EnglishDeutschFrançaisEspañolPortuguês

Databricks · DB-MLP · Professional

Databricks Machine Learning Professional — Practice Questions and Mock Exam

The Databricks Certified Machine Learning Professional exam validates advanced expertise in building production ML systems at scale on Databricks.

Practice for freeTry a practice question first
Try a practice question firstPractice for free
Start practicing DB-MLP freeBuy DB-MLP access for $29.99
59Questions per Pass-IT mock exam
120minPass-IT mock exam time limit

Checked against Databricks · October 2026 · Current exam version

Try Pass-IT

See what learning with Pass-IT is like.

Answer a DB-MLP sample question and read the explanation with its source. Right here, with no sign-up.

MLOps1 / 5

A machine learning engineer needs to implement an MLOps practice where every model training run is linked to the exact Git commit of the training code, the specific version of the training data, and the resulting model artifact. What ensures this end-to-end traceability?

AlexFull explanation from Alex

MLOps provenance chain: 1) Code → Run: Databricks Repos: automatic Git integration. MLflow: logs notebook path and Git commit SHA. Reproducibility: checkout commit, rerun notebook. 2) Data → Run: Delta Lake version: exact snapshot of training data. Data hash: fingerprint of the dataset. Feature Store version: which feature table version used. Time travel: spark.read.format('delta').option('versionAsOf', 42).table('training_data'). 3) Run → Model: MLflow run contains: parameters, metrics, artifacts. Model registered from run: linked via run_id. Model Registry: version → run → code + data. 4) Model → Deployment: deployment logs: which model version is serving. Endpoint: references model URI. Inference table: logs every prediction. 5) Full chain: prediction → model version → MLflow run → (Git commit + data version) → original source code and data. 6) Automation: autologging: captures most code/model links automatically. Data logging: requires manual mlflow.log_param for data version. Unity Catalog: adds governance layer (permissions, audit logs). 7) Benefits: reproducibility: recreate any model. Auditability: satisfy compliance requirements. Debugging: find what changed when model degraded.

Sourcedocs.databricks.com

Model Deployment2 / 5

A machine learning engineer needs to implement a solution that serves different model versions to different geographic regions. The EU region requires a model trained with GDPR-compliant data (no personal identifiers), while the US region can use all features. How is this multi-region serving configured?

AlexFull explanation from Alex

Regulatory-aware model serving: 1) GDPR requirements: data minimization: the EU-facing model may be trained only on features with no personal identifiers. Purpose limitation: features used for the stated purpose only. Right to be forgotten: the model must not memorize individual data. 2) Feature categories: PII: name, email, phone, address, IP. Quasi-identifiers: age + zip code + gender (combination may identify). Aggregated: average transaction amount (last 30 days). Behavioral: number of logins (no individual identification). 3) Why two registered versions: the compliance boundary is the training data itself, so you need two distinct trained artifacts — one trained on the GDPR-compliant feature set, one on full features — registered side by side in the Model Registry, where each version carries its own lineage, stage, and audit trail. 4) Request routing: a routing layer in front of the serving endpoint inspects the request's region header and directs EU requests to the compliant version and US requests to the full-feature version. Databricks Model Serving supports serving multiple models and multiple versions of a model at the same time and lets you query an individual served model behind an endpoint, so per-request routing to a specific registered version is a supported pattern. 5) Testing: compliant version: verify no PII features in its input signature. Full-feature version: verify all features available. Routing: verify an EU region header always resolves to the compliant version. 6) Compliance documentation: model cards per registered version document which features are used. Model Registry lineage tracks the training-data origin of each version. Access logs record which version served each request.

Sourcedocs.databricks.com

Model Development3 / 5

When performing hyperparameter tuning with Hyperopt on Databricks, what is the primary benefit of using SparkTrials instead of the default Trials class?

AlexFull explanation from Alex

Hyperopt is a hyperparameter optimization library that supports Tree-structured Parzen Estimators (TPE) and random search algorithms. The default Trials class runs trials sequentially on a single machine. SparkTrials extends this by distributing trials across a Spark cluster, where each worker trains a model with different hyperparameters in parallel. This is particularly effective for single-node ML models (scikit-learn, XGBoost) where each trial fits on one worker.

Sourcedocs.databricks.com

MLOps4 / 5

A model monitoring system detects that the input feature distributions have shifted significantly from the training data distribution. However, model performance metrics remain stable. What should the ML engineer do?

AlexFull explanation from Alex

Model monitoring distinguishes between data drift (input feature distribution changes) and concept drift (relationship between features and target changes). Data drift can occur without immediate performance impact if the model generalizes well to the shifted region. However, sustained data drift often precedes concept drift. Best practice is to monitor both drift metrics and performance metrics, set up alerts at different severity levels, and have retraining pipelines ready to trigger when performance actually degrades.

