Description

For learners studying for Databricks Certified Machine Learning Professional
What you'll learn
Practice pack content
Domain 1 - Model Development
Domain 2 – ML Ops
Domain 3 – Model Deployment
Final Exam Simulators
Exam Study Guide
Interactive Flashcards
Try 3 Machine Learning Professional practice questions
Choose an answer, explore the explanation, and see why the other options are less suitable. Three short scenarios. No sign-in required.
Model DevelopmentEnsure point-in-time correctness in feature lookups to prevent data leakage during model training and inference.
1. Preventing future feature values from leaking into training
A fraud model is trained from historical transaction labels and a feature table whose customer-risk features change throughout the day. The training set must use only feature values that were available when each transaction occurred. Which approach best prevents feature leakage?
- Join every historical transaction to the latest feature values currently stored for that customer.
- Use a time-series feature table with timestamp keys and perform point-in-time feature lookups for each transaction timestamp.
- Read the feature table with Delta time travel at the date when the training notebook was first created.
- Sort the feature table by event time before joining, but use a standard equality join on customer ID only.
Answer and full explanation
Correct answer: B — Use point-in-time feature lookups with timestamp keys.
Why B is correct
Point-in-time feature joins select feature values that were known as of each observation's event timestamp. That prevents the training set from using information that became available only after the label event occurred. Databricks time-series feature tables use timestamp keys so feature lookups can resolve the latest valid value before or at the observation time. The Professional exam guide explicitly requires point-in-time correctness in feature lookups to prevent data leakage during training and inference. This approach therefore preserves historical realism while still allowing features to evolve over time.
Why the other options are less suitable
- A — Latest feature values: Joining historical labels to today's newest features can expose the model to values that did not exist when the outcome occurred. That is a classic source of target or temporal leakage.
- C — Delta time travel at notebook creation: Delta time travel reconstructs a table snapshot at a chosen version or timestamp. A single snapshot does not automatically select the correct feature value separately for every historical transaction time.
- D — Sort before an equality join: Sorting changes row order but does not enforce temporal correctness. A join on customer ID alone can still pair an observation with a feature value recorded after that observation.
Key takeaway: Historical training data should use feature values that were actually available at each label's event time.
Read two more sample questions
MLOpsDetect data drift by comparing current data distributions to a known baseline or between successive time windows.
2. Detecting a meaningful distribution shift
A production classification system is profiled daily. The ML team needs a queryable output that shows how feature distributions have changed relative to a baseline or the preceding time window so they can alert when drift becomes significant. Which monitoring output best fits this requirement?
- The profile metrics table containing only the current window's column summary statistics.
- The MLflow model artifact directory containing the serialized production model and dependencies.
- The drift metrics table containing comparisons between baseline or consecutive-window metrics.
- The billing system table containing DBU consumption and workspace usage for the monitoring job.
Answer and full explanation
Correct answer: C — Use the drift metrics table.
Why C is correct
Databricks monitoring produces drift metrics that describe changes between metric values rather than only reporting a single window's profile. Drift can be calculated against a baseline or between consecutive time windows, which matches the comparison required in the scenario. These metrics are stored in a dedicated drift metrics table that can be queried and used to drive dashboards or alerts. The MLOps section of the Professional guide explicitly includes detecting drift by comparing current distributions with a known baseline or successive windows. The drift table is therefore the most direct source for the team's change-detection workflow.
Why the other options are less suitable
- A — Profile metrics only: Profile metrics describe characteristics such as counts, averages, or other statistics for a particular profile or window. They do not by themselves provide the explicit baseline-versus-current or window-versus-window comparison requested.
- B — MLflow artifacts: Model artifacts preserve the trained model and related files. They do not record ongoing changes in production feature distributions.
- D — Billing system table: Billing data is useful for cost and resource analysis. It does not measure statistical distribution changes in model inputs.
Key takeaway: Use drift metrics when the question is how a data distribution has changed relative to a baseline or earlier window.
Model DeploymentRegister a custom PyFunc model and log custom artifacts in Unity Catalog.
3. Serving a model with custom preprocessing logic
A production model requires a custom tokenizer artifact plus Python preprocessing and postprocessing that are not represented by the model framework's standard MLflow flavor. The team wants to package this behavior as one governed model and deploy it through Databricks Model Serving. Which approach is best?
- Create a custom MLflow PyFunc model that loads the tokenizer artifact, register it in Unity Catalog, and serve the registered model.
- Store the tokenizer in a notebook variable and require every application to call that notebook before querying the model endpoint.
- Register only the raw framework model, then duplicate the preprocessing and postprocessing logic independently in every client application.
- Save the tokenizer to a local driver path and configure the serving endpoint to read that driver's filesystem during each request.
Answer and full explanation
Correct answer: A — Package the custom logic and artifact in an MLflow PyFunc model.
Why A is correct
MLflow PyFunc is intended for models that need arbitrary Python code or custom inference behavior beyond a built-in model flavor. A custom Python model can load artifacts such as tokenizer files and execute preprocessing and postprocessing inside its prediction path. After the model is logged, Databricks recommends registering it in Unity Catalog and deploying that registered model to a Model Serving endpoint. The Professional exam guide explicitly includes registering custom PyFunc models and custom artifacts in Unity Catalog. This keeps model logic, dependencies, and governed deployment together instead of scattering inference behavior across clients.
