MLA-C01 Practice Question 17: Canary Deployment Guardrails
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 2 · question 7 of 10
A bank serves a credit-limit model from a SageMaker AI real-time endpoint and needs to roll out a retrained version. The rollout policy says the new version must first answer a small share of real customer requests on its own fleet for 10 minutes. If a CloudWatch alarm on 5XX errors or model latency fires during that window, all traffic must return to the current version automatically, with no operator involvement. Only after a clean window may the remaining traffic move over.
How should the ML engineer perform the update?
Show answer and explanation
Answer:
SageMaker AI deployment guardrails implement this policy directly. In a blue/green update, SageMaker AI provisions a new (green) fleet alongside the old (blue) one. With canary traffic shifting it routes a portion of traffic (the canary, up to 50% of the green fleet's capacity) to the new fleet and starts a baking period set by WaitIntervalInSeconds. The CloudWatch alarms named in AutoRollbackConfiguration are watched throughout; if any trips, SageMaker AI rolls all traffic back to the blue fleet. If none do, the rest of the traffic shifts to green, and the blue fleet is terminated only after a final bake. Every part of the requirement (small real share, a 10-minute window, automatic rollback, no operator) is native behavior.
Option B is the trap. Shadow testing sends a *copy* of production requests to the new version, but its responses are never returned to customers. It is excellent for validating a model without risk, but the policy explicitly requires the new version to answer real customer requests, and promotion from a shadow test is a separate step rather than an alarm-driven rollback.
Option C does expose real traffic to the new variant, but nothing rolls it back automatically: a human watches the alarms and shifts weights by hand, which is exactly the operator involvement the policy forbids.
Option D has the right guardrails in the wrong shape. All-at-once routing shifts 100% of traffic to the green fleet in a single step and then bakes, so every customer request, not a small share, is exposed to the new version during the window.
Reference: Use canary traffic shifting