Databricks-ML-Assoc Practice Test 16 — 15 Questions

Question 1

A data scientist is evaluating feature importance for a tree-based model trained on Databricks. They want to understand which features contribute most to the model's predictions. Which TWO methods are appropriate for extracting feature importance from a scikit-learn Random Forest model? (Choose two.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Access the feature_importances_ attribute of the fitted model.. The feature_importances_ attribute of a fitted Random Forest provides a quick, impurity-based measure of global feature importance. SHAP values offer a more detailed, model-agnostic approach that attributes each feature's contribution to individual predictions, which can be aggregated for global importance. Both are valid methods for understanding feature influence. The other options do not yield feature importance: predict_proba returns probabilities, decision function coefficients are for linear models, and get_params returns hyperparameters.

Answer

Examine the coefficients of the model's decision function.

Answer

Use the model's get_params method to retrieve feature importance.

Answer

Use the model's predict_proba method to derive feature importance.

Question 2

A team has a Databricks Model Serving endpoint configured with scale-to-zero enabled and min_instances set to 0. During a load test, they observe that the first request after an idle period takes roughly 40 seconds while subsequent requests complete in under 200 milliseconds. They need to eliminate this cold-start latency for a customer-facing application without over-provisioning. Which configuration change best addresses the requirement?

Accepted Answer

Set min_instances to 1 so at least one instance stays warm.. Cold-start latency occurs because a scaled-to-zero endpoint must provision a container and load the model before serving the first request. Keeping one replica always available by setting min_instances to 1 removes that provisioning and loading delay while still allowing the endpoint to scale up under load. Increasing max_instances, changing workload size, or enabling routing features does not prevent the endpoint from scaling to zero during idle periods.

Answer

Increase max_instances so more replicas can absorb the burst.

Answer

Increase the workload size from Small to Large.

Answer

Enable route optimization in the endpoint configuration.

Question 3

A team is preparing to deploy an MLflow model to a Databricks Model Serving endpoint and wants to diagnose why requests are failing before contacting support. Which two actions allow them to inspect the endpoint's behavior and errors? (Choose two.)

Accepted Answer

Query the endpoint's event logs and metrics in the Databricks workspace to review build and update events.. Inference tables capture the actual request and response payloads, exposing malformed inputs and model-side errors, while endpoint event logs and metrics show deployment, scaling, and update events. Together they cover both request-level and infrastructure-level diagnosis. Attaching clusters, SSH access, and redeploying do not provide supported visibility into a managed serving endpoint's runtime.

Answer

Re-register the model with a new signature and redeploy the endpoint to force the error to surface in the UI.

Answer

SSH into the serving container and read the application log files directly from the filesystem.

Answer

Attach a notebook to the serving cluster and run the model's predict method interactively to reproduce the error.

Question 4

A data scientist is using Databricks Feature Store to build a training set for a model that predicts customer churn. The feature table contains a column `customer_id` and several features, and the label is stored in a separate Delta table. The data scientist wants to ensure that the exact same feature values used during training are available at inference time. Which approach correctly uses Databricks Feature Store to create the training set?

Accepted Answer

Use `fs.create_training_set()` with the feature table and label DataFrame, specifying `customer_id` as the lookup key.. The `create_training_set` method is specifically designed to join feature tables with labels using a lookup key, preserving feature lineage and ensuring that the same feature computations are used during training and inference. This prevents training-serving skew and simplifies deployment.

Answer

Perform a manual join between the feature table and the label table using Spark, then log the resulting DataFrame as an MLflow artifact.

Answer

Export the feature table to a CSV file and merge it with the label data using pandas, then log the model with MLflow.

Answer

Use `fs.log_model()` with the feature table name and specify the label column; the method automatically creates the training set.

Question 5

A data scientist is using MLflow to log a model on Databricks. They want to ensure that the model can be loaded and used for inference in a different environment. Which two of the following are necessary components that must be included when logging the model to guarantee portability? (Choose two.)

Accepted Answer

The conda environment file (conda.yaml) that lists all dependencies.. For a model to be portable across environments, it must include a signature to define input/output schemas and a conda environment file to capture dependencies. These ensure that the model can be correctly loaded and executed elsewhere. Other elements like training data, run ID, or input examples are not required for the model to function.

Answer

The MLflow run ID where the model was logged.

Answer

The model's input example, which provides a sample input for testing.

Answer

The training dataset used to fit the model.

Question 6

A data scientist finishes a training notebook and wants to capture the source code revision and the Git repository URL on the MLflow run so reviewers can reproduce the exact code state. The repository is already connected to the Databricks workspace through Git integration. Which mechanism records this information automatically?

