Ask me anything with the Fiddler team is happening next Tuesday in the slack channel creatively named #ask-me-anything! Stay up to date with what's going on in the community by subscribing to our public cal.
Coffee Session
Fast Fast
It took many moons of scheduling and rescheduling to talk with the famous creator of Fast API Sebastián Ramirez. Our talk looks at what FastAPI is, how Sebastián built it, what the next big problems to tackle in ML are, and how to focus on adding value where you can.
What is Fast API? - A python framework for building web API's.It is built around all the latest features of the python language, including async support and type annotations.Many of you may already be familiar with the API that relies heavily on Open API. In fact, it seems as more time passes Fast API is becoming the go to framework and increasingly more popular than flask API's.
Why all the love? The popularity of Fast API was beyond what Sebastián ever imagined. "I thought it would be a tool I created and 50 hipsters used".
So why has it been gaining traction? Some of its features are really really nice. It's Fast, Async, easy to code, has data validation thanks to pydantic, and the kicker, automatic docs!
Fast API + Pydantic - One of Sebastians biggest suprises while building Fast API was seeing the code inside of Pydantic. Pydantic is a useful library for data parsing and validation, and Sebastián tells us stories of how things got complex rather quickly as he began intertwining the two.
Sebastián mentions how he solves problems (or trys not to) and why he feels it's important to continue using Fast API and not just building it. Let us know what you thought of it and be sure to leave a like or follow.
If the picture above doesn't say enough, lemme break you off with some hot news
from David Aponte, Harry Berg and the System Design Review crew.
The Challenge - A few years ago Uber set out to create an ads platform for the Uber Eats app that relied heavily on three pillars; Speed, Reliability, and Accuracy. Some of the technical challenges they were faced with included exactly once semantics in real time. To accomplish this goal, they created the architecture diagram above with lots of love from Flink, Kafka, Hive and Pinot. You can dig into the whole paper here to see all the reasoning for their design decisions.
One last thing - It takes us a considerable amount of time and energy to create these byte-size
summary videos. We want to make them as valuable as possible to you all. Please reply and let us know how we can make them better.
This past week we had two new posts show up on our community blog.
Feature Stores for Real-Time ML - Real-time AI/ML is on the rise and feature
stores are key to successfully deploying them. See how the choice of online store and the feature store architecture play important roles in determining its performance and cost.
Putting Together a Continuous ML Stack- Due to the increased usage of ML-based products within organizations, a new CI/CD-like paradigm is on the rise. On top of testing your code, building a package, and continuously deploying it, we must now incorporate CT (continuous training) that can be stochastically triggered by events and data and not necessarily dependent on time-scheduled triggers.
This post shows how fast and easy it is to set up a robust training-serving pipeline that will execute automatically based on production data and ongoing events. Notebook and repo included.
This was the most versatile meetup we have had to date. Haytham the CTO of Union.ai walked us through workflows and tasks in flyte. Neils then gave a quick overview of the most recent engineering labs then the teams showcased their stuff.
Speaking of the Engineering Labs - Over the past 6 weeks, we had teams ferociously competing to win the latest edition of our kaggle-esque Competition. The tool of honor for this last iteration was Flyte. Here
are three finishing teams and what they created during the last 6 weeks.
Brave Hyenas - This dynamic duo created the 'Brave Music'
project a music classification engine. Check out the github here. (Apparently they call what I listen to K-POP)
Adorable Unicorns - The team from Artefact created the 'what's cooking good looking' project - a name entity recognition pipeline. Have a gander at the github project here.
Vamos Dalhe - The winning team coming from Brasil created the 'destination similarity' project which was a recommender system that suggested other places you might want to visit if you
were looking at vacationing in certain areas. Full github project here.
You know the old iceberg analogy, where the larger portion is hidden under the surface? Well, most of us in MLOps tend to focus on the visible, the models we need to deploy and run in production.
But if we ignore resource management as our AI/ML initiatives grow, we’ll start to take on water, in the form of researchers fighting for resources, time-consuming manual workload rescheduling, and spiraling costs associated with ML inference.
In this talk, the run.ai team will show us what role resource management has in MLOps, what to strive for, and how to get buy-in from IT.