Databricks Machine Learning Professional · 44% practice allocation

Databricks Machine Learning Professional: Model Development Practice Questions

5 free sample questions from the Model Development 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.

A data scientist runs a hyperparameter sweep and wants each trial's metrics grouped under a single parent experiment run for organization. Which MLflow feature supports this?

  • A. Nested runs (a parent run with child runs per trial)✓ Correct
  • B. Lakehouse Monitoring profiles that group each trial's metrics by time window
  • C. Model Registry stages, giving each trial's model its own Staging version
  • D. A Delta Live Tables pipeline with one expectation per trial
Explanation

MLflow nested runs group child runs (one per trial) under a parent run, which keeps a sweep's results organized. Model Registry stages version and promote models; they do not group trial metrics. Lakehouse Monitoring profiles production tables and inference data, not tuning trials. Delta Live Tables is an ETL framework, and its expectations check data quality rather than organizing experiment runs.

2.

A team needs feature lookups automatically included when scoring, so callers only pass primary keys and the model retrieves features. Which Feature Store capability provides this?

  • A. Adding a Delta Live Tables expectation on the scoring table that fills in missing feature columns from the feature store
  • B. Having each downstream application join the feature tables itself and pass full feature vectors in every scoring request to the endpoint
  • C. Applying a cluster policy that attaches feature tables to every scoring cluster so models can read features by key
  • D. Logging the model with feature table lookups (a FeatureStore/feature-engineering packaged model) so lookups happen at scoring time✓ Correct
Explanation

When a model is packaged with feature-store lookups, the serving layer fetches the referenced features by key at scoring time, so clients only pass keys. Having each application join features itself brings back training/serving skew. Cluster policies govern compute and don't attach feature tables. DLT expectations are data-quality checks and don't fill in features.

3.

A hyperparameter search with Hyperopt should minimize validation loss. What must the objective function return for Hyperopt to optimize correctly?

  • A. A loss value (and status) that Hyperopt minimizes — return the metric so lower is better✓ Correct
  • B. The raw accuracy as a float, since Hyperopt maximizes whatever scalar the objective function returns
  • C. A dict containing only the sampled hyperparameters so the TPE algorithm can compare the configurations
  • D. The trained model object, which Hyperopt scores internally by calling predict on the validation data
Explanation

Hyperopt minimizes the loss returned by the objective, usually a dict with 'loss' and 'status'. So you return the validation loss such that lower is better, and you negate metrics like accuracy. Hyperopt does not score a returned model object itself. Returning raw accuracy is wrong because Hyperopt minimizes and does not maximize. Returning only the hyperparameters gives it no loss to optimize.

4.

A scikit-learn model is logged with MLflow. Which artifact allows the model to be loaded and served in different environments with the right library versions?

  • A. The logged MLflow model with its flavor and environment/conda (or requirements) specification✓ Correct
  • B. The notebook exported as HTML, which captures the pip install cells used during training
  • C. The cluster's Spark UI event log, which records the libraries installed on the cluster
  • D. The raw pickle file saved with joblib, which embeds the scikit-learn version needed to reload it anywhere
Explanation

An MLflow model stores the model in a flavor along with an environment specification (conda.yaml / requirements.txt). This lets it be reloaded and served with compatible dependencies. A bare joblib or pickle file has no environment metadata, so it can't reliably rebuild the right library versions. An exported notebook and the Spark UI event log are not deployable model artifacts.

5.

A data scientist wants to log a custom evaluation metric (e.g., a business-specific weighted error) alongside standard metrics for a run. How should they do this in MLflow?

  • A. Store it in a Python variable in the notebook, which MLflow captures at run end
  • B. Rename the logged model artifact file so its name includes the metric value
  • C. Custom metrics are not supported; only autologged metrics can be recorded
  • D. Compute the metric and log it with mlflow.log_metric()✓ Correct
Explanation

Custom metrics are logged with mlflow.log_metric() and appear next to autologged metrics in the run. MLflow fully supports custom metrics, so saying only autologged metrics are allowed is false. Putting the value in an artifact's file name does not track it as a metric. MLflow does not capture notebook variables automatically, so a value stored only in a variable is never saved to the run.

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

Model Development study guide (Machine Learning Professional) →

Other Machine Learning Professional domains

← All Databricks Machine Learning Professional study resources