Amazon SageMaker Model Monitor
Continuous monitoring of deployed models for drift and quality.
❓ What is it?
A monitoring capability that captures endpoint traffic and periodically checks it against a training-time baseline for data-quality drift, model-quality degradation, bias drift, and feature-attribution drift.
💡 Why does it exist?
Models rot silently: the world changes, inputs shift, and yesterday's 95% accuracy becomes today's coin flip with no error message. Monitoring turns silent decay into an actionable alert.
⏱️ When should you use it?
Attach it to any production endpoint whose decisions matter — pricing, fraud, recommendations — and whenever compliance requires evidence that a model still behaves as validated.
🗺️ Where does it fit?
It wraps SageMaker endpoints: data capture writes request/response samples to S3, scheduled monitoring jobs compare them to the baseline, and violations publish to CloudWatch for alarms and retraining triggers.
🔌 How do you integrate it?
Enable data capture on the endpoint, generate a baseline from the training set, create a monitoring schedule for each check type, and route CloudWatch alarms to SNS or an automated retraining pipeline.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Four monitor types to memorise: data quality, model quality, bias drift, feature attribution drift.
- "Model accuracy degraded over time — how to detect?" → Model Monitor, usually paired with automated retraining.