Parsnip's "Model already registered" error prevents attaching packages with custom models to fit_resamples and tune_grid

[chillerb]

Hello, it's me again! 🥳

So, here is the scenario: I have a package that registers custom parsnip models .onLoad, and also exports a custom yardstick metric for evaluation as S3method.

Now, I want to call:

library(tidymodels)
library(my_pkg)

plan(multisession)

...

cv_res <- fit_resamples(my_model, y ~ x, cv_splits, metrics = my_metric)

However, this fails, because my custom metric my_metric isn't available in the workers' environments.

Ok, so what about explicitly attaching my package then?

library(tidymodels)
library(my_pkg)

plan(multisession)

...

cv_res <- fit_resamples(my_model, y ~ x, cv_splits, metrics = my_metric, control = control_resamples(pkgs = "my_pkg"))

This fails too - this time, because on load, my_pkgs tries to register the custom parsnip model again. However, the workers apparently receive the hidden parsnip registry object, causing parsnip to throw a "Model already exists" error.

Is there is a good solution to this problem at the moment?
parsnip:::check_model_doesnt_exist is not exported, so the obvious route of checking in .onLoad is not available.
I think, for now, I will just wrap the set_new_model calls with try?