Preface

Data-driven decision-making has been the key to the success of any business, and Machine Learning (ML) plays a key role in achieving that and helping businesses stay ahead of the competition. Though ML helps in unlocking the true potential of a business, there are many obstacles along the way. According to a study, 90 percent of ML models never make it to production. The disconnect between model development and productionization as well as bad or mediocre ML practices are a few of the many reasons for this. This is why there are so many end-to-end ML platforms offering to make ML development easy. One of the primary goals of these platforms is to encourage data scientists/ML engineers to follow Machine Learning Operations (MLOps) standards that help in the faster productionization of a model. In recent years, feature management has been one of the aims of the ML platform – whether it is built in-house or offered as a Platform as a Service (PaaS). A feature store that provides the ability to create, share, and discover curated ML features has become an integral part of most of these ML platforms.

The aim of this book is to discuss the significance of a feature store in ML pipelines. Hence, we will start with an ML problem and try to develop a model without a feature store. We will then discuss what aspects of ML can benefit from a feature store and how a few capabilities of feature stores not only help in creating better ML practices but also help in the faster and more cost-effective development of the model. As we move from why we should use a feature store to the what? and how? aspects of it, we will go through feature engineering, model training, inference, and also productionization of batch and online models with practical examples. In the first and second sections of the book, we will use an open source feature store, Feast. In the last section, we will look for alternatives that are available on the market and also try out an end-to-end use case with a managed feature store.