Question

A retail company deploys a demand-prediction model to a SageMaker real-time endpoint. Over several weeks, the ML team notices that predictions have degraded, and they suspect the statistical properties of the incoming inference request features no longer match the data used during training. The team wants an automated, ongoing way to compare live inference traffic against the training distribution and be alerted when the input features drift significantly. Which approach requires the LEAST custom development to meet this requirement?

Accepted Answer

Enable data capture on the endpoint, generate a data quality baseline from the training dataset with SageMaker Model Monitor, and schedule a monitoring job that emits violation metrics to CloudWatch for alarms.

Answer

Write a scheduled AWS Lambda function that pulls a sample of endpoint invocations from CloudWatch Logs, computes feature histograms in NumPy, and compares them to hard-coded training statistics.

Answer

Configure SageMaker Model Monitor model quality monitoring with ground-truth labels merged in to detect changes in accuracy over time.

Answer

Enable SageMaker Clarify feature attribution drift monitoring to compute SHAP values on each request and alert when attribution rankings change.