Free Databricks ML Professional sample questions
Try these before reading the answers. Four come from the two 44% sections, matching the real weighting.
Question 1. A model validated at 0.94 AUC performs at 0.71 in production from day one. Investigation finds a feature derived from a field populated only after the outcome is known. What has occurred?
- A. Concept drift
- B. Overfitting to the training set
- C. Data leakage — the feature is unavailable at prediction time
- D. Insufficient model capacity
Show answer
Answer: C
A feature containing information unavailable at prediction time is data leakage, which inflates validation scores and collapses in production immediately rather than gradually. Concept drift develops over time, overfitting would show as a train/validation gap during development, and insufficient capacity would not produce a strong validation score.
Question 2. A production model's accuracy has declined gradually over five months. The input feature distributions are unchanged, but the relationship between those features and the outcome has shifted as customer behaviour changed. What is this?
- A. Data drift
- B. Concept drift
- C. Data leakage
- D. An undersized serving endpoint
Show answer
Answer: B
Unchanged inputs with a changed input-to-outcome relationship is concept drift, as distinct from data drift where the input distribution itself moves. Data leakage appears immediately rather than gradually, and neither an undersized endpoint nor a stale registry entry would degrade accuracy this way.
Question 3. An automated retraining pipeline promotes each newly trained model straight to production. Quality has become erratic. What control is missing?
- A. Validation of the retrained model against the current production model before promotion
- B. More frequent retraining so the model stays current
- C. A larger cluster for the retraining job
- D. Disabling automated retraining entirely
Show answer
Answer: A
A retrained model is not automatically better, so it must be evaluated against the incumbent on a consistent holdout before promotion. More frequent retraining amplifies the problem, a larger cluster affects training speed, and disabling automation abandons the benefit rather than fixing the gap.
Question 4. Regulators ask which model version produced a specific prediction eight months ago, and what data and parameters created that version. What must have been in place?
- A. Endpoint uptime and latency metrics retained for a year
- B. A description of the modelling approach in the team wiki
- C. A copy of the current model artefact in cloud storage
- D. Registered model versions linked to tracked runs, plus logging of which version served each prediction
Show answer
Answer: D
Answering this requires the registered model version linked to its tracked run, including data version, code version and parameters, together with prediction logging that records which version served each request. Endpoint uptime metrics, a wiki description and the current artefact alone cannot reconstruct a historical prediction.
Question 5. A model scores 20 million records once weekly. Latency per record is irrelevant and cost matters. Which serving approach fits?
- A. A real-time serving endpoint held continuously available
- B. Scheduled batch inference
- C. Structured streaming inference
- D. A notebook run manually each week
Show answer
Answer: B
Batch inference on a schedule processes large volumes without maintaining persistent serving capacity, which is the cheapest fit when latency does not matter. A real-time endpoint bills continuously for capacity used once a week, streaming suits continuously arriving data, and manual notebook runs are not a production deployment.
How did you do?
Questions 2, 3 and 4 are ML Ops — 44% of the exam and the section with no Associate equivalent. They are about problems that only appear after a model has been live for months.
If those felt least certain, that is the expected pattern and it is where weeks 3 and 4 of the study plan go.