FinTrust Inc. trains a tabular fraud-detection model and stores the training CSV in s3://fintrust-training/transactions/. The ML platform team must create a SageMaker Model Monitor data-quality baseline (statistics and constraints) from the training data so that daily monitoring of production inference input schema and distributions can alert on drift. Which is the most appropriate, minimal-effort way to produce a Model Monitor baseline suitable for use by a MonitoringSchedule?

The most appropriate and minimal-effort method is to use SageMaker Model Monitor's built-in suggestbaseline functionality (or the underlying CreateProcessingJob). This directly leverages the Model Monitor container to analyze the training data from the specified S3 path and automatically generate the necessary statistics and constraints JSON files, which are then saved to an S3 baseline path. This is designed precisely for this use case, ensuring compatibility and ease of integration with a MonitoringSchedule. Enabling data capture on a production endpoint for a week to auto-generate a baseline is incorrect because baselines should ideally be generated from the training data, which represents the expected data distribution. Using production data for a baseline can introduce bias if drift has already occurred. Using AWS Glue to compute statistics and manually converting them to Model Monitor's JSON format is overly complex and error-prone; Model Monitor provides a direct way to do this. Running Athena queries and storing custom CloudWatch metrics is also unnecessarily complex and doesn't directly produce the Model Monitor-specific statistics and constraints JSON required for a MonitoringSchedule.