This seems more sensible... but I'd like to hear whether there are reasons other than computational cost that others can find for this restriction existing in the first place.

In short, LGTM

weird I didn't get the error locally... must have not run the test properly..

I have no idea why I put _safe_split into utils.meta_estimator, but I don't really have a better idea now.

(I wonder whether kallable kernels were outlawed at some point when joblib didn't check for picklability and things crashed??)

for reference,

from sklearn.model_selection import cross_val_score
from sklearn.svm import SVC
from sklearn.datasets import make_blobs
import numpy as np

X, y = make_blobs()
cross_val_score(SVC(kernel=lambda x, y: np.inner(x, y)), X, y)

works, n_jobs=2 crashes with a pickling error - unsurprisingly.

Thanks for the exemple @amueller.
LGTM

Maybe we can add something about that in what's new ?

this is good for merge, right?