En Wed, 25 Feb 2009 21:24:33 -0200, <chr...@fsfe.org> escribió:
On Wed, Feb 25, 2009 at 05:05:28PM -0200, Gabriel Genellina wrote:
I'd try to move all the global stuff in that module into a function,
"init". Importing the module will always succeed - you have to manually
call init() after importing it.
i normally do that anyway and would also have done so here (afaik, it's
best practice anyway), but this is in this case not possible as the
modules in question are not under my control at all. as pep 299 has been
turned down, this can't really be considered a bug of those files.
You might ask the module author for this feature... At least, try to move
the "registration" at the end, after any code likely to raise exceptions.
anyway, it would just be another workaround around the globals
disappearing.
That makes a strange situation where the module doesn't exist in
sys.modules but its globals are still alive...
is this a known bug?
Not that I know.
i mean from a garbage collection / refcount point
of view, at least everything that is still referenced somewhere should
be preserved.
In older Python versions, an error when importing a module could leave a
module object in sys.module partially initialized. That's very bad! Now,
the entry in sys.modules is removed in such cases. Note that functions and
classes defined inside a module don't have a reference to the module
itself - their __module__ attribute is a string (although functions *do*
have a reference to the module namespace, as you have remarked). So the
entry in sys.modules is likely to be the last and only reference, and when
it's removed, the module is deleted. This is a very tricky operation and
at some stage involves setting all globals to None; if not were for any
external references (like a callback registered somewhere else, as in your
case) the module plus the functions and classes and objects defined inside
it plus its global namespace, all of these should disappear, kaput, gone.
You should report it at http://bugs.python.org -- I don't think this is
likely to be "fixed" now, but at least it's on record. And when someone
attempts to improve the module destruction sequence, there is a better
chance this is taken into account too.
--
Gabriel Genellina
--
http://mail.python.org/mailman/listinfo/python-list
