A production SageMaker endpoint runs model version v1. A new version v2 must be tested in production with minimal risk and without disrupting live traffic. What is the best approach to validate v2 before switching over?

Deploying v2 as a shadow variant is the best approach because it allows for full-scale testing of the new model version with live production traffic without any impact on end-users. The shadow variant receives a copy of all inference requests, enabling comprehensive evaluation of v2's performance, latency, and output quality against v1 in a real-world scenario. If v2 performs acceptably, it can then be promoted to handle live traffic. Routing a small percentage of live traffic (1% or 10%) to v2 (A/B testing) carries some risk of user impact if v2 has issues. Deploying to a new endpoint and sending 100% of input via a script is not a true production test and requires additional management.