An ecommerce company has a deployed SageMaker factorization machines (FM) endpoint for product recommendations. The team trained two new models (TensorFlow and PyTorch) and needs to run A/B tests with these routing rules: send 70% of traffic to the FM model, 15% to the TensorFlow model, 15% to the PyTorch model; but for customers from Europe, route all traffic to the TensorFlow model. Which architecture supports this behavior?

The correct answer leverages SageMaker's built-in A/B testing capabilities. SageMaker endpoints support multiple production variants, allowing you to define traffic splits (weights) for each model. The TargetVariant header provides a mechanism to override these weights for specific requests, directing them to a chosen variant. This directly addresses the requirement to send 70% to FM, 15% to TensorFlow, and 15% to PyTorch, while also routing all European traffic to TensorFlow. Using Application Load Balancers (ALB) or Network Load Balancers (NLB) with separate endpoints is overly complex and doesn't fully integrate with SageMaker's A/B testing features. While ALBs can handle weighted routing and rules, managing multiple independent SageMaker endpoints adds operational overhead compared to a single endpoint with variants. Auto scaling policies are relevant for endpoint scaling, but not directly for the A/B routing logic itself.