Databricks Machine Learning Associate Question 653
Single answerYou are building a real-time machine learning inference system on Databricks. Your team wants to ensure that incoming data is properly routed between different model endpoints based on specific conditions, such as customer region or device type. Which approach would be the most appropriate for splitting data between endpoints in this scenario?
- A
Use a feature store to directly route data to the appropriate model endpoint based on the input features.
- B
Implement a request routing layer that evaluates input data and sends it to the appropriate model endpoint.
- C
Use a batch processing job to split the data and send requests to the respective endpoints.
- D
Deploy all models to a single endpoint and route requests within the model code itself.
Show answer and explanation
Correct answer: B
Explanation
In real-time inference scenarios, it is crucial to have an efficient and scalable mechanism to route incoming data to the appropriate model endpoint based on specific conditions. A request routing layer serves this purpose by acting as an intermediary that evaluates the input data and directs it to the correct endpoint, ensuring low latency and flexibility. Other approaches, like using a feature store or batch processing, are not designed for real-time routing, while embedding routing logic in the model code leads to poor scalability and maintainability.
- A. Incorrect.
A feature store is used to manage and retrieve features for machine learning models but is not responsible for routing data to model endpoints.
- B. Correct.
A request routing layer is an appropriate solution for real-time inference as it can evaluate input data and make decisions to send it to the correct model endpoint. This ensures low latency and flexibility.
- C. Incorrect.
Batch processing jobs are not suitable for real-time inference as they introduce significant latency and are designed for offline or asynchronous tasks.
- D. Incorrect.
Deploying all models to a single endpoint and handling routing within the model code is not scalable or maintainable, as the logic becomes tightly coupled with the model's implementation.