If there are any broken symlinks in the same directory as a setup.py when e.g. sdist is run, findall() will fall over when attempting to os.stat() the symlink:Traceback (most recent call last):  File "setup.py", line 81, in run    _sdist.run(self)  File "/usr/lib/python2.6/distutils/command/sdist.py", line 144, in run    self.get_file_list()  File "/usr/lib/python2.6/distutils/command/sdist.py", line 238, in get_file_list    self.filelist.findall()  File "/usr/lib/python2.6/distutils/filelist.py", line 47, in findall    self.allfiles = findall(dir)  File "/usr/lib/python2.6/distutils/filelist.py", line 297, in findall    stat = os.stat(fullname)OSError: [Errno 2] No such file or directory: 'debian/tmp/usr/share/somepath/somesymlink'Solutions would include replacing the call to os.stat() with one to os.lstat() (probably backwards-incompatible), or trying one and then the other.This bug is present in Pythons 2.6.6 (Debian 6.0.2) and 2.7 (Fedora 14).When attempting to reproduce in Python 3.1.2 (on Fedora) no error was encountered. However, looking at distutils/filelist.py, the same unadulterated call to os.stat() is present. I'll presume that for whatever reason distutils in Py3.1.2 never has cause to stat my broken symlink.

Symlinks are barely supported by distutils; the only mention you can find in the doc is that the MANIFEST file can contain symlinks.  In the light of that, one could argue that broken symlinks are not supported nor ignored, and that you just should not have them.  Is there a valid use case for having broken symlinks in a Python project?(About versions: I’m removing 2.6 which only gets security fixes, and adding 3.x to add regression tests to them, even if the bug isn’t there now.)

I've come across it as I'm creating a Debian package of the Python package in the same tree — I'm happy to be told this is a Bad Idea and that they should be in different places.The broken symlinks are relative and in debian/tmp, and will point to locations provided by other Debian packages once my package is installed in the right location.FWIW, I'm getting round it at the moment by walking the directory tree and removing the files is os.path.islink(filename) and not os.path.exists(os.path.join(filename, os.readlink(filename))).I'm happy to provide tests and a patch if necessary.

> I've come across it as I'm creating a Debian package of the Python> package in the same treeI think a lot of people are doing this.> The broken symlinks are relative and in debian/tmp, and will point to> locations provided by other Debian packages once my package is> installed in the right location.It’s too bad that filelist goes into the debian subdirectory :(  The Ubuntu-originated python-distutils-extra project had a similar problem and they switched from FileList.findall to os.walk to avoid it.What happens if you ignore the debian dir in MANIFEST.in?

This issue has been monkeypatched by setuptools for 7 years:https://bitbucket.org/pypa/setuptools/commits/4556f9d08ef7

Eric, would you be willing to review the commits in the referenced patch repo (collapsed in the attached diff)? If you're happy with the implementation, I'll backport it to Python 2.7 and apply it to 3.4-3.6 and then update setuptools to only monkeypatch older distutils.

Patch looks good and would fix the reported problem.  Being a backport from setuptools also gives confidence.Does the patch change the behaviour for the handling of the MANIFEST file (not MANIFEST.in)?  My previous message mentions that the docs say that one can include symlinks in MANIFEST.

Although the code was ported from setuptools, it wasn't ported from the long stable implementation, but was ported following the recent refactor, which had a [regression](https://bitbucket.org/pypa/setuptools/issues/425/odd-failure-on-183-missing-files-but-they) revealing that the [docstring no longer matched the implementation](https://bitbucket.org/pypa/setuptools/commits/72ec36bea07c7dd1cf48e20894864e8a4e1b480f). The latest implementation provides [a more elegant approach](https://bitbucket.org/pypa/setuptools/src/738b139f3016a05084c9f34c6364d7d20b3aac9d/setuptools/__init__.py?fileviewer=file-view-default#__init__.py-154), but continues to maintain the expectation of the setuptools implementation.I still need to address Eric's question about what filelists are affected, but also now update the patch to ensure that the disparity between findall('.') and findall(anything_else) is captured.

After porting the latest released Setuptools version, I discovered yet another regression in the Setuptools implementation, specifically around its handling for this issue, so I've created yet another release of Setuptools (18.3.2) to include tests and fixes for the expected behavior on this issue. Rebasing those changes, the patch3da42b593e1f now reflects my latest proposal.

> Does the patch change the behaviour for the handling of the MANIFEST file (not MANIFEST.in)?  My previous message mentions that the docs say that one can include symlinks in MANIFEST.This change will not affect the content of MANIFEST.in (template) nor MANIFEST. Previously, if a broken symlink was provided, it would raise an error. With the patch, under that condition the process will not fail but will instead ignore those broken symlinks as if they do not exist (and omit them from MANIFEST).Non-broken symlinks (those that resolve to directories or files) will continue to be treated exactly as if their targets exist in that place. To my knowledge, there is no special handling of symlinks (such as with tarfiles that store the underlying symbolic link details).

New changeset79422fab2ef4 by Jason R. Coombs in branch 'default':Issue#12885: Merge with 3.5https://hg.python.org/cpython/rev/79422fab2ef4

This code has been stable in Setuptools for some time, so I've pushed the code into the base.

I think you mixed up the bug number. You should probably fixMisc/NEWS.New changeset941346104718 by Jason R. Coombs in branch '3.4':Issue#12285: Add test capturing failure.https://hg.python.org/cpython/rev/941346104718New changeset56c60b3d06fb by Jason R. Coombs in branch '3.4':Issue#12285: Replace implementation of findall with implementation from Setuptools7ce820d524db.https://hg.python.org/cpython/rev/56c60b3d06fbNew changeset13619a3e0737 by Jason R. Coombs in branch '3.4':Issue#12285: Update NEWShttps://hg.python.org/cpython/rev/13619a3e0737

New changeset590d0de4ff48 by Jason R. Coombs in branch '3.4':Issue#12885: Correct issue reference in NEWShttps://hg.python.org/cpython/rev/590d0de4ff48New changeset74ccec0bd442 by Jason R. Coombs in branch '3.5':Issue#12885: Merge 3.4https://hg.python.org/cpython/rev/74ccec0bd442New changesetbfff89ed356b by Jason R. Coombs in branch 'default':Issue#12885: Merge with 3.5https://hg.python.org/cpython/rev/bfff89ed356b

New changesete82b995d1a5c by Jason R. Coombs in branch '3.4':Issue#12885: Revert commits in 3.4 branch which is security-only fixes.https://hg.python.org/cpython/rev/e82b995d1a5cNew changeseta7ce98a4e9e4 by Jason R. Coombs in branch '3.5':Issue#12885: Merge with 3.4, retaining commits reverted there.https://hg.python.org/cpython/rev/a7ce98a4e9e4New changesetfa82fabe9d65 by Jason R. Coombs in branch 'default':Issue#12885: Merge with 3.5https://hg.python.org/cpython/rev/fa82fabe9d65

New changeset556a11c11edd by Jason R. Coombs in branch '3.4':Issue#27960: Revert state to 675e20c38fdac6, backing out all changes by developed for Issue#12885.https://hg.python.org/cpython/rev/556a11c11edd