Databricks Machine Learning Associate Question 378
Select 2A data science team has trained a machine learning model on Databricks and now needs to deploy it across multiple regions to serve predictions with low latency. Which of the following strategies are appropriate for distributing the model while ensuring scalability and availability?
- A
Deploy the model as a REST API using Databricks Model Serving and use a CDN (Content Delivery Network) to replicate the API across regions.
- B
Export the trained model as a serialized file (e.g., MLflow artifact) and deploy it manually on separate cloud instances in each region.
- C
Deploy the model to a centralized cluster and use a global load balancer to route prediction requests to it.
- D
Register the model in the MLflow Model Registry and deploy it to multiple regional clusters for local inference.
- E
Deploy the model using Databricks Unity Catalog for automated model replication and inference across regions.
Show answer and explanation
Correct answers: B, D
Explanation
To scale and distribute machine learning models effectively, deploying them in regional clusters ensures low-latency inference and high availability. While exporting a model for manual deployment allows flexibility, using the MLflow Model Registry simplifies the process by providing a standard framework for deployment and tracking. Centralized or CDN-based approaches are not suitable for dynamic model inference, and Unity Catalog does not handle model distribution.
- A. Incorrect.
Using a CDN is not suitable for model deployment because CDNs are designed to cache and distribute static content, not to replicate APIs or handle dynamic model inference.
- B. Correct.
Exporting the model as a serialized file and manually deploying it to cloud instances allows for regional scalability, but it requires additional operational overhead for deployment and management.
- C. Incorrect.
A centralized cluster might lead to high latency for users in distant regions due to network delays, and it does not ensure optimal availability for regional users.
- D. Correct.
Registering the model in the MLflow Model Registry and deploying it to multiple regional clusters ensures scalability and low-latency predictions by performing inference locally in each region. This approach aligns with best practices for distributed model serving.
- E. Incorrect.
Databricks Unity Catalog is used for managing data governance and access control but does not provide functionality for model replication or distribution across regions.