canary-ml 1.2.6

Drift and anomaly monitoring for production scikit-learn and Keras/TensorFlow models.

canary-ml

Drop-in drift and anomaly monitoring for production ML models.

One line wraps your model. Every .predict() call logs drift metrics, detects anomalies, and fires an alert when something shifts. Monitoring runs in a background thread — your inference latency is unaffected. No infrastructure required.

Project page · Guide & manual · Live demo

---

Install

pip install canary-ml

Requires Python 3.9–3.12. Dependencies: numpy, scipy, scikit-learn, rich.

Keras/TensorFlow model monitoring also requires TensorFlow (Python 3.9–3.12 only):

pip install canary-ml[keras]

---

Quickstart

from canary_ml import ModelMonitor

monitor = ModelMonitor(
    model=your_model,           # any sklearn-compatible model
    reference_data=X_train,     # baseline distribution
    alert_threshold=0.2,        # PSI threshold for alerts
    log_path="./canary_logs",
    verbose=True,
)

# Drop-in replacement — monitoring runs in the background
predictions = monitor.predict(X_new)
monitor.wait()  # block until background thread finishes

# Inspect the latest report
report = monitor.get_report()
print(report.summary())
# DriftReport | psi=0.41 | features_drifted=3/8 | anomaly_rate=3.2% | ALERT

# Launch the live dashboard
monitor.serve_dashboard(port=8501)
# → http://localhost:8501

---

What it monitors

• PSI — global distribution shift. < 0.1 stable · 0.1–0.2 moderate · > 0.2 alert. Requires ≥ 200 samples per batch; use drift_detected (KS-based) for smaller batches.
• KS test — per-feature Kolmogorov-Smirnov (continuous features). Drift flags are Benjamini-Hochberg corrected across all features in the batch, so the false-positive rate stays controlled at scale instead of compounding per feature.
• Chi² test — per-feature chi-squared (categorical features, ≤ 20 unique values).
• Anomaly detection — ensemble of Isolation Forest + z-score (|z| > 3).
• Confidence estimate — label-free accuracy proxy from predicted probabilities. Accurate when probabilities are well-calibrated; overestimates if the model is overconfident.

---

Alert callback

def my_alert(report):
    send_slack(f"Drift alert: PSI={report.psi_score:.2f}")

monitor = ModelMonitor(..., on_alert=my_alert)

---

Dashboard

monitor.serve_dashboard(port=8501)

Stdlib HTTP server, no extra dependencies. Auto-refreshes every 5 seconds. Can also run standalone:

python -m canary_ml.server ./canary_logs 8501

---

API reference

ModelMonitor

ModelMonitor(
    model,                      # sklearn-compatible model with .predict()
    reference_data,             # np.ndarray or pd.DataFrame, shape (n, features)
    alert_threshold=0.2,        # PSI threshold for drift alert
    performance_threshold=0.05, # accuracy drop (pp) below reference that fires a perf alert
    anomaly_contamination=0.05, # expected fraction of anomalies; alert fires at 4×
    categorical_threshold=20,   # max unique values for a feature to be treated as categorical
    store_samples=True,         # set False to skip storing raw feature rows (recommended in PII-sensitive envs)
    log_path="./canary_logs",
    verbose=True,               # default True — set False to suppress console output
    on_alert=None,              # callable(DriftReport) fired on alert
)

Method	Returns	Description
.predict(X)	same as model	Runs model; monitoring queued in background thread
.predict_proba(X)	same as model	Passthrough to model.predict_proba(); also feeds the confidence estimate
.wait()	—	Block until background monitoring tasks complete
.get_report()	DriftReport | None	Latest monitoring report
.get_history(n=50)	list[dict]	Last n raw log entries
.reset_baseline(new_data)	—	Replace the reference distribution and refit the anomaly detector
.serve_dashboard(port=8501, host="127.0.0.1")	—	Starts dashboard server in background thread. Use host="0.0.0.0" to expose beyond localhost

DriftReport

Attribute	Type	Description
psi_score	float	Global PSI vs reference
drift_detected	bool	True if any feature's KS/chi² p < 0.05 (soft warning)
ks_results	dict	Per-feature {statistic, p_value, drifted}
features_drifted	int	Count of features with p < 0.05 (computed property)
anomaly_rate	float	Fraction of samples flagged as anomalies
alert_triggered	bool	True if PSI > threshold, anomaly rate is high, or performance drops
alert_reasons	list[str]	Which conditions fired: "drift", "anomaly", "performance"
estimated_accuracy	float | None	Confidence estimate; None if no predict_proba
reference_accuracy	float | None	Confidence estimate on reference data
performance_delta	float | None	estimated_accuracy − reference_accuracy
output_ks	dict|None	KS test on prediction distribution vs. reference; None if unavailable
performance_alert	bool	True if delta < −performance_threshold
timestamp	str	ISO 8601

DriftReport is not directly JSON-serialisable. Use report.to_dict() for logging or json.dumps(report.to_dict()). Dict-style access (report["psi_score"]) is also supported.

---

License

MIT © Aitor Bazo