The Flow Brothers

Federated machine learning promises to overcome emerging privacy challenges. Hence, algorithmic aspects of the topic have gained popularity in the scientific literature. However, fundamental aspects such as scalability, robustness, security, and performance in a geographically-distributed setting remain relatively unexplored.

At Scaleout Systems, they are developing an open-core platform for federated machine learning operations that aims at bridging the gap between the scientific literature and real-world deployments. The aim of this talk is to share challenges and experiences in their development journey.

Happening on June 22nd at 5pm BST/ 9am PST/ 12pm EST. Jump in!

Last week, we sat down with Jacopo Tagliabue, the director of AI at Coveo. Jacopo is involved with many different AI efforts in the ML space; you can look him up he’s pretty famous

You don’t need a bigger boat
Jacopo walked us through his GitHub repo, Post Modern Stack, a trimmed-down version of his original, You Don’t Need a Bigger Boat, which is an end-to-end MLOps stack for “reasonable scale” (terabytes) ML systems. These repos are not cookie cutters, but rather workable solutions that you can use to build ready-to-go ML systems for your team. Battle-tested by the best in class tools (and Jacopo himself). Jacopo was even kind enough to add a multi-series blog walking through his repos and how to implement them in production. You can read all about it here.

Enough talk, where’s the code? After about 5 slides, we jumped straight into the code to build a real e-commerce recommendation system using the largest real open-source e-commerce dataset available (made available by Coveo). Jacopo walked us step by step through the entire system, from loading the raw data to transformations to model training, deployment, and cleanup, running all of the code, and answering questions along the way.

Metaflow, dbt, snowflake, reclist, s3, comet and sagemaker oh my!
Starting with the raw JSON data, we loaded it into a snowflake and ran some simple queries to ensure baseline expectations. We then utilized DBT to transform that data into structured data ready for model consumption. After transformations, we trained a model and checked out our results in the comet.ml, deployed our model live to sagemaker, made a prediction, then took it down from sagemaker and cleaned up the pipeline.

Meta who? Metaflow was at the center of this system, acting as the pipeline orchestrator tying every step together. Metaflow is a very powerful pipeline tool and comes with some very ML-first features that can change the way you build pipelines

For anyone who missed it, the entire session is on our youtube channel

As the MLOps ecosystem continues to grow and evolve its adoption to emerging workflows and use cases, It was great to have Corey Zumar, Lead MLflow Maintainer on the pod last week, to walk us through the journey of how this evolution is influencing the development of MLfow.

What's MLflow? MLflow is an open-source platform developed by Databricks to help streamline development processes in the MLOps lifecycle

Build Key Insights Corey shared one of the most important lessons learned while building MlFlow, he says "Instead of purpose building solutions for every niche tool or use case based on users' requirements, it is often easier to create an extensible and coherent core medium(API) to let developers bring their own workflows with that."

Product development with open-source This approach creates a rich medium for the maturity of the product as it grows within the community, which is based on feedback. It increases the chances of building a product to solve actual challenges practitioners face, for their various use cases. And sometimes, a joint development structure where you have a third-party platform that adopts the use of a product could also power the Product's development, like in the case of MLflow with Databricks

The vision and success in Adoption The idea behind the product's design was to have a core abstraction of components(pillars) and structure, in such a way that enables seamless flexibility for users("libertarian engineers" ) and it looks like that turned out pretty well because it is quite popular amongst practitioners mainly because of how easy it is to set up and use, some might even consider it as the "gateway drug to MLOps”

What's next? We've definitely been thinking of adding more functionalities based on what members of the community have been asking for and also trying to support more phases of the Mlops lifecycle, but what is most important right now is achieving a high level of compatibility with industry-standard tools. However, you should be expecting something new in MLflow within a few weeks.

Machine learning operations (MLOps), is becoming an exciting space as we figure out the best practices and technologies to deploy machine learning models in the real world. Keeping us with what's new can be a lot this Primer highlights a few available resources to upskill and inform yourself on the latest in the world of MLOps. A few educational resources have been listed as a start of a more comprehensive guide in the future.

MLOps enable ML teams to build responsible and scalable machine learning systems and infrastructure. This facilitates tasks that range from risk assessment to building and testing to monitoring. While still in its infancy, MLOps has attracted machine learning engineers and software engineers in general. With every new paradigm comes new challenges and opportunities to learn.

Flyte is a workflow automation platform for mission-critical machine learning and data processing at scale, It enables users to operate and maintain highly concurrent workflows. Conceived at Lyft, Flyte was open-sourced in early 2021 under LF AI & Data, and was recently promoted to a Graduate Project in January 2022.

Companies that have adopted Flyte report a "sweet spot" where the platform seamlessly fits into their ML infrastructure, drastically reducing processing time and enabling scalability while requiring minimal integration efforts from the data science teams. Flyte has been named one of the Top 10 Python Packages for MLOps this year.

To learn more about Flyte:
Watch customer success stories on YouTube
Join us on Slack, follow us on Twitter, and attend community sync events on Zoom
Subscribe to Flyte Monthly for the latest news, updates, and events
Check out the Flyte repo on Github

It’s no secret that artificial intelligence (AI) and machine learning (ML) are used by modern companies for countless use cases where data-driven insights may benefit users.

What often does remain a secret is how ML algorithms arrive at their recommendations. If asked to explain why a ML model produces a certain outcome, most organizations would be hard-pressed to provide an answer. Frequently, data goes into a model, results come out, and what happens in between is best categorized as a "black box".

This inability to explain AI and ML will soon become a huge headache for companies. New regulations are in the works in the U.S. and the European Union (EU) that focus on demystifying algorithms and protecting individuals from bias in AI.

The good news is that there’s still time to prepare. The key steps are to understand what the regulations include, know what actions should be taken to ensure compliance, and empower your organization to act now and build responsible AI solutions.