Databricks Certified Machine Learning Professional

Last checked October 22, 2026 against the official Databricks exam page.

1. Question 1Model Development

   A data scientist logs a model with MLflow autologging, but the team must be able to reproduce the exact preprocessing that was applied before training. What should be logged alongside the model?

   1. AOnly the final fitted model artifact, since the preprocessing can be rebuilt from the notebook later on
   2. BThe text output of each notebook cell from the training session as a plain text attachment on the run
   3. CA screenshot of the training loss curve from the notebook so that the run can be compared visually
   4. DThe full pipeline as a model with its input signature, together with the code version used for training
   Show answer and explanationHide answer

   Correct answer: D

   Logging the full pipeline with its signature and the code version lets the exact preprocessing and input schema be reproduced later. A bare model file, a screenshot, or cell output cannot reconstruct the transformations that produced the training features.

2. Question 2MLOps

   Which kind of test checks that a feature function returns the expected output for known inputs?

   1. AA load test, which measures how the endpoint behaves when many concurrent requests arrive together from clients across the network
   2. BA unit test, which checks one function in isolation against inputs whose expected outputs are known
   3. CA penetration test, which tries to find security weaknesses in the infrastructure from outside it
   4. DA canary deployment, which routes a small share of live traffic to the new version of the service
   Show answer and explanationHide answer

   Correct answer: B

   Unit tests verify individual functions against known inputs and expected outputs, which is what this check requires. Load tests, penetration tests, and canary deployments test performance, security, and rollout, not the correctness of a single function.

3. Question 3Model Deployment

   Which two conditions favor batch inference over a real-time endpoint? (Choose two.)

   1. APredictions are needed for millions of rows on a nightly schedule, with no per-request latency requirement
   2. BEach prediction must return within a few hundred milliseconds in response to a user request on a website
   3. CThe results are written to a Delta table for later reporting and analytics by the business teams involved
   4. DThe model is called once per user click from a web application that runs the scoring for every visitor
   5. EPredictions must be shown in a browser form one row at a time as each row is entered by a user
   Show answer and explanationHide answer

   Correct answer: A, C

   Large scheduled scoring with no latency target is well suited to batch jobs, and writing results to a table serves reporting directly. Sub-second per-request latency, per-click calls, and one-row-at-a-time forms all need an online endpoint instead.

4. Question 4Model Development

   Which Hyperopt option distributes the trials of a hyperparameter search across a Spark cluster?

   1. ATrials, which records the results of each evaluation in memory on the single driver process only
   2. BSparkTrials, which runs each trial as a Spark task so that the search uses the cluster's workers
   3. CCalling fmin on its own, which runs every evaluation sequentially on whichever worker is available
   4. DA plain Python for loop that calls the objective function once per candidate set of parameters
   Show answer and explanationHide answer

   Correct answer: B

   SparkTrials dispatches each trial to the cluster so that the search scales with the number of workers. Trials keeps results on the driver, and fmin alone or a Python loop runs evaluations on a single process rather than distributing them.

5. Question 5MLOps

   Which signal most appropriately triggers the retraining of a production model?

   1. AEvery new row that is added to the source table, so that the model always uses the freshest data available
   2. BThe calendar date alone, so that retraining happens on the same day every month regardless of change
   3. CA significant drop in production performance or a sustained shift in the distribution of the inputs
   4. DA change in the font size used in the notebook that contains the training code for the model in question
   Show answer and explanationHide answer

   Correct answer: C

   Retraining should respond to evidence that the model is degrading, such as falling metrics or sustained input drift. Fixed calendars can waste effort or miss changes, retraining on every row is costly and noisy, and notebook formatting has no bearing on model quality.