A company wants to evaluate a new ML model architecture to understand its performance before...

Step 1: Update the existing endpoint with a shadow variant. Pick a suitable duration. Start the shadow test.

Step 2: Create a shadow test.

Step 3: Use the SageMaker shadow testing dashboard to analyze the performance differences between the variants.

This is the correct order because SageMaker shadow testing can run against an existing production endpoint. The production variant continues to receive and respond to inference requests, while the shadow variant receives replicated real-time requests but does not return responses to users. That satisfies the requirement to analyze the new model with real inference traffic without affecting the production endpoint. AWS also states that the SageMaker shadow testing dashboard provides side-by-side comparison metrics for the production and shadow variants, such as latency and error rate.

Do not choose “Create a new endpoint that includes the production model and the shadow model,” because the question explicitly says the analysis must not affect the existing production endpoint. Also, do not use the Amazon SageMaker Model Monitoring dashboard here; the specific feature for comparing variants in this scenario is the SageMaker shadow testing dashboard.