You work for an organization that operates a streaming music service. You have a custom production model that is serving a 搉ext song?recommendation based on a user's recent listening history. Your mod

Creating a new Vertex AI endpoint for the new model would require updating the application logic to route traffic to the new endpoint, increasing complexity compared to using traffic splitting on an existing endpoint.

Capturing incoming prediction requests in BigQuery is a useful step for data logging and analysis, but it doesn't directly enable testing the new model in production with live traffic alongside the old model.

Deploying the new model to the existing Vertex AI endpoint and using traffic splitting is the recommended approach for safely testing a new model version in production with minimal complexity. This allows routing a small percentage of live traffic to the new model (canary deployment) to observe its real-world performance before a full rollout.

Configuring a model monitoring job is for observing a model's performance and detecting drift or anomalies after it has been deployed and is serving traffic, not for the initial testing and gradual rollout of a new model version.