Databricks ML Professional practice test
Twenty questions across the three sections of the Databricks Machine Learning Professional exam. Model development and MLOps each carry a large share, so they dominate here. Expect production judgement rather than algorithm theory.
Model Development
Question 1. A feature is available in the warehouse at training time but only arrives three days after the event in production. What is the consequence?
- A. The model will simply be slower in production
- B. The model cannot be served as evaluated; the feature is unavailable at inference time
- C. Nothing, as the warehouse will backfill it
- D. The model will be more accurate in production
Show answer
Answer: B
The feature cannot be used at inference, so training with it produces a model that cannot be served as evaluated. This is a form of leakage caused by availability timing.
Question 2. Training and serving compute a feature with subtly different logic. Predictions are worse than validation suggested. What is this called, and what fixes it?
- A. Training-serving skew; share one feature definition between both paths
- B. Concept drift; retrain more often
- C. Overfitting; regularise the model
- D. Class imbalance; resample the data
Show answer
Answer: A
Training-serving skew is resolved by a single shared feature definition used by both paths, typically through a feature store.
Question 3. A model must be retrained monthly on a rolling window. How should the training data selection be implemented?
- A. Hard-code the date range in the notebook each month
- B. Use whatever data is in the table at run time, unrecorded
- C. Parameterise the window and record the resolved table version with the run
- D. Train on all history every time regardless of the requirement
Show answer
Answer: C
Parameterising the window and recording the resolved table version per run keeps runs reproducible and comparable. Hard-coded dates drift and cannot be reproduced.
Question 4. Hyperparameter tuning across many trials must run efficiently on a cluster. What is the key consideration?
- A. Run all trials serially for reproducibility
- B. Use the largest possible instance for a single trial
- C. Disable logging to save time
- D. Distribute trials in parallel with early termination of unpromising ones
Show answer
Answer: D
Parallelising trials with early termination of unpromising ones uses the cluster efficiently. Serial execution wastes parallel capacity.
Question 5. A model shows strong aggregate performance but fails badly for one customer segment. What should the evaluation include?
- A. Metrics computed per segment, not only in aggregate
- B. A larger test set overall
- C. A longer training run
- D. A different random seed
Show answer
Answer: A
Segment-level evaluation surfaces disparities hidden by aggregate metrics, which is both a quality and a fairness requirement.
Question 6. Two teams report different accuracy for what they believe is the same model. What is the most likely cause?
- A. Different cluster sizes
- B. Different evaluation data, splits or metric definitions
- C. Different notebook themes
- D. Different cloud regions
Show answer
Answer: B
Different evaluation datasets, splits or metric definitions almost always explain divergent numbers for the same artefact.
Question 7. An ensemble improves validation score by 0.3% but triples inference cost and complexity. What is the reasonable decision?
- A. Always ship the higher-scoring model
- B. Always ship the cheaper model without analysis
- C. Decide explicitly against business value, defaulting to the simpler model absent a clear case
- D. Ship both and let users choose
Show answer
Answer: C
A marginal gain that triples cost and complexity usually fails a cost-benefit test; the decision should be made explicitly against business value.
Question 8. A training pipeline must produce identical results when re-run on the same inputs. What must be controlled?
- A. Only the random seed
- B. Only the library versions
- C. Only the input data version
- D. Seeds, library versions, data version and non-deterministic operations together
Show answer
Answer: D
Random seeds, library versions, data version and any non-deterministic operations all affect reproducibility and must be pinned or recorded.
Question 9. An experiment's best run cannot be identified because metrics were logged under different names across runs. What practice prevents this?
- A. A project-wide convention for metric names and logging
- B. Logging fewer metrics
- C. Using a larger cluster
- D. Running fewer experiments
Show answer
Answer: A
A standard metric naming and logging convention applied across the project keeps runs comparable. Ad-hoc naming makes comparison impossible after the fact.
ML Ops
Question 10. A model in production must be rolled back within minutes if quality degrades. What makes this possible?
- A. Rebuilding the previous model from source
- B. A model registry retaining the previous version with a tested rollback procedure
- C. A copy of the notebook in a browser tab
- D. A weekly backup of the workspace
Show answer
Answer: B
A registry holding the previous version with a tested rollback path allows immediate reversion. Rebuilding from source takes far too long.
