Databricks ML Associate Practice Test

Use cost-sensitive learning or other imbalance-mitigation techniques during training is correct because Imbalance-mitigation techniques such as cost-sensitive learning address the minority-class problem more directly than just adding more majority examples. The cited source, Databricks Certified Machine Learning Associate - Exam Guide, supports this answer for the Model Development scenario rather than the adjacent distractors.

Databricks Model Serving endpoints support hosting multiple models or model versions behind a single endpoint. You can configure custom traffic routing percentages (e.g. 90/10 split) to safely run A/B tests or canary deployments.

MLflow supports logging and registering models from any client environment (including local machines) as long as the client is authenticated to the Databricks workspace. The developer simply logs the model in an active run and calls register_model using the three-level catalog namespace.

The mlflow.set_experiment(experiment_path) function is used to set the active experiment by specifying its path in the workspace. If the experiment at that path does not exist, MLflow creates it automatically.

The right approach is to use `.summary()` or Databricks data summaries because the exam guide lists those as expected ways to compute summary statistics on a Spark DataFrame. This is a data-exploration step, not a deployment step.

FeatureEngineeringClient is correct because FeatureEngineeringClient is the official fit because Databricks documents it as the Python client for creating and managing feature tables in Unity Catalog. The cited source, Feature tables, supports this answer for the Databricks Machine Learning Foundations scenario rather than the adjacent distractors.

In Databricks, Hyperopt can distribute hyperparameter tuning across worker nodes using the SparkTrials class. You pass an instance of SparkTrials (with a specified level of parallelism) to the fmin() function, replacing the standard single-node Trials class.

Delta Live Tables natively supports streaming inference. The recommended approach is to load the registered MLflow model as a Spark UDF (e.g. using mlflow.pyfunc.spark_udf()) and apply it directly within the DLT Python query definitions on the streaming DataFrame.

Model aliases are the best fit because they let the team refer to a role like champion or challenger instead of hard-coding version IDs everywhere. The exam guide explicitly includes promoting a challenger to a champion using aliases.

mlflow.search_runs() returns a pandas DataFrame containing all runs matching the experiment IDs. The order_by parameter accepts an array of strings (e.g., ['metrics.val_loss ASC']), allowing the runs to be sorted so that the run with the lowest validation loss appears first (accessed via .iloc[0]).