AWS Certified Machine Learning – SpecialtyMachine Learning Implementation and OperationsMedium

A media company uses a content moderation model deployed on a SageMaker real-time endpoint. They have developed an improved version of the model (Model B) and want to deploy it with minimal risk, gradually shifting traffic from the current production model (Model A) to Model B. They need to observe Model B's performance with a small percentage of live traffic before fully committing. Which deployment strategy should they use?

  1. AShadow Deployment
  2. BRollback Deployment
  3. CBlue/Green Deployment
  4. DCanary Deployment
Show answer & explanation

Correct answer: D. Canary Deployment

Canary deployment allows a new model (Model B) to receive a small, controlled percentage of live traffic alongside the existing production model (Model A). This enables real-world performance monitoring with minimal risk, as the majority of traffic still goes to the stable Model A. If Model B performs well, the traffic can be gradually increased.

Why the other options are wrong

  • A. Shadow deployment sends production traffic to both models but only uses the incumbent model's predictions. The new model's predictions are not used in production, which doesn't allow for observing its impact on actual user experience or system metrics with live outcomes.
  • B. Rollback deployment is a strategy for reverting to a previous version if a new deployment fails, not for gradually introducing a new model.
  • C. Blue/Green deployment switches all traffic at once to the new version after testing, which might not be 'gradual' or 'minimal risk' for observing performance with a small percentage.

Canary Deployment (ML)

A deployment strategy where a new model version is introduced to a small subset of live inference traffic, allowing for real-world performance monitoring before a full rollout.

  • Gradual traffic shift from old to new model.
  • Minimizes risk by limiting exposure of new model.
  • Enables real-time performance validation with live data.

Memory trick: Canary's Song, Small Taste, Safe Rollout Strong.

More Machine Learning Implementation and Operations questions