Operations Python (ML) active http-servicebatch-job

Prediction telemetry records model identity

ops-python-ml-003

Intent

Make model-serving regressions diagnosable by recording which model produced each production prediction.

Applicability

Applies when the diff adds or changes online inference, batch scoring, or production prediction event publishing.

What to inspect

Prediction handlers, logging, event schemas, request or entity IDs, and model version fields.

Pass criteria

Prediction telemetry includes a stable request or entity identifier and the model version or artifact identifier used for that prediction.

Fail criteria

The changed prediction path emits no telemetry, or logs predictions without the model identity needed to trace behavior later.

Do not flag

Local experiments or systems that avoid raw feature logging for privacy reasons but still record request and model identity safely.

Confidence guidance

HIGH when the inference path clearly includes or omits model identity. MEDIUM when logging is routed through shared helpers. LOW when production use is unclear.

Remediation

Emit structured prediction telemetry with request or entity identity and model version.

Pass example

logger.info("prediction_made", extra={
    "request_id": request_id,
    "model_version": MODEL_VERSION,
})

Fail example

def predict(features):
    return model.predict_proba(features)[0, 1]