The test_logistic::test_dtype_match fails on 32 bit Windows as was reported in MacPython/scikit-learn-wheels#7 (comment) The implication is that for LogisticRegression, float32 and float64 input may result in somewhat different models depending on the processor architecture (and OS), which is troublesome.

This applies to Python 3.5, 3.6 but not 2.7, 3.4 and 3.7 (cf https://ci.appveyor.com/project/sklearn-wheels/scikit-learn-wheels/build/1.0.66). Linux and Mac OS are fine, and I can confirm that I can't reproduce this with a Debian unstable 32 bit Docker image.

A workaround was applied #11899 consisting in raising tolerance to 1e-2 (from 1e-6) on 32 bit Windows, but identifying the root cause of this problem would be preferable.