scikit-learn: Unstable LogisticRegression with saga solver l1 penalty under macOS

======================================================================
FAIL: sklearn.linear_model.tests.test_logistic.test_logreg_l1_sparse_data
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/nose/case.py", line 198, in runTest
    self.test(*self.arg)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/sklearn/linear_model/tests/test_logistic.py", line 961, in test_logreg_l1_sparse_data
    assert_array_almost_equal(lr_saga.coef_, lr_liblinear.coef_)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/numpy/testing/utils.py", line 962, in assert_array_almost_equal
    precision=decimal)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/numpy/testing/utils.py", line 778, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Arrays are not almost equal to 6 decimals
(mismatch 64.0%)
 x: array([[ -128.209824,  -990.760917,     0.      ,     0.      ,
            0.      ,   783.90525 ,  1345.642841,     0.      ,
           34.880916,  -287.813868,  -364.684306,  -668.47532 ,...
 y: array([[-0.764085, -0.29863 ,  0.      ,  0.269297,  0.      ,  0.      ,
         2.629204,  0.      ,  0.      ,  0.      , -0.450364, -0.61554 ,
         0.      , -0.262512,  0.161514,  0.      ,  0.314769,  0.      ,...
======================================================================
FAIL: sklearn.linear_model.tests.test_logistic.test_saga_vs_liblinear
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/nose/case.py", line 198, in runTest
    self.test(*self.arg)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/sklearn/linear_model/tests/test_logistic.py", line 1144, in test_saga_vs_liblinear
    assert_array_almost_equal(saga.coef_, liblinear.coef_, 3)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/numpy/testing/utils.py", line 962, in assert_array_almost_equal
    precision=decimal)
  File "/Users/travis/build/MacPython/scikit-learn-wheels/venv/lib/python3.6/site-packages/numpy/testing/utils.py", line 778, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Arrays are not almost equal to 3 decimals
(mismatch 100.0%)
 x: array([[ -105.601, -1199.326,  2377.574,   479.408]])
 y: array([[ 0.   , -1.06 ,  1.222,  0.   ]])

numpy and scipy are imported from the following wheels from PyPI:

• numpy-1.11.3-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (at buil time)

• numpy-1.13.1-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (at test time)

• scipy-0.19.1-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl

• Cython-0.25.2-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl

The full log of the build is here:

https://s3.amazonaws.com/archive.travis-ci.org/jobs/253231280/log.txt?X-Amz-Expires=30&X-Amz-Date=20170713T160513Z&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAJRYRXRSVGNKPKO5A/20170713/us-east-1/s3/aws4_request&X-Amz-SignedHeaders=host&X-Amz-Signature=06e6866cac7abfc10cc67aa061a0e45b8db234c71a5997c67d8cc63828838df2

The configuration scripts of this build and test run can be found here:

https://github.com/MacPython/scikit-learn-wheels

A similar problem happened with Python 2.7 on macOS:

https://travis-ci.org/MacPython/scikit-learn-wheels/jobs/253231277

but not with python 3.4 for some reason:

https://travis-ci.org/MacPython/scikit-learn-wheels/builds/253231260

The same tests pass under Linux and Windows both 64 and 32 bit for all supported versions of Python.

Maybe this is a problem caused by the Apple Accelerate implementation of BLAS.