Tracing in Production: When 1–150ms Turns Into 700ms

We run machine learning models in production serving real-time traffic. When our managed serving platform introduced built-in MLflow tracing, we wanted to enable it. Tracing promised visibility into i