Segmentation fault by importing scipy.interpolate

[matthias]

importing scipy.interpolate leads to a segmentation fault

kubuntu 13.04
yade 0.90.3 (stable) and yade_daily (0.97.0)
python 2.7.5

[Alexander Eulitz [Eugen] (kubeu)]

[Kubeu] for details see -> https://answers.launchpad.net/yade/+question/229455

[Bruno Chareyre (bruno-chareyre)]

same problem with scipy.stats

[Bruno Chareyre (bruno-chareyre)]

I found something that could help for a workaround:

ln -s yade-daily yadeimport.py

$ipython
In [1]: import yadeimp
In [2]: import scipy.interpolate
Segmentation fault

$ipython
In [1]: import scipy.interpolate
In [2]: import yadeimp
-> No problemo

Importing scipy by default at startup?...

[Anton Gladky (gladky-anton)]

Hmm, interesting.

Importing scipy at startup is OK, but we need to add scipy in
dependencies.

It is better to find out, why it crashes to escape a workaround.

Anton

2013/8/23 Bruno Chareyre <email address hidden>:
> I found something that could help for a workaround:
>
> ln -s yade-daily yadeimport.py
>
> $ipython
> In [1]: import yadeimp
> In [2]: import scipy.interpolate
> Segmentation fault
>
> $ipython
> In [1]: import scipy.interpolate
> In [2]: import yadeimp
> -> No problemo
>
> Importing scipy by default at startup?...
>
> --
> You received this bug notification because you are subscribed to Yade.
> https://bugs.launchpad.net/bugs/1183402
>
> Title:
> Segmentation fault by importing scipy.interpolate
>
> Status in Yet Another Dynamic Engine:
> New
>
> Bug description:
> importing scipy.interpolate leads to a segmentation fault
>
> kubuntu 13.04
> yade 0.90.3 (stable) and yade_daily (0.97.0)
> python 2.7.5
>
> To manage notifications about this bug go to:
> https://bugs.launchpad.net/yade/+bug/1183402/+subscriptions

[Bruno Chareyre (bruno-chareyre)]

Well, importing scipy by default was a stupid idea, since many other modules may have to be imported the same way (Raphael was importing his own fortran library, for instance).

I added an optional command line argument, so that it does not add global dependencies for yade.
https://github.com/yade/trunk/commit/317d1718bd29cba22a1d22ed9422681f36d6e2ab

@Matthias, you can type "yade -lscipy.interpolate"
@Raphaël: "yade -lnsmp1d_yade"

I've seen a few projects having similar problems with scipy, some of them quite old [,12,3].
I noticed that both scipy and Raphael's library are based on fortran code. Maybe related?
I also realize that Raphael's lib has blas/lapack/friends as dependencies, which is also the case of some modules imported by yade (matplotlib, PFV option,...). They are maybe using incompatible variants of those, as suggested by [2]?

[1] http://mail.scipy.org/pipermail/scipy-dev/2007-March/006724.html
[2] http://blog.tremily.us/posts/SciPy_and_ATLAS/
[3] http://lists.opensuse.org/opensuse-bugs/2010-11/msg00466.html

[Bruno Chareyre (bruno-chareyre)]

For the record, a similar problem found a solution in [1]. The fix mentioned there (see below) doesn't work in our case unfortunately, but it may ring a bell for someone.

import sys
flags = sys.getdlopenflags()
import pybel
sys.setdlopenflags(flags)
import scipy.linalg

[1] http://grokbase.com/t/scipy.org/scipy-tickets/12bc83e10j/scipy-1772-scipy-linalg-segfaults-when-imported-after-pybel

[Anton Gladky (gladky-anton)]

It looks like the problem is gone with in latest distributions. In Debian Jessie the problem does not exist any more.

[Bruno Chareyre (bruno-chareyre)]

Thanks Anton -> "won't fix"