Databricks-ML-Pro Model Deployment Practice Test 2 — 15 Questions

Question 1

An ML engineer has registered a model in Unity Catalog with the name `prod.ml_models.fraud_detector`. The model has version 3 tagged as `champion` via the alias `champion`. The engineer now needs to update the serving endpoint `fraud-endpoint` to serve this specific model version using the alias. Which configuration correctly sets the served entity?

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

Accepted Answer

Set `entity_name` to `prod.ml_models.fraud_detector` and `entity_version` to `champion` in the served entities configuration.. To serve a model by alias in Databricks Model Serving, the served entity must specify the full model name in `entity_name` and the alias name in `entity_version`. This allows the endpoint to dynamically resolve the alias to the correct version and update automatically when the alias is reassigned. Other options either use a static version, incorrect syntax, or a generic tag that does not reflect the alias.

Answer

Set `entity_name` to `prod.ml_models.fraud_detector` and `entity_version` to `latest` in the served entities configuration.

Answer

Set `entity_name` to `prod.ml_models.fraud_detector` and `entity_version` to `3` in the served entities configuration.

Answer

Set `entity_name` to `prod.ml_models.fraud_detector@champion` and omit `entity_version`.

Question 2

A team is deploying a model to Databricks Model Serving. They want to ensure the endpoint can handle sudden spikes in traffic without dropping requests and that the model's dependencies are correctly resolved. Which TWO actions should they take? (Choose two.)

Accepted Answer

Configure the endpoint with a minimum number of replicas greater than zero to avoid cold starts during traffic spikes.. To handle traffic spikes, the endpoint must have a minimum replica count to avoid cold starts and allow immediate scaling. Correct dependency resolution requires that the model's conda environment includes all necessary packages. These two actions directly address the stated requirements.

Answer

Deploy the model to a single-node cluster and enable autoscaling on that cluster.

Answer

Log the model with an MLflow signature that includes input and output examples to validate the schema.

Answer

Use a Large workload size for the endpoint to guarantee the highest possible throughput.

Question 3

An ML engineer is deploying a model to Databricks Model Serving and wants to ensure that the endpoint automatically scales down to zero when there is no traffic, thereby reducing cost. Which configuration should be applied to the endpoint?

Accepted Answer

Set the endpoint's workload size to 'Small' and enable autoscaling with a minimum of 1 replica.. Databricks Model Serving endpoints always maintain at least one replica; they cannot scale to zero. To reduce cost during idle periods, you can select a smaller workload size and enable autoscaling, which will scale down to one replica when there is no traffic. This balances cost and availability, ensuring the endpoint remains ready to serve requests without cold starts.

Answer

Enable scale-to-zero in the endpoint's autoscaling configuration.

Answer

Configure the endpoint to use a serverless SQL warehouse with auto-stop.

Answer

Set the min_replicas to 0 and max_replicas to a positive integer.

Question 4

A team is deploying a PyTorch model to Databricks Model Serving. The model requires a custom Python package that is not available in the default serving environment. The team has uploaded the package as a wheel file to Unity Catalog volume at `/Volumes/prod/ml_artifacts/wheels/custom_pkg-0.1.0-py3-none-any.whl`. They need to include this dependency when logging the model. Which approach correctly ensures the wheel is installed in the serving environment?

Accepted Answer

Use `mlflow.pyfunc.log_model` with the `pip_requirements` parameter set to a list containing the wheel file path.. To include a custom wheel dependency for a model deployed to Databricks Model Serving, you must log the model with the wheel specified in `pip_requirements`. This ensures the serving environment installs the package. Using `conda_env` is incorrect because it expects a Conda environment specification, not a wheel path. Init scripts are not supported in Model Serving, and placing the wheel without referencing it in `pip_requirements` will not cause installation.

Answer

Upload the wheel to DBFS and add an init script to the serving endpoint that runs `pip install` on the wheel.

Answer

Place the wheel file in the same directory as the model artifacts and reference it in the model's `requirements.txt` using a relative path.

Answer

Add the wheel file path to the `conda_env` parameter when calling `mlflow.pyfunc.log_model`.

Question 5

A team is using Databricks Model Serving to deploy a model that must handle sensitive data. They need to ensure that all inference requests and responses are logged for auditing purposes without exposing the data to unauthorized users. Which TWO configurations should they implement? (Choose two.)

Accepted Answer

Enable inference logging on the serving endpoint to capture request and response payloads.. To audit inference data, enable inference logging on the serving endpoint, which records request and response payloads to a Delta table. Then, use a Databricks SQL warehouse with access controls to ensure only authorized auditors can query these logs. This combination provides both the necessary data capture and the security to prevent unauthorized access.

Answer

Configure the serving endpoint to use a private network endpoint to isolate traffic.

Answer

Enable model monitoring to automatically detect data drift in the inference requests.

