AI-300 Practice Question 4: Mirrored Traffic Validation
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 1 · question 4 of 10
A managed online endpoint serves the blue deployment of a fraud model. A retrained green deployment is ready, and the release engineer wants to observe how it behaves on real production requests, latency and HTTP errors, before any customer sees a prediction from it. Which action meets that requirement?
Show answer and explanation
Answer:
Traffic mirroring, also called shadowing, copies a percentage of live traffic and submits it to the new deployment as well. The results returned to clients don't change, every request is still answered by blue, so you gather metrics and logs from green with no customer impact (option D). Microsoft documents it for exactly this check: whether latency is within acceptable bounds and whether there are HTTP errors. You can mirror up to 50 percent of traffic, to one deployment per endpoint.
- A is wrong because allocated traffic is *served* traffic. Ten percent of real customers would receive predictions from the untested deployment, which is the outcome the engineer wants to avoid at this stage. - B is wrong because replayed logs are stale and a separate endpoint has its own scaling and networking. It shows how green handles yesterday's request mix, not current live conditions. - C is wrong because a batch endpoint scores asynchronously on a different serving stack. It tells you nothing about the real-time latency or error behavior of an online deployment.
Reference: Safe rollout for online endpoints - Azure Machine Learning