Databricks ML Associate practice test

Question 1. What is the primary purpose of experiment tracking in an ML workflow?

• A. To schedule training jobs
• B. To reduce storage costs
• C. To record parameters, metrics and artefacts so runs are comparable and reproducible
• D. To encrypt the training data

Question 2. Which problem does a feature store primarily solve?

• A. Inconsistent feature computation between training and serving
• B. Slow cluster startup
• C. Insufficient GPU memory
• D. Dashboard rendering performance

Question 3. A model must be promoted from staging to production with an auditable record of who approved it. Which capability supports this?

• A. Copying the model file to a production folder
• B. A model registry with versioned stage transitions
• C. Renaming the notebook
• D. A comment in the training code

Question 4. Which statement about AutoML in a professional workflow is most accurate?

• A. It removes the need to understand the data
• B. It always outperforms hand-tuned models
• C. It should be used as the final model without review
• D. It quickly establishes a baseline and produces code you can review and extend

Question 5. Training data must be reproducible for an audit two years later. What should be recorded with the run?

• A. The table version or snapshot identifier of the training data
• B. The number of rows only
• C. The date the job ran
• D. The cluster’s instance type

Question 6. A team runs dozens of training variants and cannot tell which produced the best result. What is missing?

• A. A larger cluster
• B. More training variants
• C. Consistent logging of parameters and metrics for every run
• D. A faster storage tier

Question 7. Which is the correct reason to use a managed cluster runtime designed for machine learning?

• A. It is always cheaper
• B. It ships compatible ML libraries and drivers, avoiding dependency conflicts
• C. It disables logging by default
• D. It prevents the use of custom libraries

Question 8. An experiment run cannot be reproduced because a library version changed. What should have been captured?

• A. The notebook’s author
• B. The cluster’s uptime
• C. The dashboard used to view results
• D. The environment specification, including library versions

Question 9. A dataset is split randomly into train and test, but the data is a time series. What is the problem?

• A. Random splitting leaks future information into training
• B. The test set will be too small
• C. Random splitting is slower
• D. There is no problem

Question 10. A feature is computed from the target variable. What is the likely outcome?

• A. Training will fail
• B. The model will underfit
• C. Validation scores will look excellent and production performance will be poor
• D. Training will be slower

Question 11. Categorical features must be encoded consistently in training and serving. What is the safest approach?

• A. Re-fit the encoder on each serving batch
• B. Fit the encoder during training and persist it for reuse at serving
• C. Use random integers for categories
• D. Drop categorical features

Question 12. Cross-validation scores vary widely across folds. What does this suggest?

• A. The model is ready for production
• B. The learning rate is too high
• C. The dataset is too large
• D. The performance estimate is unstable and should not be relied on as-is

Question 13. A binary classifier for rare events achieves 99% accuracy but never predicts the positive class. Which metric should replace accuracy?

• A. Precision, recall or area under the precision-recall curve
• B. Accuracy on a larger sample
• C. Training loss
• D. Mean squared error

Question 14. Training loss keeps falling while validation loss rises. What is happening and what helps?

• A. Underfitting; train longer
• B. Overfitting; apply regularisation, early stopping or more data
• C. Data leakage; remove the target
• D. Class imbalance; resample

Question 15. Hyperparameter search must cover a large space within a limited budget. Which strategy is most efficient?

• A. Exhaustive grid search
• B. A single run with defaults
• C. Adaptive or Bayesian search with early stopping of unpromising trials
• D. Manual tuning by intuition

Question 16. A regression model must be evaluated where large errors are disproportionately costly. Which metric fits?

• A. Root mean squared error
• B. Mean absolute error
• C. Accuracy
• D. F1 score

Question 17. Two models perform within half a percentage point of each other; one is far simpler. Which should usually be chosen?

• A. The complex one, because it scores higher
• B. The simpler one
• C. Neither, until a third is built
• D. Both, deployed in parallel permanently

Question 18. A stakeholder asks which features drove a specific prediction. What supports this?

• A. The model’s overall accuracy
• B. The training duration
• C. Global feature importance only
• D. Per-prediction feature attribution

Question 19. Predictions are required for 60 million records overnight, with no latency constraint. Which serving mode fits?

• A. A real-time endpoint running continuously
• B. An interactive notebook run by hand
• C. A scheduled batch inference job
• D. A streaming endpoint per record

Question 20. A deployed model's accuracy declines over a year with no code change. What should monitoring detect?

• A. Data and concept drift against the training baseline
• B. Increased inference latency
• C. Rising storage costs
• D. A larger number of API callers