Scikit_learn seems to be shipping a joblib distribution, which causes the usual overhead and confusion associated with bundling. Gentoo - for instance - has been editing the sources of scikit_learn to remove all imports of scikit_learn's joblib and instead import the system joblib for a long time now, and the software runs and passes tests successfully.

If scikits_learn may need a specific joblib version at some point, that can be more appropriately handled by restricting the dependency via a minimum or maximum version.

As part of brainhack2017 in Zurich I would be happy to help by either (in my order of preference):

• removing the shipped joblib and its associated code entirely,
• preferentially importing the system joblib, and adding except statements that default to the bundled joblib import,
• preferentially importing the bundled joblib, and adding an except statement defaulting to the system joblib. This latter suggestion simply makes it easier to maintain unbundled distributions of scikit_learn - by only having to not install sklearn/externals/joblib.

Could you please let me know what contribution addressing this issue would be best received?