MLOps with Databricks

Preface

I have been working in Data and AI for over a decade, and I’ve devoted most of my career to machine learning operations (MLOps). I started doing it before MLOps became a formalized term in the late 2010s. Even though MLOps is an established term today, too few companies and teams follow proper MLOps principles.

In many teams I encounter today, there tends to be an imbalance in expertise. Some team members have deep experience in data science and know how to build and monitor machine learning (ML) models, but they are less familiar with development and operations (DevOps) practices. Others come from a strong software engineering background and follow DevOps principles rigorously, yet they often treat the ML or AI component as a black box.

Neither situation is inherently problematic, but both present challenges when the team’s goal is to build reliable ML/AI systems that enable other teams or the organization to continuously deliver value at an accelerated pace. Success requires bridging the gap: acknowledging the nondeterministic nature of many ML/AI components while also applying the discipline of DevOps practices.

This book focuses on traditional ML. About a third of it is devoted to large language model operations (LLMOps), which is a specialized subset of MLOps that focuses on applications built around large language models and, more broadly, foundational and multimodal generative models.

In that sense, MLOps provides the foundation: the principles and practices for ...