Building Machine Learning Systems with a Feature Store: Batch, Real-Time and LLM (www.oreilly.com)

Jim Dowling’s new book, Building Machine Learning Systems with a Feature Store: Batch, Real-Time and LLM (Nov 2025), presents a unified, practical approach for designing, operating, and scaling ML systems that combine batch data, real-time streams, and large language models. Aimed at intermediate-to-advanced practitioners, the 492‑page guide centers on a shared data layer—the feature store—and modular, independent ML pipelines as the architectural foundation. Dowling, CEO of Hopsworks and an associate professor at KTH, frames the feature store as the solution to “the hardest problem in ML”: making data consistent, discoverable, and usable across training, fine-tuning, and production serving. Technically, the book walks through MLOps principles for reliability and reproducibility, and shows how feature stores support incremental datasets for training and fine-tuning, low‑latency online access for serving, and retrieval-augmented generation (RAG) workflows for LLM-powered applications. The implications are practical: consistent feature definitions across training/serving, simplified lineage and governance, easier orchestration of batch and stream pipelines, and a clear pattern for integrating model fine-tuning and real-time retrieval into production. For teams building production ML and LLM systems, the book offers concrete patterns and operational practices to reduce data friction and accelerate deployment.