Databricks-ML-Pro Model Deployment — 15 Questions

Question 1

A machine learning engineer needs to deploy a custom MLflow model that requires a non-standard Python package not available in the default Databricks ML runtime. Which approach ensures the custom dependency is correctly packaged and loaded during model inference?

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

Accepted Answer

Pass the custom library paths via the extra_pip_requirements parameter when logging the model to ensure dependencies are captured in requirements.txt.. Including extra artifacts like source files and specifying conda or pip requirements within the MLflow pyfunc signature ensures the serving environment installs necessary dependencies. This matters because model serving containers lack arbitrary user libraries unless explicitly declared through MLflow configuration artifacts during the logging phase.

Answer

Manually install the package on the driver node using pip install immediately before calling mlflow.spark.log_model.

Answer

Store the custom python wheel file in DBFS and reference it using a global environment variable inside the serving cluster configuration.

Answer

Modify the cluster spark configuration to append the local package directory to the PYTHONPATH across all worker nodes.

Question 2

An ML engineer needs to schedule daily batch inference for a fraud detection model processing millions of transactions stored in a Delta table. Which deployment pattern provides the most scalable and fault-tolerant architecture within Databricks?

Accepted Answer

Use MLflow pyfunc pandas UDFs loaded via mlflow.pyfunc.spark_udf to score the Delta table distributed across Spark worker nodes.. Using a Databricks Job orchestrated with a Spark-based batch inference notebook leverages distributed computing to process massive datasets efficiently. This pattern is essential because single-node execution would fail on large datasets, whereas distributed Spark processing handles partition-level failures and utilizes cluster resources optimally.

Answer

Write a custom Python script that downloads the entire Delta table into a local pandas DataFrame on the driver node and applies the model.

Answer

Deploy a real-time Model Serving endpoint and send individual HTTP POST requests for every single transaction row using a loop.

Answer

Manually execute an interactive notebook cell each morning and export the scored results to local CSV files.

Question 3

An ML engineer needs to deploy a custom PyTorch model to Databricks Model Serving with a custom preprocessing step that requires external Python packages. Which approach ensures the dependencies and custom logic execute correctly in the serverless inference container?

Accepted Answer

Implement a custom MLflow Pyfunc model wrapper that encapsulates the preprocessing logic and explicitly defines the external conda and pip dependencies in the model flavor configuration.. Deploying models with custom logic requires packaging dependencies using MLflow's model signature and conda environment configurations, or utilizing custom pyfunc implementations with explicit artifact dependencies. This ensures the serverless runtime correctly builds the execution environment and executes the preprocessing transformations without missing library errors during production inference requests.

Answer

Register the PyTorch model without a signature and rely on the serving endpoint to automatically infer input schemas and install required pip packages dynamically at runtime.

Answer

Upload the raw PyTorch state dictionary to DBFS and write a custom scoring script that downloads the weights and initializes the model inside a shared notebook.

Answer

Store the PyTorch weights in Unity Catalog volumes and configure the cluster hosting the model to install the required packages cluster-wide upon every manual restart.

Question 4

A data engineering team is deploying a model using Databricks Model Serving. They need to ensure that the endpoint can handle sudden bursts of traffic while minimizing costs during periods of inactivity. Which configuration strategy for the 'workload_type' and 'scale_to_zero' parameters best achieves this balance?

Accepted Answer

Set the min_provisioned_throughput to 0 and max_provisioned_throughput to a high value.. Databricks Model Serving allows for automatic scaling based on the number of incoming requests. Enabling scale-to-zero allows the infrastructure to shut down completely when no traffic is detected, which is ideal for cost-saving on non-critical or intermittent workloads. Selecting an appropriate workload size ensures that when the model is active, it has sufficient compute resources to handle the expected burst concurrency without failing.

Answer

Configure a Small workload type with a fixed instance count of five to handle bursts.

Answer

Utilize a GPU-based workload type regardless of model size to ensure rapid scaling.

Answer

Disable autoscaling and manually update the endpoint configuration during peak hours.

Question 5

A company is using Databricks Feature Store to serve features for real-time inference. Which TWO components are required to ensure that the model can retrieve the latest feature values with sub-second latency? (Select TWO)

Accepted Answer

An Online Store (e.g., Azure Cosmos DB, AWS DynamoDB) where features are published.. Offline feature stores (Delta tables) are not optimized for the low-latency lookups required by real-time serving endpoints. To support real-time use cases, features must be published to an Online Store (like Cosmos DB or DynamoDB). The model must then be logged with information that allows it to look up these features automatically at inference time.