Answer

Use Unity Catalog to store the model artifacts and set permissions on the model version.

Question 6

An ML engineer is deploying a model to Databricks Model Serving. The model's MLflow signature includes a pandas DataFrame input with 50 float features. To reduce inference latency, the engineer wants to enable request batching. Which configuration is required?

Accepted Answer

Enable request batching by setting the 'batch_size' in the served model's configuration when creating or updating the endpoint.. Databricks Model Serving allows configuring request batching at the served model level. By specifying a batch size in the endpoint configuration, the serving infrastructure groups individual inference requests into batches before calling the model's predict function. This is the documented way to enable batching and reduce latency under high request volume.

Answer

Configure the served model with a 'Small' workload size and enable scale-to-zero.

Answer

Add a 'batch_size' parameter to the model's MLflow signature.

Answer

Set the environment variable MLFLOW_ENABLE_BATCHING to true in the model's conda.yaml.

Question 7

A team is deploying a model to Databricks Model Serving and needs to ensure that the endpoint can handle sudden spikes in traffic. They configure autoscaling but notice that during spikes, the latency increases significantly before the endpoint scales up. What should they do to improve responsiveness?

Accepted Answer

Enable scale-to-zero and set a minimum number of instances greater than zero.. To handle sudden traffic spikes without latency degradation, the endpoint should have a minimum number of instances greater than zero. This ensures that there are always warm workers available to absorb the initial burst while autoscaling adds more instances. Disabling scale-to-zero by setting a minimum instance count is the recommended practice for latency-sensitive applications.

Answer

Use a larger workload size to provide more resources per instance.

Answer

Increase the maximum concurrency setting for the endpoint.

Answer

Configure a longer timeout for the endpoint.

Question 8

An ML engineer has registered a model in Unity Catalog with the name `prod.ml_models.fraud_detector`. The model has version 3 tagged as `champion` via the alias `champion`. The engineer now needs to update the serving endpoint `fraud-endpoint` to serve this specific model version using the alias. Which configuration correctly sets the served entity?

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

Accepted Answer

Use the Databricks REST API endpoint `/api/2.0/serving-endpoints` with a JSON body that includes the model name and version.. Databricks Model Serving endpoints are created and managed via the REST API, specifically the `/api/2.0/serving-endpoints` endpoint. The request body must include the model name in Unity Catalog format and the version to serve. This method is fully supported and provides programmatic control. The other options reference non-existent CLI commands, incorrect MLflow deployment APIs, or unsupported environment variables.

Answer

Use the MLflow Model Registry API to transition the model to stage 'Production' and set the environment variable `DATABRICKS_SERVING_ENABLED=true`.

Answer

Use the Databricks CLI command `databricks model-serving create-endpoint` with flags `--model-name` and `--model-version`.

Answer

Use `mlflow.deployments.create_deployment()` with the target URI `databricks` and specify the model name and version in the config.

Question 9

A fraud detection model is registered in Unity Catalog as `prod.ml.fraud_model`. The model's training script uses scikit-learn version 1.3.2. An ML engineer creates a Model Serving endpoint and specifies the model version. During the first inference request, the endpoint fails with a `ModuleNotFoundError` for `scikit-learn==1.3.2`. The engineer confirms the model was logged with `mlflow.sklearn.log_model`. What is the most likely cause of this failure?

Accepted Answer

The model was logged with a conda environment that pins scikit-learn to a version incompatible with the serving environment's Python version, causing a silent fallback to an older version.. The failure occurs because the logged conda environment specifies a scikit-learn version that cannot be installed in the serving container's Python environment. Databricks Model Serving attempts to recreate the environment, but if the Python version differs, the exact package version may be unavailable. This results in a ModuleNotFoundError for the pinned version, highlighting the importance of aligning training and serving Python versions.

Answer

The endpoint's workload size does not have enough memory to load the scikit-learn library, causing the import to fail.

Answer

The model's conda environment was not logged with the model artifact, so Model Serving cannot resolve the exact scikit-learn version.

Answer

The model was logged without specifying the `pip_requirements` parameter, so Databricks Model Serving cannot install the required scikit-learn version.

Question 10

An ML engineer has registered a scikit-learn model in the Unity Catalog Model Registry as `main.default.sales_forecast`. The model was logged with an MLflow signature and a conda environment. The engineer wants to deploy this model to Databricks Model Serving for real-time inference. Which step must be completed before creating the serving endpoint?

Accepted Answer

Assign the model version to the 'Champion' alias.. When deploying a model from the Unity Catalog Model Registry to Databricks Model Serving, the model version must be referenced by an alias, such as 'Champion'. This alias serves as a stable pointer for the serving endpoint to resolve the correct model version. Without it, the endpoint cannot be created. The other steps are either unsupported or unnecessary for standard scikit-learn models.

Answer

Convert the model to a TensorFlow SavedModel format.

