AWS Certified Machine Learning – Specialty — Question 117

A company that promotes healthy sleep patterns by providing cloud-connected devices currently hosts a sleep tracking application on AWS. The application collects device usage information from device users. The company's Data Science team is building a machine learning model to predict if and when a user will stop utilizing the company's devices. Predictions from this model are used by a downstream application that determines the best approach for contacting users.
The Data Science team is building multiple versions of the machine learning model to evaluate each version against the company's business goals. To measure long-term effectiveness, the team wants to run multiple versions of the model in parallel for long periods of time, with the ability to control the portion of inferences served by the models.
Which solution satisfies these requirements with MINIMAL effort?

Answer options

• A. Build and host multiple models in Amazon SageMaker. Create multiple Amazon SageMaker endpoints, one for each model. Programmatically control invoking different models for inference at the application layer.
• B. Build and host multiple models in Amazon SageMaker. Create an Amazon SageMaker endpoint configuration with multiple production variants. Programmatically control the portion of the inferences served by the multiple models by updating the endpoint configuration.
• C. Build and host multiple models in Amazon SageMaker Neo to take into account different types of medical devices. Programmatically control which model is invoked for inference based on the medical device type.
• D. Build and host multiple models in Amazon SageMaker. Create a single endpoint that accesses multiple models. Use Amazon SageMaker batch transform to control invoking the different models through the single endpoint.

Correct answer: B

Explanation

Option B is correct because it allows the Data Science team to create an endpoint configuration with multiple production variants, enabling them to control the inference distribution easily and efficiently. Option A requires creating separate endpoints, which increases management complexity and effort. Option C does not address the requirement of running multiple versions in parallel, as it focuses on model selection based on device type. Option D involves batch processing, which is not suitable for real-time inference control.