Sourcedocs.databricks.com

Model Deployment5 / 5

A machine learning engineer needs to implement a model serving solution that returns not just a prediction but also a confidence score and the top contributing features for each prediction. The response must include all three pieces of information. How should the endpoint be designed?

AlexFull explanation from Alex

Rich model serving responses: 1) Basic: {'prediction': 1}. 2) With confidence: {'prediction': 1, 'confidence': 0.92}. 3) With explanation: {'prediction': 1, 'confidence': 0.92, 'top_features': [{'name': 'credit_score', 'impact': 0.15}, {'name': 'debt_ratio', 'impact': -0.08}]}. 4) With metadata: add model_version, timestamp, request_id. 5) Implementation patterns: pyfunc predict → returns DataFrame. Each column becomes part of the response. Complex structures: serialize to JSON string in a column. 6) Performance trade-offs: prediction only: ~10ms. Prediction + confidence: ~10ms (probabilities already computed). Prediction + SHAP: ~50-200ms (SHAP computation adds overhead). Optimize: compute SHAP only when requested (optional flag in request). 7) Client integration: REST API: JSON response with all fields. SDK: parsed into typed objects. Dashboard: display prediction + confidence gauge + feature waterfall chart.

Sourcedocs.databricks.com

How to prepare for DB-MLP with Pass-IT.

Currently 325 practice questions for DB-MLP, with new ones added regularly. Focus on what you still need to learn.

Get past difficult questions.

Understand why an answer is correct. Alex explains the concepts and answers your follow-up questions so you can apply what you learn to the next task.

Meet Alex →

Conversation preview · DB-MLP

Parallel tuning trials

  1. You

    What does SparkTrials add to Hyperopt Trials?

  2. Alex · AI tutor

    It distributes individual trial runs across Spark workers for parallel execution.

    Read the Databricks documentation ↗
  3. You

    Does that mean every model becomes a distributed training job?

  4. Alex · AI tutor

    No. The parallelism described here is between trial runs. Keep trial distribution separate from a model’s own training method.

    Read the Databricks documentation ↗

Make your study time count.

Practice at your level.
For new practice sessions, Pass-IT adjusts the difficulty based on your previous answers.

Learn from mistakes.
Incorrect answers automatically become flashcards. When you review them, you rate how confidently you remember; Pass-IT schedules the next review.

See how focused practice works →

Animated feature preview

Turn your mistakes into knowledge you can rely on.

How Pass-IT works

Your review

Answered a question incorrectly?

Pass-IT saves it for your review.

DB-MLP · Question excerpt

Parallel tuning trials

View the full question

When performing hyperparameter tuning with Hyperopt on Databricks, what is the primary benefit of using SparkTrials instead of the default Trials class?

Understand the correct answer.

SparkTrials distributes individual trial runs across Spark worker nodes for parallel execution

SparkTrials parallelizes trial runs across Spark workers.

Read the Databricks documentation ↗

How confidently do you remember?

During review, you assess how well you still know the answer.

Correct after hesitating

Pass-IT schedules your next review.

The date depends on your self-assessment and previous reviews.

Simplified demonstration. No learning progress is saved.

Practice for the moment that counts.

Take a full mock exam with a time limit and find out whether you can finish in time.

Your DB-MLP mock exam

Questions
59
Minutes
120

Illustration of an exam session

  1. Start
  2. Answer questions
  3. Finish

Length and time limit of the Pass-IT mock exam.

Go into your exam with more confidence.

Your Readiness Score helps you assess whether you are ready for the exam. Results by exam domain show you where you should keep practicing.

Your money back if you fail despite 80% Readiness.
If your Readiness Score is at least 80% when you take the exam and you still fail, we refund the Pass-IT purchase price for this certification.

For active, paid access. Submit proof of failing within 14 days.

Further guarantee conditions

You need at least 80% Readiness and active, paid access to the certification when you take the exam. Take the exam within your access period and submit official proof of failing within 14 days after the exam.

For purchases through Apple or Google, refunds are handled by the respective store. See the full terms for further details.

Read the guarantee terms →

The Readiness Score describes your practice progress, not a probability of passing.

Guarantee requirement

80%

At least 80% Readiness
when you take the exam

Money back if you fail

Question counts and mock exam settings from the current Pass-IT catalog.

Exam domains

What you practice for DB-MLP.

Exam topics, relevant practice and sources you can check.

Develop scalable models and featuresSee the guide for weighting132 QuestionsPass-IT questions

Choose training, experiment and feature systems that fit production needs.

