Package: python-numpy Version: 1:1.12.1-2 On amd64:
$ python Python 2.7.13 (default, Jan 19 2017, 14:48:08) [GCC 6.3.0 20170118] on linux2 Type "help", "copyright", "credits" or "license" for more information. >>> import numpy >>> abs(numpy.nan) nan >>> numpy.abs(numpy.nan) nan >>> On arm64: $ python Python 2.7.13 (default, Jan 19 2017, 14:48:08) [GCC 6.3.0 20170118] on linux2 Type "help", "copyright", "credits" or "license" for more information. >>> import numpy >>> abs(numpy.nan) nan >>> numpy.abs(numpy.nan) __main__:1: RuntimeWarning: invalid value encountered in absolute nan >>> I do not see why the behaviour would intentionally be different on different architectures so I suspect that this is a bug. Unfortunately I could not discover where abs/absolute is implemented so the bug might not be in python-numpy itself. This inconsistency seems to be the reason for pandas failing to build on arm64 and mips64el: https://buildd.debian.org/status/package.php?p=pandas&suite=sid If there is a simple fix that does not change the behaviour on amd64 it would be nice to have this fixed in Stretch. (Would we then get pandas in Stretch on arm64?)
