In Debian (only change from -1 should be disabling some tests, i.e.
real pandas-related failures are unlikely):
mdtraj/amd64: upstream tests pass but a warning is printed to stderr,
which autopkgtest counts as a fail
partd/ppc64el: save/load of a plain string fails
snakemake/i386: hang in test_pipes_fail
statsmodels/armhf: wrong answer in TestDFMComplex
These have all happened before, suggesting they're random, i.e. bugs
in these packages not in pandas. (For partd, the other known instance
was on a different architecture (s390x), but it looks like partd only
uses pandas when it's handling pandas objects, which this test isn't.)