Implementing MLOps in the enterprise
AI at large
Why have MLOps when you have DevOps (or not)?
I was lucky to face the inevitability to master DevOps while deploying some of my server projects, mostly using Docker and Kubernetes and Linux. MLOps is here to shatter my sense of accomplishments with a whole new set of concepts and pipelines that I have to master, again, double luck!
I was not the first to jump on the Machine Learning train, with all the bells and whistles any new technology promises the deception wagon is never too far behind, if not a total wreckage in bubbles bursting and investments disappearing à la volée. Yet Machine Learning might be this unicorn that was not promised, yet delivered what we were not expecting for another 10 years - I'm talking for myself here, but Google's co-founder Sergey Brin didn't see it either.
Most of the tech industry got to adapt and learn quickly how to use, if not tame the new cool kid in the tech town, start-ups started flourishing on every LinkedIn post, every one and their deep pocket VC or ever-supporting aunt has an AI project that will put to rest an entire work sector.
Yet the challenges that one can face when building a Machine learning centric application are much different from the way a Fortune 500 company implements Machine Learning. The term MLOps was coined à la suite of DevOps - or in opposition, depending on how your resume gets rejected when it's missing the magic, empty keywords.
This production-first focused book by Yaron Haviv & Noah Gift showcases the ecosystem and the best practices for a production-grade, robust and flexible MLOps pipeline, one that does not get obsolete a few months down the line of the early enthusiasm. As the authors quote the entrepreneur Scott Shane: "it takes 43 startups to end up with just one company that employs anyone other than the founder after ten years.", this book is a primer and a handbook for that one man and his colleagues.
A few questions this book tries to give an answer to
- Why is it important to have a clear, robust and flexible MLOps strategy?
- How do you build an ML pipeline that continuously collects data, prepares it, engineers features, trains it, delivers it, monitors it?
- When should you use a pretrained model?
Who should read this book?
- Any MLOps engineer obviously
- Any DevOps that wants to transition to MLOps, albeit it does not clearly
- Any CTO, Engineering VP in a company eyeing an in house Macine learning usage
- Any data scientist interested in the way the models are actually delivered to the end user
Takeaways
- To be a good MLOps you have to be a good DevOps first
- MLOps is hard, do not overlook the strategic benefits of using a Pretrained model
- Got started with my first MLOps project on HuggingFace 🎉
- I am not a data scientist neither do I plan to be. This book served me as a primer and a handbook for foreign concepts, even for a seasoned software developer / devops. Albeit it is not the best introduction to the whole field, it does provide a pragmatic approach to actually doing enterprise MLOps.
Bits
