How to leverage the ML tooling ecosystem

If you have hung around our slack questions and answers channel at all you undoubtedly have seen Mariya answer all of our toughest questions. This Wednesday in the meetup she will talk to us about a crucial part of MLOps... leveraging tooling.

MLOps is an extremely broad topic, thus one tool can't solve all its issues. Mariya and the Nue.ro team believes that the MLOps future is in integrating numerous specialized tools with each other. Basically, how to get everything working together nicely.

We will talk about examples of an ML project development and production workflows which are built on top of their proprietary core - Neu.ro - using a number of open-source and proprietary tools: Jupyter Notebooks, Tensorboard, FileBrowser, PyCharm Professional, Cookiecutter, Git, DVC, Airflow, Seldon, and Grafana. She will describe how to integrate each of these tools with Neu.ro, and how to improve these integrations. Sounds like fun right?

Lina was an excellent guest on the meetup explaining to us in depth about how many different bugs you can encounter when deploying models and how easy it is for these bugs to go unnoticed. She broke down each piece of the puzzle and highlighted where she found bugs along the way. To give us a bit of hope, she then explained a method of monitoring that she developed so you don't lose serious 💰 with a recommendation system that only spits out one recommendation.

Check out the video here and podcast here and here is a link to her slides. Btw, he explained how you can have bugs in every part of the image that is labeled below. 😅

After 1.5 hours of learning about dask there are a few key takeaways I can pass along.

First, if you don't need to use parallel processing, don't. It's an expensive way to solve a problem that you may be able to do by getting a better algorithm.

And it's python all the way down. You can get to a much lower level extraction than you can when using spark. So it's got that going for you.

Check out the full workshop recording here, and have a play with the exercises on the GitHub repo here. Also if you want to check out Dan's full 8hr course on Dask you can do so here