ML Professional: study scenarios by profile

For the Databricks Machine Learning Professional exam, study for the 44/44/12 split rather than for “production” in general: Model Development and ML Ops are 44% each, Model Deployment is 12%, and the 59 scored questions run for 120 minutes in English only. Databricks recommends 1+ years of hands-on experience, so the real question is which half of the paper your experience actually covers.

Example profiles, not real people. The profiles below are illustrative planning examples built from the official exam outline. They are not reviews or reports of anyone’s result.

The budgets start from the site’s six-week study plan, which assumes 8 to 10 hours a week, roughly 48 to 60 hours, with 16 to 20 hours on each large section. Hour counts here are planning assumptions, not promises.

Profile one: the Associate-level builder with one model but no live history

Starting point. This candidate can track experiments, register a model and use Unity Catalog, at the level the ML Associate tests. A model has been trained and perhaps deployed, but it has not run long enough to drift or be retrained.

Where the time goes. ML Ops (44%), the section with no Associate equivalent. The prerequisites page calls “no production ML experience” the expensive gap because it blocks 44% of the paper and questions describe symptoms to diagnose.

The trap. Assuming the Associate’s platform knowledge carries over. The Professional assumes that ground and tests none of it directly.

Suggested order. One practical cycle first: register versions, monitor against a baseline, retrain and compare. Then ML Ops study, then Model Development, then deployment.

Time budget. A reasonable planning assumption is 20 to 24 hours on ML Ops, 14 to 16 on Model Development, 6 on deployment, and 8 on practice: 48 to 54 hours, plus the time the practical cycle takes, which can stretch the plan towards eight weeks.

Profile two: the modelling specialist who has never operated a model

Starting point. This candidate validates carefully, knows leakage and tunes well, mostly on single-node tools. The outline’s Model Development section is written around SparkML pipelines, distributed tuning with Optuna and Ray, nested MLflow runs and point-in-time Feature Store lookups, which is broader than general modelling.

Where the time goes. The platform-scale half of Model Development, then the whole of ML Ops: testing, bundles, automated retraining and drift detection with Lakehouse Monitoring.

The trap. Assuming that strong modelling means 44% is secured. The distributed and Feature Store material is specific, and the other 44% is operational.

Suggested order. SparkML and distributed tuning, Feature Store and MLflow, then drift and retraining, then testing and environments.

Time budget. A reasonable planning assumption is 16 to 18 hours on Model Development, 20 on ML Ops, and 6 on deployment, plus the plan’s practice week.

Profile three: the platform engineer who automates releases for ML teams

Starting point. This candidate writes CI/CD, manages environments and deploys with bundles. Environment architecture, unit and integration testing and rollout management feel natural, which covers a large slice of ML Ops and all of deployment (12%).

Where the time goes. Model Development (44%): validation strategy, leakage, evaluation beyond one metric, and tuning. Drift statistics inside ML Ops also need study.

The trap. Putting effort into deployment because it is comfortable. It is about seven questions, and the study plan’s own advice is to cover it properly and stop.

Suggested order. Model Development first, drift and retraining second, then a short deployment pass.

Time budget. A reasonable planning assumption is 22 to 24 hours on Model Development, 12 to 14 on ML Ops gaps, and about 4 on deployment, plus the plan’s practice week.

The three ML Professional profiles compared

How to check which ML Professional profile fits before booking

Run the five checks on the prerequisites page: can you promote a version, identify which version served a prediction, monitor against a baseline, and compare a retrained model with the incumbent? Then take the sample questions and sort misses into the two 44% sections. The practice test should be taken under the full 120 minutes, and the exam format page gives the English-only delivery rules.

If you have not run a model for months, the ML Associate is the lower-cost first step, and the comparison page lays out the difference.