GCP Professional ML Engineer Practice Question 7: Endpoint Traffic Split Canary
Google Cloud Professional Machine Learning Engineer · Test 1 · question 7 of 10
Version 4 of a recommendation model has passed offline evaluation, and version 3 is currently serving all production traffic from an Agent Platform endpoint. Before a full cutover you want version 4 to answer about 10% of live requests so its business metrics can be compared against version 3. Client applications must not be reconfigured, and rollback has to be immediate.
What should you do?
Show answer and explanation
Answer:
Deploying more than one model to a single endpoint exists for exactly this workflow. The docs describe it as the way to gradually replace one model with another: the new model goes onto the existing endpoint serving a small percentage of traffic, and you raise its share until it serves 100%. Because compute resources are associated with each deployed model rather than with the endpoint, both versions run side by side, the endpoint URL never changes so clients need no reconfiguration, and rollback is a single traffic-split edit back to 100/0.
Option A works, but it is the manual reconstruction of a built-in feature, and it violates the stem's constraint: pointing clients at a load balancer in front of a new endpoint is a client-side reconfiguration, plus a second endpoint's resources and an extra load-balancing layer to operate.
Option B misunderstands what the default version does. The Model Registry default version is preselected when the model is chosen for inference (for example when creating a new deployment or a batch job), but changing it does not retarget models already deployed to an endpoint, and it offers no percentage split.
Option D is a hard cutover dressed up as a test. Every request goes to version 4 immediately, there is no concurrent baseline to compare against, and a bad model affects 100% of users until someone redeploys version 3.
Reference: Deploy a model to an endpoint