astropy: Occasional failures in vo_test.py

The Debian CI tests (which are run whenever one of the dependencies of astropy changed) sometimes show a segmentation fault on io/votable/tests/vo_test.py. I had this now the second time at the same place, so it was probably not just a glitch in our environment: here and here. The log excerpt there is

../../usr/lib/python2.7/dist-packages/astropy/io/votable/tests/vo_test.py ...........................................................................................................................................................................
bash: line 1:  1076 Segmentation fault      (core dumped)

The relevant changes that lead to the CI test were new debian release of gfortran-5.3.1 (from debian relaease 11 to 12 and 12 to 13) This may be similar to #4240 or #4352, but I am not sure. Apart from the two failures, I have however ~20 succeeding CI test runs. Platform is amd64.

About this issue

Original URL
State: closed
Created 8 years ago
Comments: 29 (27 by maintainers)

Most upvoted comments

Thanks for the triage, @kakirastern, and thanks for checking those many logs, @olebole. I agree that if there was no issue for that many, we can just close this for now. Hopefully, we did really solve it!

mhvk on Mar 10, 2019

Looking int our Debian CI logs (~300 tests), I couldn’t find that segfault in the last year for the Python 3 package. So, I would guess this is somehow fixed and would propose to close it. If it comes back, we can still reopen.

olebole on Mar 10, 2019

I have tracked the segfault down to: https://github.com/astropy/astropy/blob/master/astropy/io/votable/src/tablewriter.c#L74 In tablewriter.c the first call to stdlib realloc() segfaults.

Anyway, this is a recurrence of #2100. This may be related to the specific win64 anaconda python 3.6, but that is compiled with the same MSVC as I’m using on astropy. I’ll see if I can test this with the official CPython binaries.

IMHO, even if it is specific to anconda python 3.6, this should be fixed: This calls realloc on a Py_Object, which is allocated by Py_MemAlloc. It should not be realloced by the stdlib realloc but through the python memory manager. (Changing the realloc to PyMem_Realloc fixes my failures)

tomkooij on Jan 13, 2017

I see. Then let’s keep this issue stay open for now… Meanwhile, I will help keep an eye on the bug should it crop up during CI tests again in the near future. I understand the gravitas of a segfault. And the bug appears to affect the io/votable/tests/vo_test.py only.

kakirastern on Mar 9, 2019

I just encountered one in CircleCI:

____________ lib.linux-i686-3.5/astropy/io/votable/tests/vo_test.py ____________
[gw2] linux -- Python 3.5.2 /usr/bin/python3
Slave 'gw2' crashed while running 'lib.linux-i686-3.5/astropy/io/votable/tests/vo_test.py::TestThroughBinary2::()::test_string_test'

MSeifert04 on Oct 12, 2017

Commenting out the suggest lines, does indeed cause a crash in a different test (which also calls tablewriter.c.

I’m going to retry on python 3.5 first, to see if it is indeed only related to python 3.6. EDIT: Without recompiling installed in python 3.5: No test failures/segfaults.

tomkooij on Jan 12, 2017