Answer

Register the model in the Workspace Model Registry instead.

Answer

Manually create a REST API proxy for the model.

Question 11

An ML engineer is deploying a model to Databricks Model Serving that requires a specific version of a Python library that conflicts with the version in the default Databricks Runtime. They have logged the model with a conda environment that specifies the correct version. However, during deployment, the endpoint fails to start, and logs show a dependency conflict. What is the most likely cause?

Accepted Answer

The model serving endpoint uses a base environment that overrides the conda dependencies.. The most likely cause is that the base environment of the serving endpoint includes a version of the library that conflicts with the one specified in the conda environment. Databricks Model Serving uses a predefined base image, and if the conda environment requests an incompatible version, the endpoint may fail to start. To resolve, the engineer should ensure the conda environment aligns with the supported runtime or use a custom runtime that includes the required version.

Answer

The serving endpoint does not support custom Python libraries.

Answer

The model was logged with a requirements.txt file instead of a conda environment.

Answer

The conda environment was not correctly logged with the model.

Question 12

An ML engineer is configuring a Databricks Model Serving endpoint to serve a model that requires GPU acceleration. The endpoint must automatically scale based on traffic. Which TWO configurations are required to enable GPU serving and autoscaling? (Choose two.)

Accepted Answer

Set `min_provisioned_concurrency` to 0 and `max_provisioned_concurrency` to a value greater than 0.. To enable GPU serving, set `workload_type` to a GPU-enabled size such as `GPU_MEDIUM`. To enable autoscaling, set `min_provisioned_concurrency` to 0 and `max_provisioned_concurrency` to a value greater than 0, allowing the endpoint to scale based on traffic. Other options either use incorrect parameter names, do not enable autoscaling, or describe unsupported features.

Answer

Configure a schedule-based scaling policy using cron expressions in the endpoint settings.

Answer

Set the `workload_size` to `GPU_Medium` in the served entities configuration.

Answer

Enable autoscaling by setting `scale_to_zero_enabled` to `true` and specifying `min_provisioned_concurrency` and `max_provisioned_concurrency`.

Question 13

A team is deploying a model to Databricks Model Serving that requires a custom Python library. The library is not available in the default environment. They need to ensure the endpoint can use this library. Which TWO methods can they use to include the custom library? (Choose two.)

Accepted Answer

Specify the library as a PyPI dependency in the model's conda environment.. Databricks Model Serving allows custom Python libraries to be included either by specifying them as PyPI dependencies in the conda environment or by packaging them as wheel files and including them in the model's artifacts with a reference in the conda environment. Both methods ensure the library is available at deployment time. Other options are either unsupported or impractical.

Answer

Install the library at runtime by including a pip install command in the model's predict method.

Answer

Use a custom container image that includes the library, and specify the image when creating the endpoint.

Answer

Upload the library as a wheel file to Unity Catalog and reference it in the model's environment.

Question 14

An ML engineer has registered a model in Unity Catalog with the name `prod.ml_models.fraud_detector`. The model has version 3 tagged as `champion` via the alias `champion`. The engineer now needs to update the serving endpoint `fraud-endpoint` to serve this specific model version using the alias. Which configuration correctly sets the served entity?

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

Accepted Answer

Create a serving endpoint that references the model by its alias, such as `models:/prod.ml.forecast_model@champion`, and update the alias to point to newly approved versions.. In Unity Catalog, aliases provide a mutable named reference to a model version. By configuring the serving endpoint to use an alias, such as `@champion`, you can update the alias to point to a new version when it is approved. This allows the endpoint to serve the latest approved model without manual endpoint reconfiguration, meeting the requirement for automatic updates based on approval status.

Answer

Create a serving endpoint that references the model by its tag, such as `models:/prod.ml.forecast_model@approved`, and rely on Unity Catalog to automatically update the endpoint when a new version receives the tag.

Answer

Create a serving endpoint that points to the model version 3 directly, using the model URI `models:/prod.ml.forecast_model/3`.

Answer

Create a serving endpoint that uses the model's stage, such as `models:/prod.ml.forecast_model/Production`, and transition new versions to the Production stage.

Question 15

An ML engineer is deploying a model to Databricks Model Serving and needs to enable authentication and authorization for the endpoint. Which TWO methods can be used to secure access to the endpoint? (Choose two.)

Accepted Answer

Use OAuth tokens for authentication.. Databricks Model Serving endpoints can be secured using personal access tokens (PATs) or OAuth tokens. PATs are simple to use but are tied to a user or service principal and can be revoked. OAuth tokens are obtained via the Databricks OAuth flow and are recommended for production because they are short-lived and can be scoped. Both methods provide authentication and authorization based on the token's permissions.

Answer

Use SSH keys to authenticate to the endpoint.

Answer

Use IP allowlists to restrict access to the endpoint.

Answer

Use Unity Catalog privileges to control access to the endpoint.