An ML team has produced a new version of a recommendation model. They want only a small slice of live requests to reach it at first, with that share raised step by step as dashboards confirm quality. Which rollout approach fits?
Show answer and explanation
Answer: C
A canary rollout sends a small share of live traffic to the new model version, then raises that share in stages while monitoring latency, errors and prediction quality. If metrics degrade, traffic is shifted back quickly with little impact. Vertex AI endpoints support this through traffic splits across deployed models. Blue-green swaps everything at once, mirroring discards responses, and per-request random assignment has no controlled ramp.
Why the other options are wrong
- A. Switching all requests at once is a blue-green style cutover, not a gradual rollout. It gives no chance to observe the new version on a small share of real traffic before full exposure.
- B. Shadow or mirrored traffic sends copies of requests to the new version, but its responses are discarded, so customers never see its output. That validates behavior but is not a progressive live traffic ramp.
- D. Random per-request assignment has no fixed split, no controlled step increase, and no clear rollback point. Teams cannot reason about the new version's share of traffic when it is reassigned on every call.