drift

**Drift** Monitoring for data drift and model drift detects when input distributions or model performance change over time, triggering alerts for investigation and potential retraining to maintain model quality in production. Data drift: input feature distributions change from training data; model may perform poorly on unfamiliar inputs. Types: covariate shift (X distribution changes), label shift (Y distribution changes), and concept drift (P(Y|X) changes). Detection methods: statistical tests (KS test, chi-squared), distribution distance metrics (KL divergence, Wasserstein distance), and threshold-based monitoring. Feature monitoring: track statistics (mean, variance, min, max) and distributions per feature; alert on significant deviation. Model drift: model accuracy degrades over time even without explicit data drift; detect through performance monitoring. Performance monitoring: track metrics (accuracy, F1, latency) on live predictions; requires ground truth labels (may be delayed). Reference windows: compare current data/performance against training baseline or rolling window. Alert thresholds: balance sensitivity (catch drift early) against false positives (alert fatigue). Response: investigate drift cause, determine if retraining needed, and update reference distributions after retraining. Tools: Evidently, NannyML, Fiddler, and custom dashboards. Documentation: log all drift events, investigations, and actions taken. Drift monitoring is essential for maintaining model reliability in production.