Bug#940679: pandas: FTBFS on amd64 - TestDatetimelikeSubtype::test_astype_category[index3] crash

Rebecca N. Palmer rebecca_palmer at zoho.com
Thu Sep 19 20:21:44 BST 2019


Control: retitle -1 pandas: random test crashes

You're right that there are multiple issues here.

--- Random test crashes (this bug):

tests/indexes/interval/test_astype.py::TestDatetimelikeSubtype::test_astype_category[index3] 
on amd64; worked when retried, also works here
tests/reshape/test_tile.py::TestDatelike::test_qcut_nat[timedelta64[ns]] 
on mips64el; worked when retried

Only seen in -6, but given their intermittency this isn't proof that 
they're new.  If they are new, they're probably changes in some 
dependency, as the actual -5 to -6 change is in a different test.  Not 
found in a search of upstream issues.

--- Bootstrap issue (arch:all BD-Uninstallable):

python3-pandas is arch:all (so last successfully built in -4), and is no 
longer in the sid Packages list at all.  (I don't know if this always 
happens this quickly.)

arch:all pandas build-depends on itself via python3-statsmodels and 
python3-xarray.

These are used in some tests (which skip, not fail, if they are not 
available), and some documentation examples (which leave error messages 
in the documentation, but don't fail the build).

Hence, a solution to this is to drop those build-dependencies for one 
upload.

--- Failure with statsmodels 0.9:

This looks like something that should actually be fixed in statsmodels, 
not worked around in a pandas test, as it can also break actual use:

https://github.com/statsmodels/statsmodels/issues/4772
https://github.com/statsmodels/statsmodels/pull/4963/commits/064c5c3c59946fc702880b7fbda0e0b81a1fe30c

--------

Hence, I plan to upload (but haven't yet tested):
pandas
- Revert the -6 patch and skip the test instead
- Temporarily drop statsmodels+xarray build-dependencies
- Add xarray autopkgtest-dependency
statsmodels
- Apply the above upstream patch

I also plan to investigate what causes these test crashes, but this may 
be hard if I can't reproduce them.



More information about the debian-science-maintainers mailing list