Databricks · DB-MLP · Professional

Databricks Machine Learning Professional — Practice Questions and Mock Exam

Name: Pass-IT
Author: Pass-IT

Prepare for DB-MLP with original practice questions and clear answer explanations. Ask Alex, your AI tutor, when you need more detail, use your results to identify topics to review, and practice your pacing with timed mock exams.

Start practicing DB-MLP free Buy DB-MLP access for $29.99

Buy DB-MLP access for $29.99 Start practicing DB-MLP free

59Mock exam questions

120minTime limit

Checked against Databricks · August 2026 · Current exam version

Overview

About the exam

The Databricks Certified Machine Learning Professional exam validates advanced expertise in building production ML systems at scale on Databricks. It covers advanced model development techniques, MLOps lifecycle management including CI/CD and monitoring, and model deployment strategies — model development and MLOps each make up 44% of the exam. Candidates must demonstrate proficiency in production ML pipeline design and model lifecycle management at scale.

This certification is designed for senior ML engineers and data scientists with one or more years building production ML systems on Databricks. It demonstrates the ability to operate ML systems reliably in production at scale.

Try the practice preview

Try five DB-MLP questions

Try five practice questions from the app’s current Databricks Certified Machine Learning Professional question bank, with answers and explanations.

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

Create a free account

Question pool

325 practice questions

The Pass-IT question pool gives you material to practice for DB-MLP. A Pass-IT mock exam uses 59 questions and a 120-minute time limit; these are practice settings.

Pool details: DB-MLP

Exam details checked against DatabricksAugust 13, 2026

date of the last check against the official Databricks source

Objectives in the guide47 objectives listed in the official guide

across 3 domains in the official exam guide

Pool size325 questions

= The pool size is equivalent to 5 sets of 59 questions; this does not mean that each mock exam uses a separate set.

Blueprint domains3 domains in the exam blueprint

Model Development 132 · MLOps 118 · Model Deployment 75

Recorded as checked against sources325 of 325

questions recorded as having their answer, options, and explanation checked against official Databricks documentation

How questions are made →

Exam domains

What's on the exam

Model development and MLOps are tied at 44% apiece, together accounting for the vast majority of the exam. Model development covers distributed training with SparkML, Ray, and Optuna, point-in-time-correct feature engineering, and nested MLflow runs for complex experiments; MLOps covers CI/CD for ML, Lakehouse Monitoring, drift detection, and automated retraining triggered by performance degradation.

Model deployment is the remaining 12%, testing blue-green and canary rollout strategies and reaching custom PyFunc models through a serving endpoint rather than the UI alone. The even split between development and operations reflects what a senior ML engineer does: building models is half the job, keeping them healthy in production is the other half.

Exam blueprint: DB-MLP

Model Development44%

Advanced model development, distributed training, feature engineering at scale, experiment design, hyperparameter optimization with Ray/Optuna, and custom model architectures.

≈ 53 h

MLOps44%

MLOps practices, CI/CD for ML, Databricks Asset Bundles for ML, model testing strategies, A/B testing, model monitoring, drift detection, and pipeline automation.

≈ 53 h

Model Deployment12%

Production model serving, batch vs real-time inference, scaling serving endpoints, and deployment automation.

≈ 14 h

Format

Exam format and question types

The exam consists of 59 scored, multiple-choice questions inside a 120-minute window, with unscored pilot items possible. Expect deep scenario work: choosing Ray or Spark for a distributed training job, designing a Lakehouse Monitoring alert for drift beyond a threshold, or picking a rollout strategy for a high-traffic serving endpoint.

Question types: DB-MLP

Multiple Choice100%

Select the single answer that best meets the question’s requirements.

See Databricks for official question-format information. The shares shown describe the Pass-IT practice pool; they do not establish the proportions on the official exam.

Planning

Preparing for DB-MLP

The exam is delivered online with a remote proctor or at a test center, offered in English only. The credential holds for two years, and recertifying means passing the current version of the exam.

Preparation and logistics: DB-MLP

Preparation

Illustrative study time70–180 h

illustrative planning range: 70 h with relevant experience to 180 h when starting out; your needs may fall outside this range

LevelProfessional

Recommended backgroundNone required. 1+ years hands-on Databricks ML experience recommended.

Taking and maintaining the certification

DeliveryOnline proctored or test center

Retake policyNo mandatory waiting period. Retake fee applies.

Certification validity2 years

Recertification required every 2 years by taking the current exam version.

Official guideView on provider site

Study notes

Common pitfalls

Topics to review: DB-MLP

01Point-in-Time Correctness
Joining features to labels without enforcing point-in-time correctness, so future information leaks into training, leads to wrong answers on feature-engineering questions.
02Online Feature Tables
Serving features from a batch feature table instead of configuring an online table for low-latency lookups leads to wrong answers on real-time feature-serving questions.
03Custom Metrics Logging
Logging only the default MLflow metrics instead of custom parameters and artifacts within nested runs leads to wrong answers on advanced-experiment-tracking questions.
04DAB ML Asset Config
Not declaring a serving endpoint, an MLflow experiment, and a registered model together through Automation Bundles leads to wrong answers on scalable-environment questions.
05Endpoint Health Metrics
Monitoring only model-quality drift while ignoring endpoint infrastructure metrics like latency, error rate, and memory usage leads to incomplete answers on production-monitoring questions.

Practice with questions and explanations in Pass-IT →

Questions

Frequently asked questions

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.

Which pitfalls should I review when preparing for Databricks Certified Machine Learning Professional?

Topics to review include Point-in-Time Correctness, Online Feature Tables, Custom Metrics Logging, DAB ML Asset Config, Endpoint Health Metrics. Work through examples to check that you understand the distinctions and can explain your answer.

How is the Machine Learning Professional exam weighted?

Model development and MLOps take 44% each, with model deployment at 12%. That is an unusually blunt split: the exam is essentially two halves. Anyone who can build models but has never run them in production loses half the paper.

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.

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.

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.

One certification, 12 months

Practice for DB-MLP

Focus your practice on one certification, or choose Pro to practice across all certifications.

Buy DB-MLP access for $29.99

Start a free practice sessionTry the first 20 questions without a card to see whether the practice suits you.

For eligible purchases: money-back guarantee if you fail your exam.

View guarantee terms →