Question

A machine learning team retrains a fraud detection model weekly and produces multiple candidate model artifacts. They need a way to catalog each trained model, track its lineage and evaluation metrics, and enforce a manual review step so that only vetted models can be deployed to the production endpoint through their CI/CD pipeline. Which approach best satisfies these requirements with minimal custom engineering?

Accepted Answer

Register each candidate model as a versioned entry in a SageMaker Model Registry model package group, set the approval status to PendingManualApproval, and have the CI/CD pipeline deploy only packages marked Approved.

Answer

Store each model artifact in an S3 bucket with versioning enabled and use an S3 object tag to indicate whether a model is approved for deployment.

Answer

Save all model artifacts to a single SageMaker training job output path and rely on the most recent training job timestamp to determine which model to deploy.

Answer

Create a separate SageMaker endpoint for each candidate model and manually delete the endpoints that fail review.