Databricks Machine Learning Professional · 44% practice allocation

Databricks Machine Learning Professional: ML Ops Practice Questions

5 free sample questions from the ML Ops domain, with answers and explanations. The full Machine Learning Professional bank has 58 questions across 3 domains. Questions are scenario-based and written from the official Databricks exam guide.

1.

Which testing practice is appropriate for the FEATURE ENGINEERING code in an ML pipeline?

  • A. Unit tests that assert transformation outputs for known inputs (e.g., correct encoding, no nulls introduced)✓ Correct
  • B. Only monitoring the final model's accuracy in production and rolling back when it drops below a threshold
  • C. No unit tests, since ML pipelines are non-deterministic; rely on model validation metrics to catch feature bugs
  • D. Manually inspecting a sample of transformed rows in a notebook display once before the first job run is deployed
Explanation

Feature engineering is deterministic transformation logic, so unit tests asserting expected outputs for known inputs catch regressions early. Skipping unit tests in favor of model metrics, relying only on production accuracy monitoring, or a one-time manual inspection of rows all leave defects to surface late.

2.

Which practice BEST enables CI/CD for ML projects on Databricks, packaging notebooks, jobs, and configuration for deployment across environments (dev/staging/prod)?

  • A. Manually exporting notebooks and importing them into each target workspace per release
  • B. Emailing .dbc archives to each environment owner to import and configure job settings
  • C. Databricks Asset Bundles (define resources as code and deploy across environments)✓ Correct
  • D. Editing jobs directly in the production workspace UI and cloning them back to dev after
Explanation

Databricks Asset Bundles define jobs, pipelines, and configuration as code (YAML) for versioned, repeatable deployment across dev/staging/prod — the modern MLOps CI/CD approach. Manual export/import, emailing .dbc archives, and editing production directly in the UI are error-prone and unrepeatable.

3.

A registered model version should not be promoted to production until it passes an automated validation (e.g., meets a minimum metric on a holdout set). How is this typically implemented?

  • A. Every newly registered version is automatically given the production alias, and metrics are only reviewed after deployment
  • B. Versions are kept in the registry but never promoted, and consumers load whichever version is newest by timestamp
  • C. A CI/validation job runs on the candidate version and only transitions its stage/alias to production if checks pass✓ Correct
  • D. Team members vote on the candidate in a chat channel, and whoever sees the result updates the production alias by hand
Explanation

A gated promotion runs automated validation on the candidate and only updates the production alias/stage when checks pass, making it a controlled MLOps release. Auto-promoting every version and reviewing later risks shipping bad models. Never promoting and loading the newest version skips validation completely. Chat votes followed by manual alias updates are not an automated, reliable gate.

4.

What is the primary purpose of the MLflow Model Registry?

  • A. To distribute hyperparameter tuning trials across cluster workers and automatically select the best-performing model for each run
  • B. To track cluster CPU, memory, and GPU utilization during training jobs and alert owners when resource consumption exceeds set thresholds
  • C. To act as the versioned storage layer for raw training datasets, snapshotting Delta tables alongside each experiment run for reproducibility
  • D. To manage the lifecycle of models — versioning, stage/alias transitions, and governance — as a central store of registered models✓ Correct
Explanation

The Model Registry centrally manages registered models: versions, aliases/stages (e.g., Champion/Challenger or Staging/Production), transitions, and governance. It does not distribute tuning trials or train models, it is not a storage layer for raw training data, and it does not monitor cluster resource utilization.

5.

A model was promoted to production but a regression is discovered. Using registry aliases, what is the fastest safe remediation?

  • A. Delete the entire registered model so consumers fall back to their cached copy of the last good version
  • B. Repoint the production alias back to the previous known-good model version (rollback)✓ Correct
  • C. Retrain the model from scratch on the original training data, then register it as a new version before acting
  • D. Restart the workspace and serving endpoint so the endpoint reloads the most stable cached model artifact
Explanation

Because deployment references an alias, repointing the production alias to the prior known-good version rolls back instantly without touching consumer code. Deleting the registered model is destructive and consumers have no cached fallback to rely on. Retraining from scratch first is slow during an incident. Restarting the workspace and endpoint just reloads whatever the alias points to, so the bad model is still served.

Test yourself across every Machine Learning Professional domain

The free Machine Learning Professional practice test mixes about 20 questions weighted like the exam. The full 58-question bank adds instant grading, flashcards and a timed mock exam.

Take the free Machine Learning Professional practice testPractice the full bank

Study this domain

ML Ops study guide (Machine Learning Professional) →

Other Machine Learning Professional domains

← All Databricks Machine Learning Professional study resources