Coming from scikit-learn/scikit-learn#11230: .pkl is confusing: unless there are no numpy arrays in the joblib.dumped object, the file you write to can not be read with pickle.load.

Maybe using .jbl would be fine, better suggestions welcome!

Looking at git grep -l '\.pkl', we probably want to update at least doc/persistence.rst and examples/compressors_comparison.py. Not sure about other places, we probably don't want to change output.pkl to output.jbl in Memory ...