Free Databricks Certified Machine Learning Professional Practice Questions

Sample Practice Questions

Try these 10 sample questions from the Databricks Certified Machine Learning Professional practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1A fitted Spark PipelineModel receives a new batch of records containing the required input columns. How should the engineer generate predictions using the feature processing learned during training?

   1. A. Call fit() on the PipelineModel with the new records.
   2. B. Call transform() on the PipelineModel with the new records.
   3. C. Call fit() on each original Estimator with the new records.
   4. D. Call transform() on the unfitted Pipeline with the new records.

   Answer: B

   Explanation: The guide states, "Each stage’s transform() method updates the dataset and passes it to the next stage." A fitted PipelineModel runs its learned Transformer stages in sequence; fitting the new batch would train again rather than apply the learned pipeline.

   Reference: ML Pipelines — Pipeline — How it works

2. 2A team adopts Databricks' deploy-code strategy for moving ML work from development to production. Which promotion flow follows that strategy?

   1. A. Copy a development model binary and omit production training code.
   2. B. Promote reviewed pipeline code and train with it in production.
   3. C. Promote only a notebook's cached predictions to production.
   4. D. Reuse the developer's interactive session as the production job.

   Answer: B

   Explanation: Databricks explains that this strategy ensures "the production version of the model is trained on production code." Promoting the pipeline code subjects the ML logic to review and integration testing. Moving only binaries or predictions, or retaining a development session, does not implement that flow.

   Reference: MLOps workflows on Databricks — General recommendations — Deploy code, not models

3. 3Which traffic progression characterizes a canary rollout when introducing a new served model version?

   1. A. Send all traffic to the new version immediately after its build completes.
   2. B. Validate a standby new version, then switch all traffic in one cutover.
   3. C. Increase the new version's traffic share incrementally while monitoring it.
   4. D. Mirror traffic to the new version while the incumbent supplies all user-visible results.

   Answer: C

   Explanation: The official guide describes "incrementally increasing traffic (canary)". Canary progression gradually exposes user traffic to the new version while monitoring results. Immediate replacement and one-step standby cutover do not increase exposure incrementally; mirrored evaluation with incumbent-only user responses does not shift user-visible serving traffic to the new version.

   Reference: Databricks Certified Machine Learning Professional exam guide — Page 11, Answer Key & Explanation, Question 7

4. 4A Spark ML pipeline has two independent feature branches that converge at a downstream stage. The input and output column names define an acyclic dependency graph. In what order must its stages be supplied?

   1. A. In alphabetical order of output column names.
   2. B. With every Estimator before every Transformer.
   3. C. In topological order of the data dependencies.
   4. D. With every Transformer before every Estimator.

   Answer: C

   Explanation: The documentation states, "If the Pipeline forms a DAG, then the stages must be specified in topological order." Stage type or column-name sorting does not ensure that each required input has been produced.

   Reference: ML Pipelines — Pipeline — Details — DAG Pipelines

5. 5A training task must pass a reference through task values so that the validation task can directly load the exact model produced by this execution. Which reference should the training task pass?

   1. A. The experiment ID containing all of the training runs.
   2. B. The Champion alias URI identifying the current deployed model.
   3. C. The training run ID without a model artifact or model identifier.
   4. D. The model URI returned for this training task's model output.

   Answer: D

   Explanation: Databricks states, "The model training task yields a model URI that the model validation task can use." That reference directly identifies the current task's persisted model. An experiment identifies a run collection, Champion can point to a different model, and a run ID alone is not the model URI needed for direct model loading.

   Reference: MLOps workflows on Databricks — Development stage — 4. Train model (development)

6. 6A multi-model endpoint routes 90% of general traffic to current and 10% to challenger. A release-validation probe calls the challenger-specific served-model invocations URL. How is that probe routed?

   1. A. Directly to challenger, ignoring the endpoint traffic split.
   2. B. To current with 90% probability and challenger with 10% probability.
   3. C. To both models, with the two responses averaged.
   4. D. To the highest-numbered registered version of either model.

   Answer: A

   Explanation: Databricks states, "While querying the individual served model, the traffic settings are ignored." The model-specific URL explicitly selects challenger. It does not use weighted endpoint routing, merge responses, or resolve the latest registry version.

   Reference: Serve multiple models to a model serving endpoint — Query individual models behind an endpoint

7. 7A Spark LogisticRegression instance has maxIter set to 12 through a setter. A call to fit() passes a ParamMap assigning that same instance's maxIter to 30. Which limit does this fit use?

   1. A. 12 iterations
   2. B. 18 iterations
   3. C. 42 iterations
   4. D. 30 iterations

   Answer: D

   Explanation: The documentation states, "Any parameters in the ParamMap will override parameters previously specified via setter methods." The explicit fit-time value therefore overrides 12; it is neither added to nor subtracted from the setter value.

   Reference: ML Pipelines — Parameters

8. 8A newly trained model version is registered as prod.risk.fraud_detector but has not passed deployment checks. What does its prod catalog establish?

   1. A. Its environment and governance context, without establishing live deployment status.
   2. B. Its automatic replacement of the endpoint's currently served model version.
   3. C. Its successful completion of every pre-deployment validation requirement.
   4. D. Its assignment to the Champion alias for all inference workloads.

   Answer: A

   Explanation: Databricks states that the enclosing catalog reflects environment and governance rules, "but not its deployment status." Registration in prod therefore does not prove validation, traffic routing, or Champion assignment.

   Reference: Manage model lifecycle in Unity Catalog — Train and register Unity Catalog-compatible models

9. 9A running custom-model endpoint receives a configuration update that is still building its new serving configuration. What happens to prediction traffic before that new configuration is ready?

   1. A. It is sent to the partially built new configuration.
   2. B. It is divided equally between old and incomplete configurations.
   3. C. It is rejected until the new configuration finishes building.
   4. D. It continues to use the existing active configuration.

   Answer: D

   Explanation: Databricks states, "Until the new configuration is ready, the old configuration keeps serving prediction traffic." Pending configuration construction does not automatically expose a partial build or interrupt the active serving configuration.

   Reference: Create custom model serving endpoints — Update a model serving endpoint

10. 10An engineer needs two HashingTF stages with different input columns in the same Spark Pipeline. Which construction satisfies Spark's stage identity requirement?

    1. A. Create a separate HashingTF instance for each stage.
    2. B. Insert the same HashingTF instance into both positions.
    3. C. Change one instance's input column between insertions.
    4. D. Reuse one instance and assign different output aliases.

    Answer: A

    Explanation: The documentation states, "A Pipeline’s stages should be unique instances." Distinct instances have distinct IDs. Changing column parameters or aliases on a shared instance does not create a second stage identity.

    Reference: ML Pipelines — Pipeline — Details — Unique Pipeline stages