Home / ML Associate practice test / Databricks Machine Learning

Free · 8 questions with explanations

Databricks Machine Learning: Databricks Machine Learning Associate Practice Questions

Exam-style questions on Databricks Machine Learning. Pick your answer, then open the explanation to see why it's right — and why the other options are wrong.

Updated November 10, 2026

1 Databricks Machine Learning

A machine learning engineer creates a new Databricks Runtime ML cluster and wants to access Unity Catalog feature tables from it. Which statement is correct according to the current Databricks Runtime ML documentation?

  1. AUnity Catalog access from Databricks Runtime ML is available only when Photon is disabled, because Photon and governed data access use incompatible execution paths on ML clusters
  2. BUnity Catalog access works only on non-ML runtimes, because Databricks Runtime ML is intended for isolated experimentation and not for governed data or model workflows
  3. CTo access Unity Catalog data on Databricks Runtime ML, the compute must use Dedicated access mode, and selecting the Machine learning checkbox in the compute UI sets that access mode automatically
  4. DUnity Catalog access requires a workspace model registry first, because feature engineering and model governance must be configured in the legacy registry before catalog access is enabled
  5. EUnity Catalog access is automatic for any Databricks Runtime ML cluster regardless of access mode, because the runtime itself grants the necessary data permissions at cluster startup time
Show answer & explanation

Correct answer: C

WHY C: The current Databricks Runtime ML documentation states that Unity Catalog access on Runtime ML requires Dedicated access mode, and the compute UI sets that mode automatically when Machine learning is selected. WHY NOT A: Photon is not the deciding factor for UC access. WHY NOT B: Runtime ML supports governed workflows. WHY NOT D: UC does not require the legacy workspace registry first. WHY NOT E: Access mode and privileges still matter.

2 Databricks Machine Learning

A manager opens an MLflow run page in Databricks and wants to know what information can be inspected there. Which answer is correct?

  1. AThe run page shows the final model URI, while parameters, metrics, source notebook details, and artifacts are retrieved through the MLflow Python client instead
  2. BThe run page shows details such as run ID, parameters, metrics, tags, and artifacts, together with a link back to the notebook source that created the run
  3. CThe run page shows artifact file names, while run parameters and metrics are kept on the experiment page and are not repeated on the individual run details page
  4. DThe run page shows the model signature, while source notebook links, run tags, and artifact navigation are found on the Unity Catalog model page for that model
  5. EThe run page shows notebook comments and cluster cost estimates, serving as a collaboration space rather than a place to inspect training metadata or outputs
Show answer & explanation

Correct answer: B

WHY B: Databricks MLflow run pages expose run metadata including parameters, metrics, tags, artifacts, run details, and source notebook links. WHY NOT A: Those details are visible in the UI, not just through the Python client. WHY NOT E: The run page is not limited to collaboration notes; it is where training metadata and outputs are inspected. WHY NOT C: Parameters and metrics are available on the run page as well as on the experiment page. WHY NOT D: Source links and tags are also available at the run level, along with artifact navigation.

3 Databricks Machine Learning

A data scientist has a labeled dataset but is unsure which algorithm family and hyperparameter ranges are likely to work best. She wants a Databricks capability that can produce strong baseline candidates quickly without manually coding every trial. How does AutoML help in this situation?

  1. AAutoML permanently edits the source Delta tables so only statistically important columns remain, guaranteeing that all later training pipelines will use exactly the same reduced feature set without review
  2. BAutoML creates a single fixed notebook that trains one default algorithm and prevents later inspection of preprocessing or tuning choices, because reproducibility requires hiding internal trial logic from users
  3. CAutoML replaces feature engineering entirely by publishing every input column to an online store first, then using feature serving latency as the criterion for model and feature selection
  4. DAutoML selects a winning model by comparing notebook execution time only, because Databricks treats the fastest training job as the most cost-effective proxy for generalization quality
  5. EAutoML orchestrates trials across multiple algorithms and hyperparameter settings, prepares the data, ranks candidate models with evaluation metrics, and generates notebooks so the user can inspect and refine the selected approach
Show answer & explanation

Correct answer: E

WHY E: Databricks AutoML cleans and prepares data, trains and tunes multiple algorithms, ranks results with evaluation metrics, and generates source notebooks for inspection and refinement. WHY NOT A: AutoML does not permanently rewrite source tables to enforce feature selection. WHY NOT B: It is not limited to one hidden default algorithm. WHY NOT C: Online feature publishing is unrelated to AutoML baseline selection. WHY NOT D: Model choice is based on evaluation, not training speed alone.

