This book is licensed for the personal use of the buyer. Please do not share or resell the file. Short quotations with attribution are welcome.
The information in this book is provided for education. It is not legal, financial or other professional advice. Product and company names are used only to describe them and belong to their owners.
This is a free sample: the introduction and first chapter.
IntroductionIntroduction
Training a model that works in a notebook is the beginning, not the end. Most of the work, and most of the failures, in machine-learning systems happen afterwards: when the model must be reproduced, deployed, monitored, retrained and governed, month after month, as data and requirements change. MLOps is the engineering discipline that makes this work reliable.
This book covers that discipline end to end. It explains how to version data, code and models; build training pipelines; track experiments; package and serve models in batch, online and streaming settings; release them safely with shadow and canary deployments; monitor data, predictions and outcomes; detect drift; retrain; control costs; and document everything regulators and auditors increasingly expect. Separate chapters address operations for language-model applications, which bring their own challenges.
The book is tool-agnostic. It focuses on principles and practices that apply whatever platform you use, illustrated with failure stories that show why each practice exists.
How the book is organised
Chapters 1 to 8: why ML systems are different, the life cycle and maturity.
Chapters 9 to 18: reproducibility: versioning, pipelines, experiments, features and testing.
Chapters 19 to 29: deployment: packaging, serving patterns, release strategies and scaling.
Chapters 30 to 42: monitoring, drift, retraining, incidents and costs.
Chapters 43 to 49: language-model operations, data validation, and pipelines and monitoring in code.
Chapters 50 to 56: data contracts, online experiments, explainability, fairness monitoring, human review, sustainability and organisational practices.
Chapters 57 and 58: costs of language-model applications and choosing tools.
Chapters 59 to 64: case studies, a roadmap and a checklist.
Chapter 1Why ML Systems Are Different
Traditional software behaves according to its code. Machine-learning systems behave according to code and data. Change the data and behaviour changes, even if no line of code does.
Consequences:
Bugs can live in data, labels and features as well as code.
Performance degrades silently as the world changes.
Testing must cover data and model quality, not just logic.
Reproducing a result requires the exact data, code, configuration and environment.
You have reached the end of the free sample. The full book continues with 63 more chapters, plus the appendices.