I suspect we made a mistake in letting return_train_score default to True (see #9619 for example) in GridSearchCV, as it can sometimes be expensive to score over a training set.

I think it would be useful to users if we issue a warning (in _fit_and_score? at the end of fitting?) if train scoring is greater than ?10% of fit and test score time and is more than a few seconds for the entire grid search... or something.

It's hard to come up with a precise heuristic, and it's hard to know when to issue it.

The other option is to make return_train_score=True stop being the default...