4 Databricks Machine Learning

A regression experiment evaluates models with RMSE, and lower values are better. The engineer wants the single best active run using the MLflow API rather than the UI. Which option reflects the correct logic?

  1. ACall MlflowClient.search_registered_models and order by model version number ascending, because lower version numbers in Unity Catalog correspond to earlier and better training results
  2. BSearch the experiment runs with a filter for active runs and order by the RMSE metric ascending, because the first result is then the run with the lowest error for that objective
  3. CUse MlflowClient.restore_experiment to make deleted runs visible and then select the last created run, because the newest run is the one MLflow treats as best after metric logging
  4. DUse MlflowClient.set_terminated on poorly performing runs first, because terminated runs are the ones the search API includes in metric comparisons and active runs are skipped
  5. ERead the artifact directory directly from DBFS and choose the run with the largest model file, because a larger saved model indicates lower regression error and better feature use
Show answer & explanation

Correct answer: B

WHY B: For metrics where lower is better, the correct approach is to search runs and order the metric ascending (optionally filtered to active runs), then take the top result. WHY NOT A: Model registry search is not how experiment runs are ranked, and version numbers say nothing about quality. WHY NOT C: Newest is not equal to best; MLflow does not treat the last created run as the winner. WHY NOT D: Active runs can still be searched and compared; set_terminated only changes run status. WHY NOT E: Artifact file size is not a valid model-quality criterion.

5 Databricks Machine Learning

During a custom evaluation job, an engineer saves a confusion-matrix image to a local path and wants it attached to the current MLflow run together with the trained sklearn model. What should she do?

  1. AWithin the active run, copy the image to a Unity Catalog volume and log its path as a param, as files written to local disk in the session are not accepted as run artifacts
  2. BWithin the active run, log metrics and params with MLflow and upload the image through Catalog Explorer, as the run UI does not retain arbitrary files beside a logged model
  3. CRegister the sklearn model first and share the image with the team by email, as model registration closes the active run and blocks further artifact uploads to it
  4. DWithin the active run, call mlflow.log_artifact with the saved image path and log the trained model with the sklearn MLflow flavor, placing both under the same run
  5. EWithin the active run, store the confusion matrix as a run tag value and serialize the model as JSON text, as the tracking server does not accept binary files
Show answer & explanation

Correct answer: D

WHY D: mlflow.log_artifact attaches local files such as images to the active run, and MLflow model flavors are used to log trained models in the same run, so both appear in the run metadata and artifacts. WHY NOT A: Local files can be logged as artifacts, including ones generated in the same notebook session. WHY NOT B: Arbitrary artifacts are supported and are displayed and retained alongside a logged model. WHY NOT C: Logging artifacts and models can happen within the run workflow; registration does not close the run. WHY NOT E: Tags are not a substitute for binary artifacts or model logging.

6 Databricks Machine Learning

A development workspace needs to compare a newly trained candidate with a production baseline model registered elsewhere, but both workspaces are attached to the same metastore. Which Unity Catalog benefit is most relevant?