Why the other options are less suitable
- B — Notebook preprocessing: Notebooks are useful for development and jobs, but forcing each online request through separate notebook logic complicates serving and does not package the complete inference contract as one model artifact.
- C — Duplicate logic in every client: Reimplementing preprocessing and postprocessing across applications creates consistency and versioning risks. It also separates critical inference behavior from the governed model package.
- D — Driver-local artifact: A serving endpoint should not depend on a transient notebook or cluster driver's local filesystem. Required artifacts should be packaged or referenced through the model so serving instances can load them consistently.
Key takeaway: Custom PyFunc models package nonstandard Python inference logic and artifacts into a deployable MLflow model.
Preview complete
You have tried all three samples
Review the explanations or try the samples again. This short preview covers only a small part of the exam objectives.
Your sample score resets when this page reloads.
Your practice pack includes
More ways to make each study session count
Focus on a topic, practise under a time limit, or return to the supporting resources before your next attempt.
Domain-based practice
Timed practice sets
Study guide
Flashcards
Requirements
- Databricks requires no prerequisite certification, but highly recommends related training and about one year of hands-on Databricks experience, plus working knowledge of Python, scikit-learn, SparkML, MLflow, monitoring, and Model Serving.
- Learners with gaps should combine question practice with foundational study and their own exercises involving SparkML, distributed tuning, Feature Store, MLflow, automated retraining, drift monitoring, model serving, testing, and CI/CD.
- You will need an internet-connected device, a current web browser, and an ExamsDigest account to use the online practice product; this does not imply offline availability or a separate mobile application.
Know the exam you’re studying for
These details describe Databricks’s official Machine Learning Professional exam.
What our students say
“The practice exams helped me identify the areas I still needed to improve, especially around feature engineering, MLflow, model monitoring, and deployment. The explanations were detailed without being confusing and helped me understand why each option was right or wrong.”
Martin Keller
✓ Verified Review
“The questions covered a strong mix of model development, MLOps, feature management, drift detection, and production deployment scenarios. Reviewing my incorrect answers made it much easier to see which Databricks topics I needed to spend more time on.”
Alicia Norton
✓ Verified Review
“I had already gone through the Databricks training materials before using ExamsDigest. What I found most useful was how the explanations compared different approaches, especially for point-in-time features, MLflow PyFunc models, monitoring, and serving models in production.”
Garrett Cole
✓ Verified Review
Your questions, answered.
Which Machine Learning Professional exam does this practice target?
These questions target the currently live Databricks Certified Machine Learning Professional exam described by the guide effective September 30, 2025. Databricks does not publish a separate exam code for this certification. Before your scheduled exam, check the certification page and linked guide again because Databricks identifies the exam guide as the source of truth for the content on the currently live version.
Who is this Professional-level practice intended for?
It is intended for experienced machine-learning engineers who design and operate enterprise-scale Databricks ML systems. Databricks requires no formal prerequisite certification but recommends at least one year of hands-on platform experience. Useful background includes Python, SparkML, scikit-learn, MLflow, distributed training, Feature Store, model lifecycle management, testing, automated retraining, drift monitoring, and Databricks Model Serving.
How should I combine practice questions with hands-on Databricks work?
Use practice alongside the current guide and your own production-style exercises. Build SparkML pipelines, experiment with Optuna or Ray, create time-aware feature lookups, use nested MLflow runs, configure automated retraining, profile drift, create alerts, register custom PyFunc models, and deploy serving endpoints. The exam is aimed at enterprise-scale decision-making, so being able to explain trade-offs and operational consequences matters as much as recognizing feature names.
What should I do when I answer a question incorrectly?
Use the explanation to identify the exact engineering decision you missed, then return to the matching guide objective and reproduce the workflow when practical. Determine whether the mistake involved distributed training, feature correctness, MLflow, testing, monitoring, automated retraining, or deployment. Compare every distractor against the scenario’s scalability, governance, latency, or reliability requirements rather than memorizing an answer letter. Practice results can reveal gaps but are not an official pass prediction.
What question format does the official exam use?
The official Machine Learning Professional exam uses multiple-choice questions, with 59 scored items listed in the current guide. Databricks may also include unidentified unscored questions for future statistical use, and those items do not affect your result. The three samples here are single-answer multiple-choice examples and therefore demonstrate the basic interaction type without claiming to reproduce Databricks’ live questions, topic frequency, or exact exam difficulty.
Does completing this practice earn the Databricks certification?
No. Completing an ExamsDigest practice product does not award the Databricks Certified Machine Learning Professional credential or replace the official proctored examination. Earning the certification requires registering for and passing Databricks’ current live exam through its authorized testing process. Databricks states that certifications remain valid for two years and that recertification requires taking the full current exam again.
30-day conditional refund policy
Eligible requests must be made within 30 calendar days, with no more than 20% of any included simulator completed. The purchase must not already be refunded or under a payment dispute or chargeback. Other provisions apply; mandatory consumer rights remain unaffected.
Useful links for your preparation



