Databricks Machine Learning
Free Databricks Certified Machine Learning Associate practice — 6 questions on Databricks Machine Learning, with explanations. No sign-up. Full 12-question mixed test →
A data scientist is using Feature Engineering in Unity Catalog to build a training set. Each observation in the label table has an event timestamp, and the data scientist must ensure that only feature values available AS OF that timestamp are joined in, to prevent label leakage. Which FeatureLookup parameter must be specified to achieve this point-in-time correct join?
timestamp_lookup_key tells create_training_set to perform a point-in-time join, matching each observation only with feature values that existed at or before its timestamp, which prevents future data (label leakage) from entering the training set.
A team has registered a model to Unity Catalog and wants to designate the newest version as ready for production traffic using the current, non-deprecated Unity Catalog model versioning approach. What should they do?
Models registered in Unity Catalog use aliases (e.g., 'Champion', 'Challenger') instead of the legacy stage system (Staging/Production/Archived). Assigning an alias is the supported way to mark a version as the one currently used for a purpose like production serving.
A data scientist runs an AutoML classification experiment on a customer churn dataset and wants to modify the preprocessing steps (e.g., imputation strategy) used by the single best-performing trial, then retrain it. What is the correct approach?
AutoML produces a fully editable, glass-box source code notebook for each trial (including the best trial). Data scientists are expected to open this notebook, modify preprocessing/feature engineering or hyperparameters directly in the code, and rerun it to produce a customized model — this is the documented intended workflow.
A data scientist uses Hyperopt with SparkTrials to run distributed hyperparameter tuning across a cluster and wants each individual hyperparameter combination logged as its own MLflow run, grouped underneath a single parent run for the overall tuning job. Which MLflow capability enables this run hierarchy?
MLflow nested runs, started with mlflow.start_run(nested=True) inside an active parent run, let each Hyperopt/SparkTrials trial log its own params, metrics, and artifacts as a child run visible in the tracking UI under the parent run — exactly the hierarchy Hyperopt integration with MLflow uses.
A team wants to train a deep learning model with distributed GPU acceleration on a Databricks cluster, using preinstalled CUDA/cuDNN libraries and GPU-optimized versions of frameworks like TensorFlow and PyTorch. Which cluster configuration should they select?
Databricks Runtime ML GPU editions ship with CUDA, cuDNN, and GPU-compatible builds of deep learning frameworks preinstalled and configured; combined with a GPU instance type, this is the only listed option that provides working GPU-accelerated training out of the box.
A company registers ML models in Unity Catalog under a three-level namespace (catalog.schema.model) and wants to restrict read/inference access to a specific registered model so that only the fraud-detection team can use it, without granting broader workspace-wide access. Which mechanism enforces this?
Unity Catalog governs models the same way it governs tables and other securable objects: fine-grained permissions are granted via SQL GRANT statements at the catalog, schema, or individual model level, allowing precise control such as granting EXECUTE/USE privileges only to the fraud-detection team's group.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →