About these questions
These five questions are original. We wrote them in the style of the exam objectives, with the sections in rough proportion to their weights: two on Databricks Machine Learning, and one each on data processing, model development, and deployment. They are not real exam questions, and the real exam will word things differently.
How to use these
Try each question without looking at your notes, and pick an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Question 3 asks for two answers. If a section felt shaky, go back to the domains explained page.
Question 1 (Databricks Machine Learning)
A bank retrains a fraud model every night. Policy says transaction data may never be copied out of the production workspace. The team debates what to move from the development workspace into production. Which approach fits the policy?
- A. Copy a sample of the transactions into development, train there, and promote the model
- B. Promote the trained model from development and retrain it nightly in development
- C. Promote the training code to production and train the model there on the production data
- D. Promote the model, then rerun it against staging data to refresh it
Show the answer and explanation
Correct answer: C. When the data cannot leave production, the code is what moves between environments, and the model is produced inside production by running that code on the real data. Promoting a model artifact only works when training can happen where the data is available to the development side. Option A breaks the policy outright, because it copies production data out. Option B, the closest wrong option, keeps training in development, where the nightly production data is not allowed to go. Option D confuses scoring with retraining.
Question 2 (Databricks Machine Learning)
A nightly job calculates fresh loyalty-point features only for the shoppers who bought something today. It must write them into an existing feature table in Unity Catalog, keyed by shopper ID, and leave every other shopper’s features alone. Which write behavior is needed?
- A. Overwrite, which replaces the table content with the new batch
- B. Merge, which updates existing keys, inserts new ones, and leaves other rows untouched
- C. A full rebuild from the incoming batch alone
- D. Append, which adds the batch as new rows next to the existing ones
Show the answer and explanation
Correct answer: B. A merge matches rows on the primary key. Shoppers in the batch get their features updated, new shoppers are inserted, and everyone else stays as they were. Append, the closest wrong option, keeps old rows but would add a second row for each shopper who is already in the table, so the key would no longer be unique. Overwrite and a full rebuild both replace the whole table, which would wipe out the features of everyone who did not shop today.
Question 3 (Data Processing)
A numeric column has 15 percent of its values missing. The team fills every missing value with the mean of the column. Which two statements about this choice are correct? Choose two.
- A. The mean of the column changes noticeably, because the filled value comes from the missing rows
- B. The relationship between the column and the target is left exactly as it was
- C. A tree-based model can no longer use the column, because it needs complete cases when it is fitted
- D. The spread of the column shrinks, because many rows now hold exactly the same value
- E. Adding an indicator column that marks which rows were filled preserves the fact that a value was missing
Show the answer and explanation
Correct answers: D and E. Filling many rows with one number piles them onto a single point, so the variance of the column drops and its link to the target gets weaker. A boolean indicator for the filled rows keeps the information that the value was absent, which is sometimes predictive in itself. Statement A is the closest wrong option. Filling with the mean leaves the mean of the column essentially unchanged, so it does not shift. Statement B is wrong because the spread loss weakens the relationship, and statement C is wrong because trees work fine once the gaps are filled.
Question 4 (Model Development)
One team trains a house-price model on the log of the price and reports an RMSE of 0.21. Another team trains on the raw price and reports an RMSE of 48,000. A manager compares the two numbers in a review. What is the problem?
- A. There is no problem, because a log transformation is monotonic and preserves the ranking of errors
- B. The comparison only fails if the teams used different metrics, not different transformations
- C. There is no problem, because the log-model error equals the raw error divided by its mean
- D. The errors are on different scales, so the log model’s predictions must be exponentiated back before its error is compared
Show the answer and explanation
Correct answer: D. An error on the log scale describes a multiplicative deviation, while an error on the raw scale is in currency units, so the two numbers cannot be compared even though both are called RMSE. The log model’s predictions have to be converted back with the exponential function, and its error recomputed in the original units. Option B is the closest wrong option. It is true that different metrics would also break the comparison, but the same metric on different scales fails too. Options A and C describe properties that do not make the numbers comparable.
Question 5 (Model Deployment)
A model serving endpoint currently sends 75 percent of its traffic to version 1 of a model and 25 percent to version 2. The team wants an even 50/50 split and the endpoint must keep answering requests meanwhile. What should they do?
- A. Delete the endpoint and recreate it with the new percentages, because the split is fixed at creation
- B. Remove version 1, because traffic is assigned by the order in which versions were added
- C. Create a second endpoint at the new split and ask every caller to change the URL they use
- D. Update the endpoint’s served entities configuration with the new percentages
Show the answer and explanation
Correct answer: D. Traffic percentages are part of the endpoint configuration, and updating that configuration changes the split while the endpoint continues to serve requests. That is also what makes gradual rollouts and fast rollbacks possible. Option A, the closest wrong option, would work only with downtime and for no good reason. Option C forces every caller to change. Option B is wrong because the split is set by the configured percentages, not by the order of versions.
More guides for DATABRICKS-ML-ASSOCIATE
- Databricks ML Associate Study Plan
- Is the Databricks ML Associate Exam Hard?
- Databricks ML Associate Prerequisites
- Databricks ML Associate Renewal and Retakes
- Is Databricks ML Associate Worth It?
- Databricks ML Associate Cheat Sheet
- Databricks ML Associate vs Professional
- GenAI vs ML Associate: Which Databricks Exam?
Next step
If these felt comfortable, read the exam format page and plan your booking. If a section felt shaky, try more questions in the Databricks ML Associate practice tests.