Accepted Answer

MLflow Git-based source tracking, which stores the repository URL and commit hash as run tags when the notebook runs from a Git folder.. Databricks integrates Git folders with MLflow so that runs executed from a connected repository automatically receive source tags including the repository URL and commit hash. This gives reviewers a precise revision to check out for reproduction. Manual tagging is not automatic, autologging targets framework artifacts rather than Git metadata, and model registration does not populate run source tags.

Answer

Enabling autologging, which always injects Git metadata regardless of where the notebook is executed.

Answer

Manually calling mlflow.set_tag() with a hand-copied commit hash after each training run.

Answer

Registering the model in Unity Catalog, which copies the workspace Git configuration into model version metadata.

Question 7

A machine learning engineer is using MLflow to log a model built with XGBoost. They call mlflow.xgboost.log_model(xgb_model, 'model') and then attempt to load the model in a different environment using mlflow.pyfunc.load_model('runs:/<run_id>/model'). The load fails with an error about missing dependencies. Which action should they take to ensure the model can be loaded in the new environment?

Accepted Answer

They should include the XGBoost library in the conda environment when logging the model, either by passing a custom conda_env or by ensuring the library is installed in the current environment so MLflow can infer it.. When logging a model with MLflow, dependencies are recorded in a conda environment file. If a required library like XGBoost is missing from that file, loading in a new environment fails. Ensuring the library is installed during logging allows MLflow to capture it, or you can specify a custom conda environment. This guarantees the model can be reproduced.

Answer

They should log the model with the registered_model_name parameter to ensure dependencies are captured.

Answer

They should save the model in the ONNX format instead, because ONNX models have no dependencies.

Answer

They should use mlflow.xgboost.load_model() instead of mlflow.pyfunc.load_model(), because the latter does not support XGBoost models.

Question 8

An ML engineer is configuring a Databricks Job to retrain a model daily. The job must first run a notebook that creates a feature table, then run a notebook that trains and registers the model. The engineer wants to ensure the training notebook only runs if the feature table creation succeeds. Which Databricks Workflows feature should they use to define this dependency?

Accepted Answer

Set the training task's depends_on field to the feature table task.. In Databricks Workflows, task dependencies are defined using the depends_on field, which lists upstream tasks that must complete successfully before the task runs. This creates a DAG and ensures the training task executes only after the feature table task succeeds. Other options either do not establish the dependency or rely on additional configuration that still requires depends_on.

Answer

Set the training task's run_if condition to ALL_SUCCESS, which automatically includes the feature table task.

Answer

Use a conditional task with a run_if condition based on the feature table task's output.

Answer

Configure the feature table task to call the training notebook at the end of its run.

Question 9

A machine learning engineer has registered a model in the Databricks Model Registry and wants to expose it as a REST API with automatic scaling and no server management. The model's Python dependencies are captured in a conda environment file logged with the run. Which Databricks capability should the engineer use to serve this model with minimal operational overhead?

Accepted Answer

Databricks Model Serving with a model version URI from the Model Registry. Databricks Model Serving is the managed capability that takes a registered model version URI, reconstructs the environment from logged dependencies, and publishes a scalable REST endpoint without server administration. The other choices describe batch, interactive, or SQL compute that cannot deliver managed low-latency REST inference for a registered MLflow model.

Answer

A Databricks job that runs the model on a schedule and writes predictions to a Delta table

Answer

A SQL warehouse configured with the model's conda environment installed on all nodes

Answer

An all-purpose cluster running an MLflow model server process started manually

Question 10

A data science team is deploying a model to a Databricks Model Serving endpoint. They want to enable inference logging to capture the input data and predictions for monitoring and debugging. Which TWO configurations are required to enable inference logging for a serving endpoint? (Choose two.)

Accepted Answer

Specify a Delta table to store the inference logs.. To enable inference logging for a Model Serving endpoint, you must explicitly enable it in the endpoint configuration and specify a Delta table to store the logs. The endpoint's identity needs write access to that table. Other factors like workload size or model signature do not affect logging. This allows capturing request and response data for monitoring and debugging.

Answer

Set the endpoint's workload size to at least 'Medium'.

Answer

Ensure the model signature includes a field for logging.

Answer

Configure the model to output logs in JSON format.

Question 11

A data scientist trains a model with MLflow on Databricks and logs it using mlflow.sklearn.log_model with a registered_model_name. A downstream batch job loads the model by stage using models:/<name>/Staging. Weeks later, a colleague promotes a new version to Staging and the batch job's predictions change without any code deployment. Which change best prevents unintended downstream consumption while keeping promotion workflows intact?

