MLA-C01 Question 250
Select 3You are deploying a new version of a machine learning model to an Amazon SageMaker endpoint in production. To minimize the risk of user impact while ensuring you can quickly revert to the previous version if issues arise, which of the following deployment strategies should you implement?
- A
Canary deployment with gradual traffic shifting to the new model version
- B
Blue/Green deployment to maintain two separate environments and switch only when ready
- C
Deploy the new model version directly to the production endpoint without testing
- D
Enable endpoint versioning to maintain multiple versions of the model
- E
Roll back to the previous version only after manually verifying logs for errors
Show answer and explanation
Correct answers: A, B, D
Explanation
Deployment best practices in machine learning ensure reliability and minimize the risk of user impact. Strategies like canary deployment, blue/green deployment, and endpoint versioning are commonly used to safely transition to new model versions with options for rollback. These practices allow for testing and monitoring in a controlled way, reducing the chance of disruption in production environments.
- A. Correct.
Correct: Canary deployment allows you to gradually shift traffic to the new model version, enabling monitoring of its performance and quick rollback if needed. This minimizes user impact.
- B. Correct.
Correct: Blue/Green deployment ensures the old version remains intact in a separate environment, allowing a safe transition and quick rollback if issues occur.
- C. Incorrect.
Incorrect: Deploying directly to production without testing increases the risk of issues affecting users and is not considered a best practice.
- D. Correct.
Correct: Endpoint versioning in SageMaker allows you to maintain and roll back to multiple model versions, facilitating safe experimentation and recovery.
- E. Incorrect.
Incorrect: While rollback is important, relying solely on manual log verification before rolling back is inefficient and prone to delays in addressing user impact.