To continuously monitor machine learning models and data pipelines for shifts in data distribution or concept drift, triggering alerts for retraining or investigation.

To alert ML engineers and MLOps practitioners to potential model degradation so they can investigate causes and retrain models as needed.

When the distribution of input features shifts, when the relationship between features and target changes, when model performance metrics degrade, or when new data deviates significantly from the training distribution.

Machine learning model monitoring, MLOps pipelines, data drift detection systems, online learning systems, AI observability platforms.

monitor data distribution → detect statistical shift → trigger alert or retraining

drift_score = statistical_distance(reference_distribution, current_distribution)

• data drift
• concept drift
• model drift
• feature drift
• prediction drift

• drift detection
• drift monitoring

Confusing drift detection with simple accuracy monitoring; ignoring feature-level drift; using inappropriate statistical tests for high‑dimensional data.

model performance monitoring, outlier detection, data validation

Confusing drift detection with data validation checks; applying drift detection to non‑stationary processes without adjusting for expected changes.

• model monitoring
• data drift
• concept drift
• MLOps monitoring

Drift detection focuses on statistical shifts in data distribution or concept, not merely on accuracy drops; it can detect degradation before performance metrics fall.

Triggers alerts that prompt model retraining, data collection, or investigation into upstream data pipeline changes.

Think of a model's performance dropping when the world changes—drift detection spots the change before the performance drops.

Drift detection methods differ in statistical assumptions and computational cost; select tests appropriate for data dimensionality and sample size.

Advanced drift detection techniques (e.g., ADWIN, DDM, EDDM, Page‑Hinkley)