Doubley imported module caused devastating bug
Zac Burns
'instance' of a module. One notable example is "import __init__" from
a package. See http://stackoverflow.com/questions/436497/python-import-the-containing-package
This recently caused a devastating bug in some of my code. What I have
is support for the Perforce global options as a context for a perforce
module. http://www.perforce.com/perforce/doc.072/manuals/cmdref/o.gopts.html#1040647
This way one can call functions that call many perforce command and
have them execute on a different client for example.
So, in module A and module B both imported the Perforce module, but
they turned out not to be the same module. Module A did "with
Perforce.GlobalOptions(client=client): B.function()"
B.function did not receive the new GlobalOptions because of this
problem. As a result important files on the original client were
overwritten (OUCH).
I would like to propose that it be made impossible in the Python
source to import two instances of the same module.
--
Zachary Burns
(407)590-4814
Aim - Zac256FL
Production Engineer (Digital Overlord)
Zindagi Games
Ethan Furman
I believe that modules are imported only once, and my toy example
demonstrates that (python 2.5):
test_import.py:
"""testing multiple imports"""
CONSTANT = 928
version = (2, 0, 9)
plug_ins = []
random_text = 'some text here'
def set_text(new_text):
global random_text
random_text = new_text
plug_ins.append(new_text)
A.py:
import test_import
print test_import.version
test_import.set_text('hello!')
print test_import.plug_ins
print test_import.CONSTANT
test_import.CONSTANT = 'changed!'
B.py:
import test_import
print test_import.version
test_import.set_text('world!')
print test_import.plug_ins
print test_import.CONSTANT
Running...
In [1]: import A
(2, 0, 9)
['hello!']
928
In [2]: import B
(2, 0, 9)
['hello!', 'world!']
changed!
As you can see, module A made a change to test_import.CONSTANT, and if
they were different things then B would not have seen it, yet B *did*
see it.
This makes me wonder if A) Perforce.GlobalOptions isn't actually setting
module level variables, or B) B.function is using copies of those
variables that were set when B was originally imported, so is not seeing
the changes.... or C) there is yet another wrinkle here that I don't
know about. ;-)
~Ethan~
Carl Banks
> Currently it is possible to import a file of one path to more than one
> 'instance' of a module. One notable example is "import __init__" from
>
> This recently caused a devastating bug in some of my code. What I have
> is support for the Perforce global options as a context for a perforce
> This way one can call functions that call many perforce command and
> have them execute on a different client for example.
>
> So, in module A and module B both imported the Perforce module, but
> they turned out not to be the same module. Module A did "with
> Perforce.GlobalOptions(client=client): B.function()"
>
> B.function did not receive the new GlobalOptions because of this
> problem. As a result important files on the original client were
> overwritten (OUCH).
>
> I would like to propose that it be made impossible in the Python
> source to import two instances of the same module.
Impossible's a pretty strong word.
It's a reasonable request, but with Python's importing the way it is
it'd be kind of hard to do. A Python file can be visible in multiple
ways.
However, anyone who does "import __init__" (or "from . import
__init__" with relative import) is asking for trouble, I can't think
of any valid reason to do it, and I wouldn't mind seeing that
forbidden, but it's simple to avoid. Someone probably did that
because they didn't know how to import a containing package from one
of its modules, failing to realize that it created a new module.
Carl Banks
Zac Burns
> http://mail.python.org/mailman/listinfo/python-list
>
There are corner cases. The corner case that I ran into was that there
were two ways to find the module on PATH because one value of PATH was
over another. Since then this problem has been removed and it wasn't
too much trouble to work around - but finding the problem was a real
pain.
I am not intimately familiar with the import code and trust your
judgment that it is difficult. If people are in agreement that this
should be changed though it could be put in a list somewhere waiting
for some ambitious person to figure out the implementation, no?
Personally I think it would be worthwhile.
Terry Reedy
/__init__.py is basically an implementation hack to make a directory
also 'be' a file. Use at one own risk, I say.
> There are corner cases. The corner case that I ran into was that there
> were two ways to find the module on PATH because one value of PATH was
> over another. Since then this problem has been removed and it wasn't
> too much trouble to work around - but finding the problem was a real
> pain.
>
> I am not intimately familiar with the import code and trust your
> judgment that it is difficult. If people are in agreement that this
> should be changed though it could be put in a list somewhere waiting
> for some ambitious person to figure out the implementation, no?
1. It would slow down all imports, at least a bit.
2. It would kill code that intentionally makes use of duplicate modules
(but this could be considered exploitation of a bug, perhaps). It would
also make forced module reloads harder, it not impossible. Currently,
just delete the entry in sys.modules.
3. The language itself does not specify how and where from an
implementation 'initializes' a module on first import. Indeed, CPython
has at least three options (.py, .zip, and .dll or .pyd (Windows)), with
hooks for more. Lets a take the request as specifically preventing the
creation of duplicate module objects from a particular .py file.
One implementatin *might* be add a set to sys, say sys.mod_files for
x.py or x.pyc files already used to initialize a module. The .py or .pyc
or .pyo would be stripped but the name otherwise should be the absolute
path. (Including drive letter, on Windows).
This would not cover the case when files are symlinked (or copied). For
*nix, a set of inode numbers could be used, but not for Windows. I
suspect there might be other system-specific problems I have not thought of.
Terry Jan Reedy
Francis Carr
> source to import two instances of the same module.
A fully-automatic solution is more difficult than it might seem at
first:
http://www.python.org/dev/peps/pep-0328/
But there is a simple code-discipline solution: never ever use
relative imports, even between code in the same package.
We got "bit" by double-imports (a mix of relative, absolute, and even
cross-imports A "import B" and B "import A") early on in one of our
projects. The symptom was that the imported module would be
initialized *twice*, once for a relative import and once for an
absolute. This is not a happy situation for pseudo-singletons like
the "logging" module --- esp. if one is hacking the internals! :-) We
no longer use relative imports *EVER*, even within the same package.
Perhaps Perforce is doing something tricky with scoping or importing,
and you've just managed to stumble across this trickiness because of a
double-import. Good luck, these things are a bugger to debug.