. The first is inputs moving; the second is the relationship between inputs and the right answer changing.
• Establishing a baseline — you cannot detect drift without knowing what normal looked like.
• Why infrastructure monitoring tells you nothing about model quality: a degraded model is up, fast and wrong.
• Setting thresholds that mean something, and alerting on them.

Week 4 — lifecycle and automation:

• Model versioning and the registry: what is deployed, what it replaced, and how to roll back.
• Retraining: scheduled versus triggered by a monitored signal, and how to decide.
• Validating a retrained model against the incumbent before promoting it — a new model is not automatically better.
• Automating the pipeline from retraining through evaluation to promotion.
• Governance of models in Unity Catalog: ownership, access, lineage.

Week 5: deployment (12%), then start practising

The smallest section — a few focused hours, not a week.

• Batch, real-time and streaming inference, and matching the pattern to the requirement.
• Serving a registered model from the registry.
• Scaling, latency and cost of serving.
• Deployment patterns that limit risk: staged rollout, shadow traffic, rapid rollback.

Then begin practice questions across all three sections.

Week 6: practice and repair

• Full practice exam under real conditions: 120 minutes, 59 questions.
• With only three sections, the score report is blunt — a weak 44% section is immediately visible and immediately urgent.
• Rebuild the weaker of the two large sections, then take a second practice exam.

Where the hours go

The question to keep asking

For every ML Ops topic: how would I know?

How would I know this model has degraded? How would I know the input data has shifted? How would I know the retrained version is actually better? How would I know which version produced last Tuesday’s prediction?

If you can answer all four for a model you have worked on, you are ready for 44% of this exam. If you cannot, that gap is exactly what weeks 3 and 4 are for.

Do not over-prepare deployment. Seven questions. The temptation is real because it feels like the “production” part, but ML Ops is where production actually lives on this exam.