Scenario: Multiple recommendation models must be evaluated using A/B testing in production. The system must route live inference traffic, monitor…

Amazon SageMaker multi-variant endpoints are the purpose-built, fully managed solution for this use case. They allow multiple models (variants) to be deployed behind a single endpoint, simplifying the client application. SageMaker handles the weighted distribution of live inference traffic to each variant for A/B testing. Performance metrics for each variant are emitted to Amazon CloudWatch automatically. Once the best-performing model is identified through engagement metrics, the endpoint configuration can be updated via a single API call to route 100% of the traffic to that variant. This approach abstracts away the underlying infrastructure management, offering the highest operational efficiency.

B. Using separate SageMaker endpoints with API Gateway introduces unnecessary architectural complexity and management overhead compared to the integrated, single-endpoint solution provided by SageMaker multi-variant endpoints.

C. AWS CodeDeploy with a blue/green strategy is primarily for safe, zero-downtime application updates by shifting traffic entirely from an old to a new version, not for simultaneous A/B testing.

D. This manual IaaS approach has the highest operational overhead. It requires managing EC2 instances, model deployments, scaling, and load balancer configurations, making it the least operationally efficient option.

1. Amazon SageMaker Developer Guide: Under the section "Test models in production," the guide explicitly describes using multi-variant endpoints for A/B testing. It states, "With SageMaker, you can test multiple models by deploying them to the same endpoint... SageMaker can then distribute traffic between your models by weight that you specify." This directly supports option A as the intended and most efficient method. (See: Amazon SageMaker Developer Guide -> Deploy models for inference -> Test models in production -> A/B Testing with Amazon SageMaker).

2. Amazon SageMaker API Reference: The UpdateEndpointWeightsAndCapacities API action demonstrates the operational simplicity of shifting traffic. The documentation shows how to "update the weights and capacity of one or more variants associated with an endpoint," which is the mechanism to seamlessly direct 100% of traffic to the winning model with a single command.

3. AWS Blog - A/B test machine learning models in production using Amazon SageMaker: This official blog post details the exact scenario. It highlights, "Amazon SageMaker multi-variant endpoints are a great way to test multiple models in production... This is a fully managed option that takes away the undifferentiated heavy lifting of managing the underlying infrastructure." This confirms the superior operational efficiency of the SageMaker native solution over manual setups.