Question

Your team runs a real-time model on an Azure Machine Learning managed online endpoint. The current version 'blue' serves 100% of traffic. You have trained an improved model and created a new deployment 'green' under the same endpoint. Before fully switching over, you must expose 'green' to a small percentage of live production traffic to validate its behavior, while retaining the ability to instantly revert if error rates rise. Which approach meets these requirements with minimal disruption?

Accepted Answer

Set the endpoint traffic allocation to route 90% to 'blue' and 10% to 'green', monitor 'green', then adjust the split or set 'blue' back to 100% to roll back.

Answer

Delete the 'blue' deployment and immediately allocate 100% of traffic to 'green', restoring 'blue' from source if problems occur.

Answer

Create a second endpoint for 'green' and manually update client applications to send 10% of requests to the new endpoint URL.

Answer

Use the mirror_traffic setting to send 100% of live traffic to both 'blue' and 'green' and serve responses from whichever deployment returns first.