Hard — Databricks-ML-Associate Practice Quiz

Question 1

A data scientist has created a feature table in Unity Catalog named prod.retail.customer_features with primary key customer_id. They build a FeatureLookup to join these features onto a labeled training DataFrame using FeatureEngineeringClient.create_training_set(). When they call training_set.load_df(), the resulting DataFrame contains null values for all feature columns even though the feature table is fully populated. What is the most likely cause?

Accepted Answer

The lookup_key specified in the FeatureLookup does not match the values or column present in the labeled DataFrame, so no rows join to the feature table's primary key.

Answer

The feature table must be published to an online store before create_training_set() can retrieve any values.

Answer

MLflow autologging was disabled, which prevents the Feature Engineering client from reading feature values.

Answer

The Databricks Runtime for ML version is too old to support Unity Catalog feature tables, silently returning nulls.

Question 2

A data scientist is building a training dataset using a Databricks Feature Store feature table that stores time-varying customer metrics (e.g., rolling 30-day spend) with a timestamp column. Each training label event also has its own event timestamp. When creating the training set with FeatureLookup, the scientist wants to ensure that each label row is joined only with feature values that were known at or before the label's event time, preventing data leakage from future feature updates. Which approach correctly achieves this?

Accepted Answer

Specify the timestamp_lookup_key in the FeatureLookup so the feature table is joined using a point-in-time lookup against the label's event timestamp.

Answer

Filter the feature table to keep only the most recent row per primary key before creating the training set.

Answer

Sort the training DataFrame by timestamp and use a standard inner join on the primary key only.

Answer

Enable online serving on the feature table so historical values are automatically time-aligned.

Question 3

A data scientist is using Hyperopt to tune a gradient boosting model. They want to search for the learning rate over a range from 0.001 to 0.1, and they know that this parameter's effect is multiplicative rather than additive—values like 0.001, 0.01, and 0.1 should each get comparable exploration. Which search space definition should they use for the learning rate?

Accepted Answer

hp.loguniform('learning_rate', np.log(0.001), np.log(0.1))

Answer

hp.uniform('learning_rate', 0.001, 0.1)

Answer

hp.quniform('learning_rate', 0.001, 0.1, 0.01)

Answer

hp.choice('learning_rate', [0.001, 0.1])

Question 4

A data scientist is tuning a Spark ML pipeline using CrossValidator with a 3-fold cross-validation and a parameter grid of 4 combinations. After fitting, they want to inspect the individual model trained for every fold and every parameter combination (not just the final best model refit on all data) to analyze fold-level variance. By default, what does the fitted CrossValidatorModel expose, and what must be changed to access every fold's model?

Accepted Answer

By default only the best model (refit on the full dataset) and average metrics are retained; setting collectSubModels=True before fitting retains all per-fold sub-models in subModels.

Answer

By default all sub-models are always retained in the bestModel attribute; no configuration change is needed.

Answer

By default only the average metrics are retained; setting parallelism greater than 1 causes all per-fold models to be stored.

Answer

By default the CrossValidatorModel stores only the last fold's model; enabling checkpointing retains the rest.

Question 5

A data scientist is training a gradient-boosted tree classifier on a 500 GB dataset stored in a Delta table that does not fit in the memory of a single worker. They want to tune 6 continuous and integer hyperparameters efficiently, exploring the search space adaptively rather than exhaustively, while keeping all training distributed across the cluster. Which approach best fits these requirements?

Accepted Answer

Use Hyperopt's fmin with the TPE algorithm, where the objective function fits a Spark ML pipeline on the full distributed dataset, running trials sequentially

Answer

Use Spark ML's estimator with CrossValidator and a ParamGridBuilder specifying all combinations of the 6 hyperparameters

Answer

Use Hyperopt's fmin with the TPE algorithm and SparkTrials, wrapping a scikit-learn model trained on a pandas DataFrame loaded from the Delta table

Answer

Use Spark ML's estimator with TrainValidationSplit and a ParamGridBuilder covering the full grid of the 6 hyperparameters

Question 6

A data engineer needs to score a 200-million-row Spark DataFrame using a registered scikit-learn model in Unity Catalog. Before prediction, each row must be enriched by calling a preprocessing function that depends on grouping records by 'store_id' to compute a rolling group-level statistic that the model expects as a feature. The engineer wants the enrichment and prediction to run in a single distributed pass. Which approach best satisfies these requirements?

Accepted Answer

Group the DataFrame by 'store_id' and use applyInPandas with a function that computes the group statistic and calls the loaded model's predict method

Answer

Load the model with mlflow.pyfunc.spark_udf and apply it directly to the DataFrame with a withColumn call

Answer

Collect the DataFrame to the driver, compute the group statistics in pandas, then broadcast the result back for scoring

Answer

Deploy the model to a Databricks Model Serving endpoint and send each row as a REST request from a Spark map operation

Question 7

A data engineer needs to score 500 million records in a nightly Spark batch job using an MLflow model registered in Unity Catalog. Each record must first be grouped by 'store_id' so that a store-specific normalization step (implemented as custom Python logic requiring the full group in memory) is applied before the model's predictions are generated. Which approach correctly handles the per-group preprocessing and inference in a distributed manner?

Accepted Answer

Group the DataFrame by 'store_id' and use applyInPandas with a function that performs the per-group normalization and then calls the loaded model to predict, returning a pandas DataFrame with a defined output schema.

Answer

Use mlflow.pyfunc.spark_udf to load the model and call it directly on the DataFrame, relying on the UDF to group and normalize the data automatically.

Answer

Collect the entire DataFrame to the driver with .toPandas(), apply normalization per group in a loop, and score with the model in memory.

Answer

Register the model as a spark_udf and wrap it in a Structured Streaming query so each store's records are processed as a separate micro-batch.

Question 8

A data scientist is preparing a customer churn dataset that contains a high-cardinality categorical feature, 'zip_code' (over 30,000 unique values). They decide to use target (mean) encoding, replacing each zip code with the average churn rate observed for that zip code. To avoid inflating validation performance, which approach should they use when computing the encoding?

Accepted Answer

Compute the mean churn rate per zip code using only the training split, then map those values onto both the training and test splits (using a global mean for unseen categories).

Answer

Compute the mean churn rate per zip code using the full dataset before splitting, so every row gets a stable, consistent encoding.

Answer

Compute the mean churn rate per zip code separately within the training split and within the test split, so each split reflects its own churn behavior.

Answer

One-hot encode the zip code feature instead, since target encoding always introduces leakage regardless of how it is computed.