Question

Your team runs a fraud-detection model behind a Vertex AI Endpoint serving live transactions. You have trained an improved model version and want to validate it in production before full rollout. You need to route a small fraction of live traffic to the new model, monitor its behavior, and be able to instantly revert if error rates spike — all without exposing users to a new URL or causing downtime. What is the most appropriate approach?

Accepted Answer

Deploy the new model to the same endpoint alongside the current model and configure a traffic split (e.g., 90/10), gradually increasing the new model's share while monitoring metrics.

Answer

Create a second Vertex AI Endpoint for the new model and update your application code to send 10% of requests to the new endpoint URL.

Answer

Undeploy the current model and deploy the new model to the endpoint, then roll back manually if issues appear.

Answer

Run the new model as a batch prediction job against a sample of recent transactions and compare results offline before replacing the online model.