model monitoring

Model monitoring tracks deployed model performance, detecting degradation and triggering retraining or alerts. **What to monitor**: **Performance metrics**: Accuracy, latency, throughput over time. Compared against baseline. **Data quality**: Input distribution shifts, missing values, outliers, schema violations. **Model behavior**: Prediction distributions, confidence scores, feature importance changes. **Infrastructure**: GPU utilization, memory, queue depth, error rates. **Why monitoring matters**: Models degrade over time due to data drift, concept drift, or system issues. Silent failures are costly. **Alerting**: Set thresholds for key metrics, alert on degradation, escalation policies. **Tools**: Evidently AI, WhyLabs, Arize, Fiddler, custom dashboards with Prometheus/Grafana. **Ground truth delay**: Labels may arrive days/weeks later. Use proxy metrics, statistical tests in meantime. **Dashboard design**: Real-time performance, trend analysis, comparison to baseline, segmented analysis. **Response to alerts**: Investigate root cause, consider rollback, trigger retraining if needed. **SLAs**: Define acceptable performance ranges, document monitoring coverage.