You want to route a small portion of traffic to a new model version for preview users, evaluate it, and then gradually shift all traffic to the new model over a fixed timeframe with minimal operational overhead. Which SageMaker approach accomplishes this with the least operational work?

The correct approach leverages SageMaker's built-in traffic shifting capabilities with UpdateEndpointWeightsAndCapacities. This allows you to update an existing endpoint to include a new model variant with an initial weight of 0, effectively making it inactive for general traffic. Preview users can then be directed to this specific variant using the TargetVariant parameter in their InvokeEndpoint calls. As confidence grows, you can gradually increase the DesiredWeight of the new variant, shifting traffic until it serves all users, all within a single SageMaker endpoint. The first incorrect option, using CreateEndpointConfig, creates a new endpoint configuration, which would require updating the endpoint itself or creating a new endpoint, adding operational overhead. The second and fourth options, using an Application Load Balancer or Route 53 with two separate endpoints, introduce external infrastructure management and increase operational complexity compared to SageMaker's native endpoint update features.