Answer

A high-concurrency Spark cluster to perform real-time joins on Delta tables.

Answer

A dedicated streaming job that constantly updates the Workspace Model Registry.

Answer

A pre-warmed cache of all possible feature combinations in the Model Serving memory.

Question 6

An organization wants to implement a 'Shadow Deployment' for a new fraud detection model. How is this typically achieved in Databricks Model Serving, and what is its primary benefit?

Accepted Answer

By configuring the endpoint to log all requests, then running the new model offline on those logs.. In a Shadow Deployment, the new model receives the same production traffic as the current 'Champion' model, but its results are not used to make decisions. This allows teams to compare the performance and behavior of the new model against the old one in a real-world setting without risking any negative impact on the business process.

Answer

By routing 100% of traffic to both models and returning the average of their scores.

Answer

By using a 50/50 traffic split and manually discarding half of the results.

Answer

By deploying the model in 'ReadOnly' mode which prevents it from accessing the database.

Question 7

What happens when a model serving endpoint is configured with a 'Small' workload type but the model's Python environment requires 16GB of RAM to load the weights?

Accepted Answer

The endpoint will fail to initialize, and logs will show an OutOfMemory (OOM) error.. If the model's memory requirements exceed the resources allocated to the workload type, the container will fail to start or will crash during the loading phase due to OutOfMemory (OOM) errors. It is essential to match the workload size to the specific resource demands of the model's architecture and its associated libraries.

Answer

The endpoint will automatically upgrade the workload type to 'Medium' to accommodate the model.

Answer

The model will load successfully but will use virtual disk swap, significantly increasing latency.

Answer

The model weights will be partially loaded, resulting in decreased prediction accuracy.

Question 8

A data science team is transitioning from batch prediction to real-time serving using MLflow. They need to ensure that the deployment process is reproducible and allows for easy rollback. Which TWO actions should the team take to meet these requirements?

Accepted Answer

Log the model with all required dependencies and register it in the MLflow Model Registry.. Using MLflow Models and the Model Registry is fundamental for production lifecycle management. By logging the model with its environment dependencies (conda.yaml or requirements.txt), the team ensures the deployment environment matches the training environment. Registering the model version allows for strict version control, which is essential for auditability and implementing seamless rollback strategies when performance degradation is detected after deploying a new model version to production.

Answer

Hardcode the model file path in the production inference script to avoid registry lookups.

Answer

Manually copy the model artifact file to the target production server's local file system.

Answer

Use a local Python script to perform the model inference instead of Databricks Model Serving.

Question 9

A machine learning engineer has registered a model in the Unity Catalog Model Registry and wants to serve it with Databricks Model Serving. The model's requirements.txt includes a private Python package hosted in an internal PyPI repository that requires authentication. Which configuration step is necessary to ensure the serving endpoint can install the package?

Accepted Answer

Specify the private PyPI repository URL and credentials in the serving endpoint's environment configuration using the `pip_repository` field.. The serving endpoint's environment configuration supports a `pip_repository` section where you can specify the URL and credentials for a private PyPI repository. This allows the endpoint build to authenticate and install the required package. Other methods, such as environment variables or custom Docker images, either do not affect pip authentication or add unnecessary complexity.

Answer

Configure the serving endpoint to use a custom Docker image that already has the private package installed.

Answer

Include the package as a wheel file in the model artifacts and reference it with a relative path in requirements.txt.

Answer

Add the private PyPI repository URL and credentials to the environment variables of the serving endpoint.

Question 10

An ML engineer is deploying a model to a Databricks Model Serving endpoint and needs to ensure that only authenticated users with specific permissions can invoke the endpoint. Which mechanism should they use?

Accepted Answer

Configure Unity Catalog privileges on the model and endpoint.. Unity Catalog privileges on the model and endpoint enable fine-grained, identity-based access control, ensuring only authorized users can invoke the endpoint. Other options either do not enforce user-level permissions or rely on insecure practices.

Answer

Enable token-based authentication on the endpoint.

Answer

Use a network security group to restrict IP addresses that can access the endpoint.

