Databricks ML Professional prerequisites

There are no formal prerequisites for the Databricks Certified Machine Learning Professional exam. No certification required, nothing checked at registration, and the ML Associate is not needed.

What Databricks does state is 1+ years of hands-on experience performing the machine learning tasks in the exam guide — double the Associate’s six months, and the most demanding experience expectation of any Databricks certification covered here.

Why the year matters

Because 44% of the exam is ML Ops, and ML Ops is the study of things that go wrong slowly.

Drift, degradation, retraining decisions, version confusion, training/serving skew — none of these appear during a project. They appear months after a model goes live. You can read about them, but the questions are written from the perspective of someone who has watched a model quietly get worse and had to work out why.

That is the gap the year represents, and it is not easily closed by study.

What is assumed

Platform fluency. MLflow tracking and registry, AutoML, Unity Catalog. The Associate devotes 38% of its exam to these; the Professional assumes them entirely and tests none of them directly.

Python. Reading ML code fluently — libraries, pipelines, training scripts, evaluation.

Machine learning fundamentals at depth. Not just which algorithm, but validation strategy, evaluation beyond a single metric, and what leakage looks like.

Production experience. Having deployed something and lived with it.

What you do not need

• No certification, including the Associate.
• No deep mathematics. You are not examined on the derivations behind algorithms.
• No infrastructure engineering. Deployment is 12% and at a conceptual level.
• No deep learning specialism. The exam covers general ML practice.

Ranking your gaps

1. No production ML experience. The expensive one — it blocks 44% of the exam. Close it with time, not reading.
2. Never used MLflow properly. Assumed throughout, tested indirectly everywhere.
3. Weak validation practice. If you have never worried about leakage, the model development section will be harder than it looks.
4. No exposure to retraining decisions. Concentrated in ML Ops.
5. No deployment experience. Only 12%, the most forgiving gap.

If you are short of the year

Take the ML Associate first. It covers the platform knowledge this exam assumes, at a level matching six months of experience. Its 38% platform section is genuinely preparatory.

Or operate something for a few months. Deploy a model, monitor it, watch it drift, retrain it, and compare the new version against the old. One cycle of that teaches more of the ML Ops section than any course.

What does not work is reading about drift. The questions are diagnostic — a scenario describes symptoms and asks what is happening — and diagnosis comes from having seen the disease.

Cost of preparing

You need a Databricks workspace, and more compute than the Associate requires: training jobs, a registry with several versions, and ideally a model running long enough to monitor. Terminate clusters when you finish.

Before you book

1. Register three versions of a model and promote one.
2. Identify which version served a given prediction.
3. Set up monitoring against a training baseline.
4. Retrain and compare the new model against the incumbent.
5. Try the free sample questions.

If steps 2 to 4 are unfamiliar, that is the one-year gap showing. Build it before booking rather than reading about it after.