Real-time data drift detection for machine learning pipelines.
etsi-watchdog is a production-ready Python library for drift detection, version comparison, and real-time monitoring of data streams. Designed for ML practitioners, data scientists, and AI engineers who need reliable data quality insights.
- PSI-based Drift Detection (more algorithms coming)
- Rolling Monitoring with time-frequency windowing
- Version Drift Comparison between model/data snapshots
- Built-in Visualization & JSON Export
- Minimal Dependencies & Fast Performance
- Clear API, suitable for both research and production
pip install etsi-watchdogfrom etsi.watchdog import DriftCheck
import pandas as pd
ref = pd.read_csv("reference.csv")
live = pd.read_csv("current.csv")
check = DriftCheck(ref)
results = check.run(live, features=["age", "salary"])
for feat, result in results.items():
print(result.summary())
result.plot()from etsi.watchdog import Monitor
monitor = Monitor(reference_df=ref)
monitor.enable_logging("logs/rolling_log.csv")
results = monitor.watch_rolling(
df=live_data_stream,
window=50,
freq="D",
features=["age", "salary"]
)from etsi.watchdog import DriftComparator
check = DriftCheck(ref)
v1 = check.run(live1, features=["age", "salary"])
v2 = check.run(live2, features=["age", "salary"])
comp = DriftComparator(v1, v2)
print(comp.diff())| Feature Area | Description |
|---|---|
| Multi-Algorithm Support | Support for additional drift metrics: - Jensen–Shannon Divergence (JSD) - Wasserstein Distance - Kolmogorov–Smirnov Test (K-S) - Tree-based drift (e.g., DeepDrift) |
| Plug-in Architecture | Drift algorithms will be fully plug-and-play. Custom metric support via register_drift_function() API. |
| Real-Time Stream Hooks | Support for Kafka/Redis/WebSockets to detect drift on live data streams. |
| Concept Drift Detection | Integration with models to detect label or concept drift, not just feature distribution shift. |
| CLI & YAML Configs | Full CLI support:etsi-watchdog detect --ref ref.csv --live live.csv --features age salary+ YAML-based configuration for automated pipelines. |
| Benchmark Suite | Built-in benchmarking with synthetic datasets to evaluate metric sensitivity and response time. |
| Dashboard UI (Optional) | Lightweight dashboard (Streamlit/FastAPI) for monitoring drift over time visually. |
| Drift Summary Reports | Generate PDF/HTML reports with drift summary, top features, histograms, and timestamps. |
| Sklearn & Pandas Integration | DriftCheck will support .fit()/.transform() methods like Scikit-learn transformers. |
Contributions are welcome! 🙌
If you have an idea for a drift metric, integration, or improvement, feel free to open an issue or pull request.
Please refer to CONTRIBUTING.md for guidelines.
Licensed under the BSD-2-Clause License.
See the LICENSE file for details.