Before fully releasing a new model, a company needs to validate it online on 10% of production traffic. The model is served by a SageMaker endpoint behind an Application Load Balancer. Which setup provides the required 10% online validation with the least operational overhead?

Adding the new model as a production variant to the existing SageMaker endpoint with a weight of 0.1 (10%) is the most efficient solution. SageMaker's production variants are designed for A/B testing and canary deployments, allowing you to route a specified percentage of inference requests to a new model without modifying your application code or infrastructure. This minimizes operational overhead. Monitoring with CloudWatch is standard practice for SageMaker endpoints. Setting the variant weight to 1 would send 100% of traffic, not 10%. Creating a separate endpoint adds unnecessary operational complexity and management overhead for a simple 10% validation. Using an ALB to split traffic is possible but more complex to configure and manage than SageMaker's built-in variant routing, and CloudTrail monitors API calls, not model invocations.