Ville Tuulos - Effective Data Science Infrastructure: How to make data scientists productive

I first met the author, Ville Tuulos, in 2012 when I was trying to understand the hype around Hadoop. At the time, Ville was working on Disco, an Erlang-based solution to map-reduce that made it easy to interact with Python. Peter Wang and I had just started Continuum Analytics Inc., and Villes work was a big part of the motivation for releasing Anaconda, our distribution of Python for Big Data.

As a founder of NumPy and Anaconda, Ive watched with interest as the explosion of ML Ops tools emerged over the past six to seven years in response to the incredible opportunities that machine learning presents. There are an incredible variety of choices and many marketing dollars are spent to convince you to choose one tool over another. My teams at Quansight and OpenTeams are constantly evaluating new tools and approaches to recommend to our customers.

It is comforting to have trusted people like Ville and the teams at Netflix and outerbounds.co that created and maintain Metaflow. I am excited by this book because it covers Metaflow in some detail and provides an excellent overview of why data infrastructure and machine learning operations are so important in a data-enriched world. Whatever MLOps framework you use, Im confident you will learn how to make your machine learning operations more efficient and productive by reading and referring to this book.

Travis Oliphant author of NumPy, founder of Anaconda, PyData, and NumFocus

As a teenager, I was deeply intrigued by artificial intelligence. I trained my first artificial neural network at 13. I hacked simple training algorithms in C and C++ from scratch, which was the only way to explore the field in the 1990s. I went on to study computer science, mathematics, and psychology to better understand the underpinnings of this sprawling topic. Often, the way machine learning (the term data science didnt exist yet) was applied seemed more like alchemy than real science or principled engineering.

My journey took me from academia to large companies and startups, where I kept building systems to support machine learning. I was heavily influenced by open source projects like Linux and the then-nascent Python data ecosystem, which provided packages like NumPy that made it massively easier to build high-performance code compared to C or C++. Besides the technical merits of open source, I observed how incredibly innovative, vibrant, and welcoming communities formed around these projects.

When I joined Netflix in 2017 with a mandate to build new machine learning infrastructure from scratch, I had three tenets in mind. First, we needed a principled understanding of the full stackdata science and machine learning needed to become a real engineering discipline, not alchemy. Second, I was convinced that Python was the right foundation for the new platform, both technically as well as due to its massive, inclusive community. Third, ultimately data science and machine learning are tools to be used by human beings. The sole purpose of a tool is to make its users more effective and, in success, provide a delightful user experience.

Tools are shaped by the culture that creates them. Netflixs culture was highly influential in shaping Metaflow, an open source tool that I started, which has since become a vibrant open source project. The evolutionary pressure at Netflix made sure that Metaflow, and our understanding of the full stack of data science, was grounded on the pragmatic needs of practicing data scientists.

Netflix grants a high degree of autonomy to its data scientists, who are typically not software engineers by training. This forced us to think carefully about all challenges, small and large, that data scientists face as they develop projects and eventually deploy them to production. Our understanding of the stack was also deeply influenced by top-notch engineering teams at Netflix who had been using cloud computing for over a decade, forming a massive body of knowledge about its strengths and weaknesses.

I wanted to write this book to share these experiences with the wider world. I have learned so much from open source communities, amazingly insightful and selfless individuals, and wicked smart data scientists that I feel obliged to try to give back. This book is surely not the end of my learning journey but merely a milestone. Hence, I would love to hear from you. Dont hesitate to reach out to me and share your experiences, ideas, and feedback!

This book wouldnt be possible without all the data scientists and engineers at Netflix and many other companies who have patiently explained their pain points, shared feedback, and allowed me to peek into their projects. Thank you! Keep the feedback coming.