Model Monitoring
Model monitoring is the ongoing practice of tracking a deployed machine learning model's performance, input data characteristics, and predictions in production to detect degradation, drift, or failures.
Definition
Model monitoring is the ongoing practice of tracking a deployed machine learning model's performance, input data characteristics, and predictions in production to detect degradation, drift, or failures.
Overview
A model's performance at deployment time is not guaranteed to persist — the real world changes, input data distributions shift, and the relationships a model learned during training can become outdated, a phenomenon captured by concept drift and data drift. Model monitoring is the practice of continuously observing a model's behavior after deployment to catch these problems early, before they cause meaningful business or user harm. Model monitoring typically operates across several layers. Operational monitoring tracks infrastructure-level metrics like latency, throughput, and error rates, similar to monitoring for any production service. Data quality monitoring checks that incoming features match expected schemas, ranges, and distributions, flagging missing values, unexpected categories, or upstream pipeline failures. Statistical drift monitoring compares the distribution of production inputs (and sometimes outputs) against the training distribution using metrics like population stability index (PSI) or Kolmogorov-Smirnov tests, to detect data drift. Performance monitoring tracks accuracy-related metrics directly, which requires eventually obtaining ground-truth labels for production predictions — sometimes available quickly (e.g., whether a fraud flag was correct), sometimes delayed by weeks or months (e.g., loan default outcomes), and sometimes never available at all, which is why proxy metrics and drift detection matter so much. Dedicated MLOps monitoring tools — such as Evidently AI, Arize, Fiddler, WhyLabs, and cloud-native offerings within SageMaker Model Monitor or Vertex AI Model Monitoring — provide dashboards, alerting, and automated drift detection so teams don't have to build this tooling from scratch. For large language model applications specifically, monitoring has expanded to include tracking hallucination rates, toxicity, latency per token, and cost per request. Effective model monitoring often feeds directly into retraining triggers and human-in-the-loop review workflows, closing the loop between production observation and model improvement.
Key Concepts
- Tracks operational metrics like latency, throughput, and error rates in production
- Monitors input data quality against expected schemas and ranges
- Detects data drift by comparing production and training feature distributions
- Tracks performance metrics against ground-truth labels when available
- Uses statistical tests such as population stability index and Kolmogorov-Smirnov
- Supported by dedicated tools like Evidently AI, Arize, Fiddler, and WhyLabs
- Extended for LLM applications to track hallucination rate, toxicity, and cost
- Feeds into retraining triggers and human review workflows