The ML Code Fraction

Only a tiny fraction of real-world ML systems is composed of actual ML code. The required surrounding infrastructure is vast and complex.

A mature system might be at most 5% machine learning code and at least 95% supporting infrastructure: data collection, feature extraction, serving, monitoring, configuration management, and the glue code connecting it all.

This ratio has implications for where teams spend their time, where bugs live, and what skills matter in production ML work. The model is the smallest part of the system.

Related:, 05-molecule—ml-system-anti-patterns