I have been flirting with the idea of doing less meetups a month. I asked the question in slack and would love to hear what you're ideal number of meetups would be?
Also, if you were forwarded this email, consider subscribing here
Past Meetup
What MLOps Taught Me
MLOps Learnings
Ewan Nicolson gave us his top learnings over the past year and a half since diving into MLOps. One of the key takeaways from my side was how MLOps has given all the different teams working in ML a common vernacular. Nicolson went a step deeper to explain how this common language has helped him in different situations to destroy the isolation teams may feel and even avoid the whole "lets throw it over the fence" mentality.
With a plethora of
experience to draw from this talk and live coding session with Ewan is highly recommended. Check out the GitHub repo he put together so you can follow along with him in the video.
Reading Club + Alexey's Zoom Course Reading group is back! Last Friday the crew met once again and discussed this quick read on how to Avoid ML Pitfalls.
For some context, the reading group meets every other week on Friday. You can see all the info about the next sessions here or in the slack channel #reading-group.
Han-chung Lee who attended the session mentioned the paper is a good introductory for ML research. He then went on to break it down section by section.
Section 2: Don't look at all your data. Hide the testing set for final model evaluation well, not include them in EDA. And translating to production/industry practices, partition data by time so there's no leakage.
Section 3: pay attention to data leakage of different features, e.g., time lagged features, aggregation features, etc. and don't overfit towards the final model evaluation testing set.
Section 4: in addition for using the correct statistical metrics, for industry usage, translate the model impact to KPIs for stakeholders
Section 5: A/B test your models and monitor not just statistics/compute, but also KPI of the model for industry applications.
Zoom Bootcamp In other news, my arch-nemesis Alexey Grigorev, the author of Machine Learning Bookcamp and creator of Data Talks Club, is running a free online course based on his book. It covers topics from linear and logistic regressions to model deployment with AWS Lambda and Kubeflow. The course starts in September. Hit the link below to register!
Container Homes for Models For all those who were looking to Dockerize yo ML, look no further. Tomorrow we will be talking to the one who had the idea for this whole MLOps community, Luke Marsden.
You may know him as my old boss, or that guy from the MLOps Stacks channel, but he is so much more than that. Luke is also a passionate technology leader. He has a proven ability to conceive and execute a product vision from strategy to implementation while iterating on product-market fit.
Luke has a deep understanding of AI/ML, infrastructure software and systems programming, containers, microservices, storage, networking, distributed systems, DevOps, MLOps, and CI/CD workflows. He also left me without a job so don't apply if he is hiring 😆
In this session, expect a little live coding as Luke takes us through building ML models into container images so that you can run them in
production for inference.
There are various questions around doing this: Who should build the images and when? What should they contain? How should data science & ML teams interact with DevOps teams? If you build images specific to one platform, will you get locked in? If you try to build your containers inside a container, what happens, and why is this a security challenge?
Based on Luke's experience setting up ML container builds for many clients, he'll propose a set of best practices for ensuring secure, multi-tenant image builds that avoid lock-in, and he'll also share some tooling (chassis.ml) and a standard (openmodel.ml) for doing this.
See you tomorrow at the usual time 9am PST/5pm BST. Pro-tip, we have a public cal you can subscribe to.
Advanced Git: Community member Alexandra Johnson wrote a great post about how to use Git in more advanced ways in a practical, product-oriented
setting. Lots of awesome tips.
Nubank ML Monitoring: This one will definitely make it to Mega-Ops newsletter! The amazing ML group down at Nubank shared 9 tips
they learned about ML model monitoring.
MLOps hilarity flowing: Community member Aditya Soni shared a hilarious meme about all the "-flow" tools out there, which led
to even more hilarity. Check it out for a good time.