Answer

Embed the endpoint URL in a private application and rely on obscurity.

Question 11

A machine learning engineer is deploying a model to Databricks Model Serving and needs to ensure that the endpoint can automatically scale based on traffic. The engineer wants to configure autoscaling to handle variable load. Which TWO actions should the engineer take? (Choose two.)

Accepted Answer

Set the minimum and maximum number of concurrent requests for the endpoint.. To enable autoscaling for a Model Serving endpoint, the engineer must enable autoscaling and set the minimum and maximum concurrent requests. These settings allow the endpoint to dynamically adjust its capacity based on traffic, ensuring optimal performance and cost efficiency. Other options like schedules or cluster attachment are not applicable to Model Serving.

Answer

Attach a Databricks cluster with autoscaling enabled to the serving endpoint.

Answer

Enable GPU acceleration to allow the endpoint to handle more requests per second.

Answer

Configure a schedule for the endpoint to scale up during business hours and down after hours.

Question 12

An ML engineer is deploying a model to Databricks Model Serving and needs to ensure the endpoint can access sensitive data stored in Unity Catalog during inference. The model requires reading a feature table at request time. Which TWO configurations are necessary to enable this securely? (Choose two.)

Accepted Answer

Grant the serving endpoint's service principal SELECT privileges on the feature table.. To securely access Unity Catalog data during inference, the serving endpoint's service principal must have SELECT privileges on the feature table, and the endpoint must be configured to allow Unity Catalog access. These two configurations ensure both authorization and connectivity, enabling the model to read features securely.

Answer

Store the feature table as a Delta table in the workspace's default database.

Answer

Use a personal access token (PAT) embedded in the model to authenticate to Unity Catalog.

Answer

Configure the endpoint to run with a cluster that has the necessary libraries installed.

Question 13

A data scientist has registered a model in Unity Catalog and wants to deploy it to a Databricks Model Serving endpoint. They need to ensure that only the latest version of the model is served and that the endpoint automatically updates when a new version is registered. Which configuration should they use?

Accepted Answer

Use a model alias (e.g., 'Champion') in the endpoint configuration, and update the alias to point to the new version when it is registered.. Model aliases in Unity Catalog provide a stable reference to a model version. By pointing the serving endpoint to an alias such as 'Champion', you can promote a new model version by moving the alias, and the endpoint will automatically serve the new version. This enables continuous deployment and easy rollback without endpoint reconfiguration.

Answer

Deploy the model using a job that periodically checks for new versions and restarts the endpoint with the updated version.

Answer

Configure the endpoint to serve the model by referencing the model name with the 'latest' version alias.

Answer

Create an endpoint that serves the model version by specifying the version number directly, and update the endpoint configuration when a new version is registered.

Question 14

An ML engineer is deploying a scikit-learn model to Databricks Model Serving. The model must be served on a GPU-enabled workload to meet latency requirements. Which combination of configuration settings should be used when creating the serving endpoint?

Accepted Answer

Set workload_type to 'GPU_SMALL' and workload_size to 'Small' in the endpoint configuration.. Databricks Model Serving allows specifying a GPU workload type to accelerate inference for compatible models. The correct configuration uses workload_type 'GPU_SMALL' with an allowed workload_size such as 'Small'. This ensures the endpoint runs on GPU hardware, meeting the latency requirement. Other options either use invalid values or rely solely on CPU scaling, which does not provide GPU acceleration.

Answer

Set workload_type to 'GPU' and workload_size to 'Medium' in the endpoint configuration.

Answer

Set workload_type to 'GPU_SMALL' and workload_size to 'Large' in the endpoint configuration.

Answer

Set workload_type to 'CPU' and workload_size to 'Large' to provide more CPU cores for parallel inference.

Question 15

A data scientist has registered a model in Unity Catalog and wants to serve it in real-time. They need to ensure that the endpoint can be updated without downtime when a new model version is ready. Which deployment mode should they use?

Accepted Answer

Blue-green deployment using aliases. Blue-green deployment using aliases in Unity Catalog enables seamless model updates by switching the alias to a new version. This approach avoids downtime and allows for quick rollback if issues arise. It is the recommended method for real-time serving when continuous availability is required.

Answer

Direct deployment with version pinning

Answer

Canary deployment with traffic splitting

Answer

Serverless deployment with auto-scaling