Blue/green, canary, linear — and rolling back
Lesson 2 of 4 in MLOps Pipelines: CI/CD for Models.
Shipping a model is a bet; a deployment strategy decides the stake. SageMaker AI's native machinery is deployment guardrails on UpdateEndpoint: a blue/green deployment stands up the new fleet (green) beside the old (blue) and shifts traffic in one of three modes — all at once, canary (a small slice first), or linear (n equal steps) — each watched by pre-specified CloudWatch alarms during a baking period. An alarm trip triggers automatic rollback to blue. A separate rolling mode replaces instances in batches instead of running two full fleets. The exam plays this as pick-the-strategy: match the risk sentence to the mode.
| Strategy | Traffic story | Rollback story | The tell |
|---|---|---|---|
Blue/green, all at once | 100% flips to green after it stands up | Instant — blue is kept through the baking period | "Fast cutover but keep the old fleet ready" |
Blue/green, canary | A small slice bakes on green first; rest follows on success | Automatic on alarm; only the canary slice ever saw the bad model | "Minimize the number of users a bad model can affect" |
Blue/green, linear | Equal steps (e.g., 10% at a time), each with its own bake | Automatic on alarm at whichever step tripped it | "Gradual, predictable shift in stages" |
Rolling | Instance batches are replaced in place | Alarm stops the roll; no second full fleet needed | "Capacity-conscious — avoid doubling the fleet during deploys" |
Interactive sorting exercise: Match the requirement in each stem to the traffic-shifting mode it is asking for.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.