Meridian Line wants to promote a new version of the customs-compliance agent cautiously: send it 10% of live traffic while the existing version keeps handling the rest, and watch quality before cutting over fully. What Model Serving capability supports this directly on one endpoint?
Select an answer to reveal the explanation.
Short Explanation
One endpoint can serve more than one model version at once, each with a traffic-percentage weight. That's the built-in canary lever — no client-side coin flip required.
Full Explanation
A single Model Serving endpoint can host multiple served entities (model versions) simultaneously, each configured with a traffic-percentage weight, so the endpoint itself splits incoming requests according to those weights — 10% to the new version, 90% to the incumbent, with no change needed on the caller's side. This is the direct, supported mechanism for staged promotion. Building the split into client code by calling two separate endpoint URLs at a 10/90 ratio works around the platform instead of using it, adds client-side complexity, and loses the single-endpoint metrics/logging view that makes comparing the two versions straightforward. Aliases are pointers to a specific version, not a traffic-shaping mechanism — assigning an alias named 'canary' has no special meaning to Model Serving and does not create an automatic percentage cap; the name is just a label. Scale-to-zero controls whether an endpoint's compute spins down to zero when idle to save cost; it has nothing to do with proportionally routing traffic between two versions, and 'throttles it down to 90% of prior capacity' is not a real scale-to-zero behavior. Exam framing: traffic splitting is the named capability under 'manage model versions and aliases, and roll out changes safely (traffic splitting, staged promotion).'