Designing Machine Learning Systems
The book for people who think the model is the hard part. Most of the work is data, drift and interfaces.
A production view of machine learning: data engineering, feature stores, training pipelines, deployment patterns, monitoring, and the organisational question of who owns a model after it ships.
The idea I took
Treating a model as a system whose inputs decay is the framing that stuck. It makes monitoring a first class requirement rather than an operational afterthought, and it gives a straightforward answer to what happens after go live, which is the question control functions actually want answered.
Where it stops
It is a systems book written before generative AI reshaped the tooling. The principles hold. Some of the specific stack advice has aged.