Tidymodels in prodution

Machine Learning and Modeling
1

I have kind of a "philosophical" question.
I have been reading a bit about using R for production and got a little worried.
I love tidymodels, but i have never deployed a model, but I hope to do so soon.

However, problems mentioned in this post Don’t use R in production. But if you do, here is how! | by Christian Freischlag | Towards Data Science got me worried, specially the licencing part. Is it possible to use tidymodels in the company where I work without having to share the code?

Also, would like to know the opinion in general of anyone who has used tidymodels in production, or if posit has any product/recomendation for that situation.

2

First, I'll say that you can deploy in any way that works for you. That said, here are some things:

  • You may want to trim down the model object to what is only essential for prediction. The butcher package can do this for you. ]
  • If you are deploying a model built with xgboost, lightgbm, keras, or a few others, you should know that using save() doesn't grab all of the information that is needed by the model (due to where those packages stash data in memory). The bundle package will fix this.
  • You might be able to deploy the model in pure SQL code. The orbital package can convert the prediction equation for some model objects to SQL. There was a presentation on this at the 2024 posit conf and a recent blog post.

Finally, there is a package that does one thing(deployment) and does it well for both R and python: vetiver. I would definitely recommend that.

1 Like