etsi-watchdog 0.2.0

Real-time drift detection for ML pipelines using PSI

etsi-watchdog

Real-time Drift Detection for Machine Learning Pipelines

etsi-watchdog is a lightweight, pluggable Python library that detects distribution shifts between training and production data. It supports both numeric and categorical features using Population Stability Index (PSI), and allows users to trigger custom actions on drift detection.

---

Installation

Install the latest version:

pip install etsi-watchdog

---

Features

Supports numeric and categorical features

Uses PSI (Population Stability Index) for drift detection

Safe binning and smoothing to avoid division errors

Per-column drift scoring

Customizable hooks via .on_drift() to trigger actions

Simple and extensible API

---

Quick Start

from etsi.watchdog import DriftMonitor
import pandas as pd

# Reference (training) data
X_ref = pd.read_csv("train.csv")

# Incoming (production) batch
X_live = pd.read_csv("production.csv")

# Initialize and monitor
monitor = DriftMonitor(reference=X_ref)

# Run check and handle drift
monitor.watch(X_live, threshold=0.2).on_drift(lambda: print("Drift detected"))

---

Use Cases

Monitor batch data for concept drift

Trigger model retraining workflows

Add drift alerts to production pipelines

Validate incoming data quality in staging or CI pipelines