DataWave Inc. runs a low-latency fraud model on a SageMaker real-time endpoint (single production variant). They want to perform a shadow test by routing 5% of live production traffic to a new model variant (the shadow) while 95% continues to the current variant; responses from the shadow model should not be returned to clients (they will be logged for offline comparison). What is the most appropriate, least-intrusive way to implement this within SageMaker?

The most appropriate method is to add the new model as a production variant with a low InitialVariantWeight (e.g., 0.05) to the existing SageMaker endpoint. SageMaker's built-in traffic splitting will automatically route the specified percentage of requests to the new variant. Since the goal is a shadow test, the responses from this new variant are for internal logging and comparison, not for clients. This is the least intrusive as it leverages SageMaker's native A/B testing capabilities without requiring external infrastructure or complex client-side logic. Incorrect options: Discarding shadow responses on the client using client-side logic is intrusive and adds unnecessary complexity to the client application. Using an ALB with request mirroring is overly complex and introduces external infrastructure when SageMaker can handle this natively. A multi-model endpoint is designed for hosting many models on shared infrastructure, not for A/B testing or shadow testing with traffic splitting between distinct models. The TargetModel header is for routing to specific models within the multi-model endpoint, not for percentage-based traffic splitting of live requests to a shadow model.