  • Select SparkML or single-node models for the data and batch, streaming or real-time inference needs; build SparkML pipelines with suitable estimators and transformers, tune with MLlib, evaluate and score them.
  • Scale training with SparkML, pandas Function APIs or UDFs and compare Ray with Spark; choose vertical or horizontal scaling and model or data parallelism, tune with Optuna plus MLflow or Ray, and train and score group-specific models in parallel.
  • Track complex experiments with nested MLflow runs and programmatically log custom metrics, parameters and artifacts.
  • Create custom model objects with real-time feature engineering, preserve point-in-time correctness to prevent leakage and automate feature computation with FeatureEngineering Client.
  • Configure online tables with Databricks SDK, build streaming feature pipelines and serve on-demand features consistently across training and production.
Official Databricks objectives
Operate the model lifecycleSee the guide for weighting118 QuestionsPass-IT questions

Automate testing, monitoring and retraining across environments.

  • Design and implement lifecycle pipelines for promoting code across environments, map Databricks features to lifecycle tasks and configure scalable ML environments and assets with Databricks Asset Bundles.
  • Organize and implement unit tests for notebook functions and integration tests spanning features, training, evaluation, deployment and inference, choosing tests for each environment.
  • Trigger automated retraining from drift or performance alerts and select the strongest replacement models.
  • Choose and build Lakehouse Monitoring monitors for snapshot, time-series or inference tables; apply statistical drift tests to numerical or categorical data, compare baselines or time windows and alert on significant changes.
  • Build logging, drift, performance and health monitoring; evaluate inference trends, custom metrics, granularities and feature slices and track endpoint latency, traffic, errors, CPU and memory.
Official Databricks objectives
Roll out and serve custom modelsSee the guide for weighting75 QuestionsPass-IT questions

Choose a deployment strategy and expose models to applications.

  • Compare blue-green and canary strategies for high-traffic applications and implement the rollout with Databricks Model Serving.
  • Register custom PyFunc models and their artifacts in Unity Catalog.
  • Deploy custom model objects through MLflow Deployments SDK, REST API or the UI and query them through the SDK or REST API.
Official Databricks objectives
Question formats you can practice
Multiple Choice
Select one answer.

These are Pass-IT practice formats. Check the Databricks certification page for the official exam information.

Frequently asked questions

Your preparation

How much study time should I plan for DB-MLP?

For planning, Pass-IT estimates roughly 70–180 hours. Relevant experience may put you toward the lower end; starting from scratch may put you toward the higher end. Your needs may fall outside this range. This is not a time requirement from Databricks or a promise that you will be exam-ready within that time.

What experience does the Machine Learning Professional assume?

One year or more of hands-on machine learning work on Databricks. The catalog budget is 120 hours on top of that experience, not instead of it. Nothing gates the booking, so the gate is effectively self-imposed.

Using Pass-IT

What language are Pass-IT questions in?

Questions in the app are currently in English. Translations of the sample questions on this page apply only to this preview. Check with Databricks for the languages available for the official exam.

How long does my DB-MLP access last?

Access covers this certification for 12 months. You pay once, with no automatic renewal.

When does the money-back guarantee apply?

The guarantee refunds the Pass-IT purchase price for this certification. You need eligible paid access and a Readiness Score of at least 80% from your practice results when you take the exam. This score can only reach 80% after at least 50 answered questions and is not a validated probability of passing. Take the exam during your access period and submit official proof of failing within 14 days. Guarantee terms

About the DB-MLP exam

Where can I take the DB-MLP exam?

Online proctored or test center.

How long is the Databricks Certified Machine Learning Professional exam?

The Databricks Certified Machine Learning Professional exam has 59 questions and a 120-minute time limit.

Do you need the Machine Learning Associate first?

No, Databricks sets no certification prerequisite. The difference is where the weight sits: the associate exam puts 38% on the platform tooling, while the professional exam puts 44% on MLOps. If pipelines, monitoring and retraining are not yet part of your week, the associate exam is the more honest starting point.

How long is the Machine Learning Professional valid?

Two years from the pass date, renewed by sitting the current version again. Databricks has no continuing-education programme, so plan the resit rather than expecting a credit route to appear.

Your preparation for DB-MLP

$29.99one-time

12 months of access for DB-MLP · No automatic renewal

Buy DB-MLP access

Questions in the app are currently in English.

Try 20 questions for freeFree account required. No credit card.

Already have an account? Log in

Prefer a subscription? View subscriptions →

Money back if you fail with at least 80% Readiness, for eligible purchases under the guarantee terms.

Guarantee terms

The guarantee refunds the Pass-IT purchase price for this certification. You need eligible paid access and a Readiness Score of at least 80% from your practice results when you take the exam. This score can only reach 80% after at least 50 answered questions and is not a validated probability of passing. Take the exam during your access period and submit official proof of failing within 14 days.

Full guarantee terms →

Related certifications

Databricks Certified Associate Developer for Apache SparkDatabricks Certified Data Analyst AssociateDatabricks Certified Data Engineer AssociateDatabricks Certified Data Engineer ProfessionalAll Databricks certifications →