Accepted Answer

Load the model by explicit version URI instead of by stage.. Stage-based URIs are mutable pointers: whatever version is transitioned into a stage becomes the artifact that stage references. Pinning a consumer to an explicit model version makes its dependency immutable, so later promotions affect only consumers that intentionally follow stages. This preserves Registry-based promotion while eliminating silent behaviour changes in the batch job.

Answer

Archive the previous versions in the Model Registry before each promotion.

Answer

Add a model signature and re-log the model with registered_model_name.

Answer

Set the batch job to load the model from the run's artifact URI in the tracking server.

Question 12

A data scientist is training a gradient boosting model using Spark MLlib on a large dataset in Databricks. They notice that the model's performance on a validation set is significantly worse than on the training set, and they suspect overfitting. They want to use MLflow to track hyperparameters and metrics to diagnose the issue. Which combination of MLflow logging practices will best help them identify overfitting across multiple runs?

Accepted Answer

Log training and validation metrics separately for each run, and log hyperparameters such as maxDepth and minInstancesPerNode.. To diagnose overfitting, it is essential to compare training and validation performance across runs. Logging both metrics separately, along with relevant hyperparameters that control model complexity, allows the data scientist to visualize the gap and tune accordingly. MLflow's tracking and comparison features then make it straightforward to identify runs where validation performance lags.

Answer

Log only the training metric and the model artifact, and rely on the MLflow UI to automatically compute validation metrics.

Answer

Log hyperparameters and the final model, but log metrics only at the end of training to reduce clutter.

Answer

Use MLflow autologging for Spark MLlib, which automatically logs training and validation metrics and hyperparameters.

Question 13

A data scientist is tuning a scikit-learn GradientBoostingClassifier on Databricks. They use Hyperopt with the fmin function and the SparkTrials backend, but they notice that the best model returned by fmin is not identical to the model they get when they retrain with the same hyperparameters. They also observe that the logged metrics from each trial vary slightly even when the same hyperparameters are used. What is the most likely cause?

Accepted Answer

The random seed for the classifier is not fixed, so each trial and the final retrain produce different random splits and initialization, leading to slight variations.. The mismatch and metric variation occur because the model training is non-deterministic. GradientBoostingClassifier uses random processes for feature subsampling and sample selection when subsample is less than 1.0. Without setting a fixed random_state, each run produces a slightly different model, even with identical hyperparameters. This explains both the differing trial metrics and the final model discrepancy.

Answer

SparkTrials runs each trial on a different worker node, and the data is shuffled differently on each node, causing divergent results.

Answer

Hyperopt's fmin function does not support scikit-learn models; it only works with MLflow models, so the returned model is a placeholder.

Answer

SparkTrials caches the training data on each worker, and the cache is not invalidated between trials, causing stale data to be used.

Question 14

An ML engineer is using Databricks Jobs to orchestrate a multi-step ML pipeline. The pipeline includes a task that trains a model and logs it to MLflow, followed by a task that registers the model in the Model Registry. The engineer wants to ensure that the model is only registered if its accuracy exceeds a threshold. Which approach best implements this conditional logic?

Accepted Answer

In the training task, evaluate the model and if accuracy exceeds the threshold, call the MLflow Model Registry API to register the model.. The most straightforward way to conditionally register a model based on accuracy is to perform the evaluation within the training task and call the Model Registry API only if the threshold is met. This keeps the logic in one place and avoids unnecessary task dependencies. Other approaches either do not enforce the metric condition or require additional complexity.

Answer

Configure the Model Registry to automatically register models only if they meet a performance threshold set in the model's metadata.

Answer

Use a Databricks Job with a conditional task that checks a metric value stored in a Delta table and then triggers registration.

Answer

Use a Databricks Job with multiple tasks and define a dependency that runs the registration task only if the training task succeeds.

Question 15

A machine learning engineer needs to capture every request and response payload sent to a Databricks Model Serving endpoint so that the team can later join predictions with ground-truth labels for monitoring. Which Databricks feature should they enable on the endpoint?

Accepted Answer

Inference tables. Inference tables are the Databricks Model Serving feature that persists request and response payloads to a Delta table. Enabling them allows teams to join predictions with ground-truth labels over time, monitor drift, and debug production behavior. Signature enforcement, MLflow artifact logging, and audit logs serve different purposes and do not capture payloads.

Answer

Model signature enforcement

Answer

Endpoint access logs in the workspace audit log

Answer

MLflow model logging to the run's artifact store