Databricks Certified Machine Learning Professional

About the Databricks Machine Learning Professional exam

The Databricks Certified Machine Learning Professional exam is the senior step after the ML Associate. It is less about training a single model and more about running machine learning as a production system: training at scale with Spark, tuning in parallel, moving models safely from development to production, watching them for drift and retraining them automatically when they degrade. Databricks recommends at least a year of hands-on experience with the tasks in the exam guide.

Most questions describe a team with a working model and a production problem. Training is too slow, a feature is computed differently online and offline, a new model version needs a canary rollout, or monitoring shows the input data has shifted. You choose the Databricks or MLflow approach that solves it cleanly. The exam also tests SQL, written to ANSI SQL standards.

Exam format

• Questions: 59 scored multiple-choice questions. The exam may also include unscored items, with extra time allowed for them.
• Time: 120 minutes.
• Cost: 200 USD.
• Delivery: Proctored, online or at a test center. No test aids are allowed.
• Language: English.
• Validity: 2 years. To recertify you take the current version of the exam again.

Databricks does not publish a passing score on the exam page.

What the exam covers

The exam outline has three sections, two of them large. The official outline gives only the weights, so the topics below are grouped by where they most naturally fit:

• Model Development (44%). Spark ML pipelines and how to keep preprocessing inside them to avoid leakage, when Spark ML is worth it over a single-node library, class weights, caching training data, distributed hyperparameter tuning with Optuna, training one model per group, and multi-node deep learning with TorchDistributor. It also covers advanced MLflow, such as custom pyfunc models, models from code and dataset lineage, and Feature Store topics like point-in-time lookups and on-demand features.
• ML Ops (44%). Environment management with Declarative Automation Bundles and per-target variables, the "deploy code, not models" pattern, unit and integration tests, data quality expectations, model lifecycle and promotion across Unity Catalog catalogs, Lakehouse Monitoring for drift with custom metrics, and automated retraining triggered by table updates and gated by evaluation thresholds and approval.
• Model Deployment (12%). Model Serving strategies: canary traffic splits, scale-to-zero and its cold starts, publishing features to an online store, granting access to serve a model, and inference tables for logging requests.

How the practice questions here are organised

There are three 10-question practice tests. Each one samples from across the outline, with most questions on model development and MLOps, matching the exam's weighting. Topics are labelled at a finer level, for example Scaling and Tuning, Advanced MLflow, Environment Architectures, Drift Detection and Lakehouse Monitoring, so you can see where your weak spots are. Each question explains why the right answer works and why each other option falls short, with a link to the Databricks or MLflow documentation it is based on.

A study plan that works

• Be solid on the ML Associate material first. This exam assumes you already know MLflow tracking, the Unity Catalog model registry and basic Model Serving.
• Take Test 1 cold, then read every explanation, including the ones you got right by guessing.
• Set up a small project with a bundle that deploys a training job to a development and a production target. Seeing the targets and variables in practice makes the environment questions much easier.
• Create a monitor on an inference table and read the drift metrics it produces, so the monitoring questions describe something you have seen.
• Expect to spend time on distributed training and tuning. Know when work should spread across the cluster and when a single node is the better choice.
• Work through Tests 2 and 3, and retake any test you scored under 70% after a few days.

These questions are original, written from the public Databricks and MLflow documentation and the official exam guide. They are not real exam questions, and this site is not affiliated with Databricks.