driftwatch 0.4.0

Lightweight ML drift monitoring, built for real-world pipelines

DriftWatch

Lightweight ML drift monitoring, built for real-world pipelines

---

📖 Documentation

Read the full documentation here: vincentcotella.github.io/DriftWatch

🚀 Features

• Multi-Drift Monitoring:
  • 📊 Feature Drift: Monitor input data distribution changes (P(X)).
  • 🎯 Prediction Drift: Monitor model output changes (P(Ŷ)).
  • 🧠 Concept Drift: Monitor model performance degradation (P(Y|X)).
• Unified Interface: DriftSuite combines all monitors in one simple API.
• 7 Statistical Detectors:
  • PSI, KS-Test, Wasserstein, Jensen-Shannon, Anderson-Darling, Cramér-von Mises, Chi-Squared.
• Explainability: Built-in statistical explanation (DriftExplainer) and visualization (DriftVisualizer).
• Production Integrations:
  • ⚡ FastAPI Middleware
  • 📈 MLflow Tracking
  • 🔔 Slack & Email Alerts
• Lightweight & Robust: Minimal dependencies, 100% type-safe.

📦 Installation

pip install driftwatch

For specific extras:

pip install driftwatch[viz]     # Visualization support
pip install driftwatch[mlflow]  # MLflow integration
pip install driftwatch[all]     # CLI, API, Alerting, etc.

⚡ Quick Start

DriftWatch v0.4.0 introduces DriftSuite for unified monitoring:

from driftwatch import DriftSuite, DriftType
import pandas as pd

# 1. Initialize suite with reference data (e.g., training set)
suite = DriftSuite(
    reference_data=X_train,
    reference_predictions=y_val_pred,
    task="classification",  # or "regression"
    model_version="v1.0"
)

# 2. Check production batch
report = suite.check(
    production_data=X_prod,
    production_predictions=y_prod_pred
)

# 3. Act on specific drift types
drift_types = report.drift_types_detected()

if DriftType.CONCEPT in drift_types:
    print("🚨 CRITICAL: Concept drift detected — Retrain model!")
elif DriftType.PREDICTION in drift_types:
    print("⚠️ WARNING: Prediction drift — Check model outputs.")
elif DriftType.FEATURE in drift_types:
    print(f"📊 INFO: Feature drift in {report.feature_report.drifted_features()}")
else:
    print("✅ All systems normal.")

🛠️ Usage Scenarios

Scenario	Solution
Unified Monitoring	Use DriftSuite to track Feature, Prediction, and Concept drift in one go.
Experiment Tracking	Log all drift metrics to MLflow for long-term trend analysis.
Real-time API	Use DriftMiddleware in FastAPI to monitor every request.
Alerting	Send critical alerts via Slack or Email when model performance degrades.
CI/CD	Block deployments if DriftType.PREDICTION is detected in staging.

📓 Interactive Tutorials

• Multi-Drift Tutorial — Step-by-step guide to Feature, Prediction, and Concept drift.
• Complete Showcase — Tour of all detectors, visualizers, and integrations.

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

1. Fork the repo.
2. Install dev dependencies: pip install -e ".[dev,all]"
3. Run tests: pytest
4. Submit a PR!

📄 License

MIT © Vincent Cotella