Noah Gift - Practical MLOps: Operationalizing Machine Learning Models

• 2021-09-14: First Release

Weve both spent most of our careers automating things. When we first met and Alfredo didnt know Python, Noah suggested automating one task per week. Automation is a core pillar for MLOps, DevOps, and this book throughout. You should take all the examples and opinions in this book in the context of future automation.

If Noah could summarize how he spent 20002020, it was automating just about anything he could, from film pipelines to software installation to machine learning pipelines. As an engineering manager and CTO at startups in the Bay Area, he built many data science teams from scratch. As a result, he saw many of the core problems in getting machine learning to production in the early stages of the AI/ML revolution.

Noah has been an adjunct professor at Duke, Northwestern, and UC Davis in the last several years, teaching topics that primarily focus on cloud computing, data science, and machine learning engineering. This teaching and work experience gives him a unique perspective about the issues involved in the real-world deployment of machine learning solutions.

Alfredo has a heavy ops background from his Systems Administrator days, with a similar passion for automation. It is not possible to build resilient infrastructure without push-button automation. There is nothing more gratifying when disaster situations happen than rerunning a script or a pipeline to re-create what crashed.

When COVID-19 hit, it accelerated a question we both had, which was, why arent we putting more models into production? Noah touched on some of these issues in an article he wrote for Forbes. The summarized premise of the article is that something is wrong with data science because organizations are not seeing returns on their investments.

Later at OReillys Foo Camp, Noah led a session on Why can we not be 10X faster at ML in production? where we had a great discussion with many people, including Tim OReilly, Mike Loukides, Roger Magoulas, and others. The result of that discussion was: Yes, we can go 10X faster. So thanks to Tim and Mike for stirring such a fascinating discussion and getting this book on its way.

Machine learning feels a lot like many other technologies that have appeared in the past several decades. At first, it takes years to get results. Steve Jobs talked about how NeXT wanted to make it 10X faster to build software (and he did). You can watch the interview on YouTube. What are some of the problems with machine learning currently ?

• Focus on the code and technical details versus the business problem

• Lack of automation

• HiPPO (Highest Paid Persons Opinions)

• Not cloud native

• Lack of urgency to solve solvable problems

Quoting one of the things Noah brought up in the discussion: Im anti-elitism across the board. Programming is a human right. The idea that there is some priesthood that is only allowed to do it is just wrong. Similar to machine learning, it is too crucial for technology to lie only in the hands of a select group of people. With MLOps and AutoML, these technologies can go into the publics hands. We can do better with machine learning and artificial intelligence by democratizing this technology. Real AI/ML practitioners ship models to production, and in the real future, people such as doctors, lawyers, mechanics, and teachers will use AI/ML to help them do their jobs.