Question 11. Ground-truth labels arrive 30 days after prediction. How should quality be monitored meanwhile?
- A. Wait for labels and do nothing
- B. Retrain daily as a precaution
- C. Monitor input and prediction distributions as proxies, then compute true metrics when labels arrive
- D. Assume performance is stable
Show answer
Answer: C
Proxy signals such as prediction distribution and input drift give early warning, with true metrics computed once labels arrive. Waiting leaves a month blind.
Question 12. A new model version must be validated on live traffic without risking all users. Which approach fits?
- A. Canary or shadow deployment with measured comparison
- B. Full replacement with close monitoring
- C. Deployment to development only
- D. Deployment during a maintenance window to all users
Show answer
Answer: A
A canary or shadow deployment exposes the new version to a small share of traffic, or runs it alongside without serving its output, allowing measurement before full rollout.
Question 13. An automated retraining pipeline occasionally promotes a worse model. What control is missing?
- A. More frequent retraining
- B. A larger training set
- C. More alerting on job failures
- D. A promotion gate comparing candidate against incumbent on a held-out set
Show answer
Answer: D
A promotion gate that compares the candidate against the incumbent on a held-out set, with a minimum improvement threshold, prevents regressions.
Question 14. Which monitoring signal distinguishes concept drift from data drift?
- A. Latency versus throughput
- B. Input distribution change versus change in the input-to-target relationship
- C. CPU usage versus memory usage
- D. Batch size versus learning rate
Show answer
Answer: B
Data drift is a change in the input distribution; concept drift is a change in the relationship between inputs and the target, visible only once labels arrive.
Question 15. An incident requires reconstructing why a specific prediction was made six months ago. What must have been retained?
- A. Monthly aggregate accuracy
- B. The cluster’s event log
- C. Model version, served input features and configuration for that request
- D. The name of the on-call engineer
Show answer
Answer: C
The model version, input features as served, and configuration at that time are required to reconstruct the decision. Aggregate metrics cannot explain one prediction.
Question 16. A feature pipeline and a training pipeline are owned by different teams and break each other regularly. What practice addresses this?
- A. A validated schema and semantics contract between the pipelines, enforced in CI
- B. Weekly coordination meetings only
- C. Merging both teams
- D. Freezing both pipelines
Show answer
Answer: A
A contract on feature schema and semantics, validated automatically in CI, prevents one team’s change breaking the other silently.
Question 17. Costs rise sharply after a model is promoted. Which factor should be checked first?
- A. The training data size
- B. Serving instance type, minimum replicas and autoscaling bounds
- C. The number of features in the model
- D. The notebook’s line count
Show answer
Answer: B
Serving configuration such as instance type, minimum replicas and autoscaling bounds usually explains a step change in cost on promotion.
Question 18. Which practice most reduces the risk of an unreviewed change reaching production?
- A. Monitoring production closely after each deployment
- B. Deploying only on quiet days
- C. Requiring a rollback plan without review
- D. Code review and automated tests as mandatory gates before deployment
Show answer
Answer: D
Requiring code review and automated tests as gates before deployment catches changes before they ship. Post-deployment monitoring detects problems only after users are affected.
Model Deployment
Question 19. An endpoint must serve sub-100ms predictions and compute several features from a feature store at request time. What is the main design risk?
- A. Online feature lookup latency consuming the latency budget
- B. The model file size on disk
- C. The training cluster’s configuration
- D. The number of experiments recorded
Show answer
Answer: A
Feature lookup latency can dominate the budget, so online feature serving latency must be measured and engineered for, not assumed.
Question 20. A model must serve both a nightly batch of 80 million records and low-latency single requests. What is the appropriate design?
- A. One real-time endpoint used for both
- B. One batch job used for both
- C. Separate batch and real-time serving paths sharing the same registered model version
- D. Two independently trained models
Show answer
Answer: C
Separate serving paths sized to each workload, sharing the same registered model version, meets both requirements without compromising either.
How did you do?
Sixteen or more correct suggests you are close. Below fourteen, the section guides here are the fastest fix — this exam rewards operational judgement, and the professional level is a real step up from the associate. Databricks does not publish a passing score.