On 02/03/11 16:28, Tom Lane wrote:
> Jan Urbański <wulczer@wulczer.org> writes:
>> On 02/03/11 14:25, Robert Haas wrote:
>>> But does bumping the ref count then create a leak the rest of the time?
>
>> Not really, because you never want to garbage collect the spiexceptions
>> module (just like you don't want to GC th plpy module, or the plpy.info
>> function etc.). So the reference count of that module should never drop
>> to zero, but apparently on some machines it does. So just reffing
>> artificailly is kind of a valid solution, I'm just uneasy with not
>> knowing why it fails on some machines and does not on others.
>
> Yeah, that last point makes me nervous too. A look into the Fedora
> repository shows that the python version shipped in F13 is rather
> heavily patched:
> http://pkgs.fedoraproject.org/gitweb/?p=python.git;a=tree;h=refs/heads/f13/master;hb=refs/heads/f13/master
> It's not clear to me which of their changes from a stock build might
> be at issue, though, and even less clear whether they introduced a
> bug or did something to expose a bug of ours.
FWIW I looked at these patches yesterday when I was trying to reproduce
the bug, but did not find anything interesting. It's mostly for stuff in
the standard library. I haven't tried building Python with all of of
these patches though, and did not find an easy way to rebuild a SRPM on
a Debian system. I'm also wondering if it can be a 32 vs 64 bit issue?...
Jan
