Many people come to machine learning by training a model on a tidy dataset, and then meet a harder problem: making that model work for real users, on fresh data, every day. Jim Dowling’s O’Reilly book, Building Machine Learning Systems with a Feature Store, is written for that moment. Dowling, CEO of Hopsworks, based the book on a course he taught at KTH in Stockholm, so it reads like a guided walk through building real systems.
The heart of the book is a simple way of thinking about any AI system. You break it into three parts. One part prepares the data into the useful inputs a model needs, which the book calls features. A second part uses those inputs to train the model. A third part puts the model to work, making predictions for whatever your application needs. The feature store of the title is the shared storage in the middle that lets the three parts connect. Seen this way, each part is small enough to build and test on its own.
The book sorts AI systems into three kinds, and its examples grow out of that scheme. A batch system makes predictions on a schedule, shown through an air quality forecaster you point at a sensor near your home. A real-time system responds the instant a request arrives, and the book builds two: a credit card fraud detector, and a video recommender modeled on TikTok that ranks clips from your recent activity. Agentic capabilities arrive as a further layer, where a large language model uses live data and tools to pursue a goal. The book adds that layer on top of the air quality service and the recommender, using agents built with LlamaIndex. Seeing the same building blocks reused across systems is where the method sinks in.
The examples are all in Python, using free and open source tools throughout, so a reader can follow along without buying anything. Each chapter ends with exercises, and the projects run on free tiers of common cloud services. Some comfort with Python and basic SQL will carry you through. Beyond that, it assumes you want to build, and hands you the scaffolding to start.
A few ideas give the book its own flavor. The book is careful about where each kind of data preparation belongs, so the same work runs the same way in training and in live predictions, which spares beginners a subtle class of bugs. It also questions habits newcomers pick up too early, pointing them toward a lighter path to a first running system. The message throughout is to get something working, then improve it.
One thing worth knowing going in: Dowling helped build Hopsworks, the platform the examples run on, and the book notes a collaboration between O’Reilly and Hopsworks. The lessons carry over regardless. Because the tools underneath are standard open source, a reader can carry the same methods to other platforms once the core ideas settle.
The last chapter is where it all comes together. It builds the TikTok-style recommender from end to end, then lists common mistakes that keep AI projects from reaching users, with a short word on ethics.
For someone getting started, the appeal is a repeatable method that turns a pile of scattered tricks into a single way of working. The theory stays light and the building stays central. Read it, work the exercises, and you come away able to take an idea and stand up a system of your own.
