Integration with MLFlow?

[harrytruckit]

Hi guys,

I'm fairly new to MLOps and I'm trying to gather some opinions on the matter. I've come across this library and I really like its features. In addition to the cool data cleaning/preprocessing features its offering, is it attempting to compete with frameworks like MLFlow for experiment tracking, etc?

Thanks!

[rcap107]

Hey @harrytruckit, thanks for checking out skrub!

No, skrub isn't trying to compete with MLFlow for experiment tracking. The idea behind Data Ops is recording your operations so that they can be repeated reliably on new data. From my understanding of MLFlow, it's a tool for tracking experimental results; there is some of that in skrub, but it's mostly for inspecting the hyperparameter results in a param search. Skrub is really more about "putting all your operations into a computational graph, and use that to make sure that they are executed in the same way in potentially different environments".

I think skrub may complement MLFlow, with MLFlow being used for tracking Data Ops learners. I am not completely sure the two libraries are fully compatible at the moment, but it should be possible to connect the two so that learners can be tracked properly with MLFlow.

[harrytruckit]

Thank you for clearing that up!

[GaelVaroquaux]

No, skrub isn't trying to compete with MLFlow for experiment tracking.

Yes. One day, we hope that there might be some integration

[harrytruckit]

From my understanding, would Skore already be trying to do this? Additionally, SkoreHub seems to be trying to create a team friendly way of collaborating on data science projects.

[GaelVaroquaux]

Skore is doing some experiment tracking, in a way that is complementary to MLops. It does not have right integration with skrub, but will if there is interest (it's always a question of prioritizing thing), and integration with MLflow is planned