Detect drift with SageMaker Model Monitor

Steps

1. Enable data capture on the inference endpoint by adding a DataCaptureConfig to the endpoint configuration, specifying a destination S3 URI and a capture percentage
2. Create a baseline from a representative dataset using a DefaultModelMonitor.suggest_baseline() job, which generates constraint and statistics JSON files in S3
3. Create a monitoring schedule: monitor.create_monitoring_schedule(monitor_schedule_name='my-schedule', endpoint_input=endpoint_name, output_s3_uri='s3://your-bucket/monitor-output', statistics=baseline_stats_uri, constraints=baseline_constraints_uri, schedule_cron_expression='cron(0 * ? * * *)')
4. After scheduled executions run, retrieve results with monitor.list_executions() and inspect violation reports in the output S3 path
5. Set up a CloudWatch alarm on the metric aws/sagemaker/Endpoints/data-metrics to receive notifications when constraint violations exceed a threshold
6. Stop the monitoring schedule when the endpoint is decommissioned: monitor.stop_monitoring_schedule()