  1. AUnity Catalog keeps models visible inside the workspace where they were created, which protects production baselines from accidental comparison in development environments
  2. BUnity Catalog has each workspace copy a model version locally before loading it, which improves reproducibility by keeping development and production on separate artifacts
  3. CUnity Catalog provides cross-workspace access for fine-tuned LLMs, while traditional ML models still rely on the workspace registry and are not loadable by name elsewhere
  4. DUnity Catalog changes where notebook artifacts are stored, while model discovery, privileges, and reuse stay scoped to each workspace even on a shared governed metastore
  5. EModels in Unity Catalog are accessible from workspaces attached to the same metastore, which makes it easier to compare candidates against governed baselines without duplicates
Show answer & explanation

Correct answer: E

WHY E: Unity Catalog model registration supports cross-workspace access within the same metastore, which is valuable for governed comparisons and reuse without duplicating registry entries. WHY NOT A: Unity Catalog does not restrict model visibility to the creating workspace; that is the legacy workspace registry's behavior. WHY NOT B: Local copying is not required; a model version is loaded by its three-level name from any attached workspace. WHY NOT C: Unity Catalog supports traditional ML models as well as fine-tuned LLMs. WHY NOT D: Unity Catalog materially improves governance, discovery, and reuse of models, not just artifact storage.

7 Databricks Machine Learning

A company has three Databricks workspaces for development, staging, and production. Multiple teams need to discover and reuse governed feature tables across those workspaces without recreating them locally. Why is storing feature tables in Unity Catalog generally preferable to legacy workspace-level Feature Store tables?

  1. AWorkspace-level feature tables are preferable, as they are queryable from each workspace by default, whereas Unity Catalog feature tables stay hidden outside the workspace where they were first created
  2. BWorkspace-level feature tables are preferable, as lineage, permissions, and discovery are more granular there, while Unity Catalog limits a table to one owner and blocks cross-workspace sharing by design
  3. CUnity Catalog feature tables are preferable for teams without centralized governance needs, as account-level feature tables are harder to discover and are not taggable or searchable in the Databricks UI
  4. DUnity Catalog feature tables are preferable, as they provide centralized governance, lineage, and cross-workspace access through the metastore and are easier to share than workspace-scoped feature assets
  5. ENeither is preferable in practice, as account-level and workspace-level feature tables are stored with identical sharing, privilege, and model-serving behavior in current Databricks releases
Show answer & explanation

Correct answer: D

WHY D: Unity Catalog feature tables are governed centrally and can be accessed across workspaces attached to the same metastore, with lineage and consistent privileges. WHY NOT A: The sharing behavior is the opposite: workspace-level tables are scoped to one workspace, and Unity Catalog tables are the ones visible across workspaces. WHY NOT B: Unity Catalog improves governance and sharing rather than weakening it; it does not block cross-workspace sharing. WHY NOT C: Unity Catalog is the choice for centralized governance, and UC feature tables can be discovered, searched, and tagged. WHY NOT E: There are meaningful governance and cross-workspace differences between the two.

8 Databricks Machine Learning

A team wants full control over what is recorded during a training run rather than relying on autologging. Which pattern correctly performs manual MLflow logging?

  1. ACreate a notebook experiment, fit the model, and let the hosted tracking server infer the desired parameters, metrics, artifacts, and model once the run has finished
  2. BWrite metrics and parameters to a Delta table and register the model in Unity Catalog, as manual MLflow logging covers artifacts but not scalar values in a run
  3. COpen an MLflow run and explicitly call APIs such as log_param, log_metric, log_artifact, and the appropriate log_model flavor to record the chosen metadata, files, and model
  4. DOpen an MLflow run and call MlflowClient.set_tag for each result, as parameters, metrics, artifacts, and models are stored internally as tags on the run page
  5. ECall create_training_set before fitting and log through the feature store client, as manual MLflow logging is offered for feature-store-backed models
Show answer & explanation

Correct answer: C

WHY C: Manual MLflow logging is done by explicitly logging params, metrics, artifacts, and models within a run. WHY NOT A: MLflow does not infer all desired metadata after the fact. WHY NOT B: MLflow can log scalar metrics and parameters directly. WHY NOT D: Tags are not substitutes for metrics, artifacts, and models. WHY NOT E: Manual logging works broadly (scikit-learn, XGBoost, custom Python workflows), not just for feature-store-backed models.

Take the full ML Associate practice test →

From the free full-length test

17 more Databricks Machine Learning questions with answers

Each opens on its own page with the correct answer and the full explanation.

All free Machine Learning Associate exam questions & answers →