The Godfather Of MLOps

Today A Reader Tomorrow A Leader

Shout out to Charlie You for rallying 22 ML practitioners and students to joined to discuss the Google article "Rules of ML". As community member Daniel Cooper said

"Eugene made a really great point that MLOps is arguably in the early adopter phase and while a lot of us recognize the need for this stuff... it’s not clear that there is market demand for some of the solutions / practices. My counterpoint was that the reason this might be true is because so many companies are not successfully integrating ML into production, so they haven’t discovered things like drift... but it’s likely because they haven’t put the right infrastructure / processes in place (MLOps) in order to get models deployed and build confidence in them (chicken vs. egg) kinda problem."

Tons of insightful points brought up during the discussion that lasted >20m later than was scheduled. I encourage anyone who likes to learn with others to jump in this group! We meet bi-weekly so the next one will be next week!

Like Chatting With an Old Friend

Daniel was a great sport last Wednesday when he came through the meetup and talked to us about how his team has been able to operationalize 100% of the data products they wanted to get out into the wild!

The Primitives: One of the insights was extracting primitives of the data science process and making DevOps work with that. The three primitives that make a successful data science product regardless of the algorithm, regardless of the architecture in Daniel's view are:

Training Pipeline: Takes feature engineering code and outputs as part of the pipeline a model object.
Scoring Pipeline: Takes feature engineering code as well as model objects as inputs and outputs a prediction.
Monitoring Pipeline: Takes label code from training pipeline as well as the predictions from the scoring pipeline and compares the two along with the ground truth.

+ The way that Daniel has structured the ML systems is fascinating to hear about especially knowing the industry he works in is heavily regulated. Some quick tips I found useful using git, and github actions to trigger pipelines, and treating the monitoring pipeline as a first-class citizen (aka its super important)

+ Video here and podcast here.

The Godfather Of MLOps

If the name D. Sculley does not ring a bell that is because you, like me, have read his papers but never even realized it. Check out a list of all the papers he has contributed to here if you are still wondering who he is.

What's Changed? Right before we hit record I let D. know how grateful we were for him to come on the pod and share his wisdom with us especially considering he hasn't been doing many appearances as of late. He responded with this "I just got tired of talking about this stuff" That's right, the Godfather of MLOps since 2015 has been talking about the need for attention on the Ops side of things. Like an oracle, he revealed to us a glimpse of the future then stepped out of the limelight to move on to the next hard problem while we all continue to learn from his papers written all those years ago. So we asked him, "what's changed since he wrote these papers?" His answer, "well to start, now we have an MLOps community."

+This is easily the most influential person in MLOps that we have had the pleasure of interviewing. Aside from being an A+ human, he was a wealth of information. I cannot stress enough, go check out this interview.

Missing Link In ML Infra?

Learning MLOps for long enough you've undoubtedly come across the full stack deep learning course as it manages to give some shelter from the storm.

Meetup: Machine learning is quickly becoming a product engineering discipline. Although several new categories of infrastructure and tools have emerged to help teams turn their models into production systems, doing so is still extremely challenging for most companies. In our chat with Josh, we survey the tooling landscape and point out several parts of the machine learning lifecycle that are still underserved.

Josh will propose a new category of tool that could help alleviate these challenges and connect the fragmented production ML tooling ecosystem. Last but not least we will finish off the conversation by discussing similarities and differences between his proposed system and those of a few top companies.

Bio: Josh Tobin is the founder and CEO of a stealth machine learning startup. Previously, Josh worked as a deep learning & robotics researcher at OpenAI and as a management consultant at McKinsey.

He is also the creator of Full Stack Deep Learning, the first course focused on the emerging engineering discipline of production machine learning. Josh did his PhD in Computer Science at UC Berkeley advised by Pieter Abbeel.

+ Pre reading to get pumped up for this session can be found here on slides Josh made for the MLsys course

+ As always see you at 5pm GMT / 9am PST tomorrow, Wednesday by clicking the link below.