MLA-C01 Question 350
Select 3You are deploying a new version of a machine learning model to production using Amazon SageMaker. The model serves real-time predictions for a critical application, and you want to minimize the risk of introducing a faulty model while ensuring a smooth transition for users. Which deployment strategy should you choose to achieve this, and which rollback action would be most effective if issues are detected?
- A
Use a blue/green deployment strategy and switch traffic back to the old model if issues occur.
- B
Use a canary deployment strategy, starting with a small percentage of traffic for the new model, and roll back if errors are detected.
- C
Use a linear deployment strategy to gradually increase traffic to the new model over time, with the option to stop and roll back if issues arise.
- D
Deploy the new model directly to production and monitor performance closely, rolling back manually if needed.
Show answer and explanation
Correct answers: A, B, C
Explanation
When deploying a new machine learning model, strategies like blue/green, canary, and linear deployments are designed to minimize risk and provide rollback options. Blue/green ensures a complete separation of old and new environments, canary tests the new model with a small subset of users, and linear gradually shifts traffic, all allowing controlled rollbacks if issues arise. Direct deployment, while simpler, lacks the safety mechanisms of these strategies and is not ideal for critical applications.
- A. Correct.
Blue/green deployment is a safe approach where the new model is deployed alongside the old one, and traffic is shifted only after verifying the new model's performance. This allows for an immediate rollback to the old model if needed.
- B. Correct.
Canary deployment is an effective strategy to minimize risk, as it starts with a small percentage of traffic for the new model. If issues are detected, it is straightforward to roll back to the previous version.
- C. Correct.
Linear deployment allows for gradual traffic increases, providing an opportunity to monitor the new model's behavior over time. If issues arise, the deployment can be stopped, and traffic can be reverted to the old model.
- D. Incorrect.
Direct deployment to production carries higher risk, as it does not involve gradual traffic shifting or verification steps. Although manual rollback is possible, it is less controlled and not recommended for critical applications.