RetailStream LLC has a classification endpoint that captures predictions to s3://retailstream-capture/predictions/. They also have labeled ground truth produced by a SageMaker Ground Truth labeling job stored as a manifest in s3://retailstream-labels/manifest/. The data-science team needs model-quality monitoring that compares predictions with the true labels. What is the correct way to supply ground truth to SageMaker Model Monitor so the MonitoringSchedule can compute model quality metrics?

The correct approach is to pre-process and merge the prediction and ground truth data. SageMaker Model Monitor requires a single input dataset containing both the model's predictions and the corresponding true labels to compute model quality metrics. A SageMaker Processing job or Lambda function can effectively perform this ETL (Extract, Transform, Load) task by joining the captured predictions with the ground truth manifest using a common record identifier. The resulting merged dataset, with both prediction and truelabel columns, is then stored in S3 and provided as input to the Model Quality Monitoring job. Incorrect options: Model Monitor does not automatically join separate ground truth manifests with captured predictions; it expects a unified dataset. EventBridge and CloudWatch logs are not designed for this data merging task, and Model Monitor doesn't automatically fetch predictions from logs or merge them with a manifest this way. Simply uploading the manifest to the same S3 prefix without merging is insufficient, as Model Monitor requires a single, joined input file, and it does not infer joins based on timestamps across separate files.