In Lib/ntpath.py:

# realpath is a no-op on systems without islink support
realpath = abspath


However, Windows Vista and newer support symbolic links and other Python methods support them.

(noticed this through source code inspection; haven't actually tested it)

The issue #12799 has been marked as a duplicate of this issue:
"From a look at the code, realpath() does not seem to resolve symlinks under Windows, even though we have the _getfinalpathname function to do that. Please indulge with me if I'm wrong :)"

This is a patch for the os.path.realpath() bug under Windows,  http://bugs.python.org/issue9949  "os.path.realpath on Windows does not follow symbolic links"

ntpath.diff fixes the realpath() function to resolve symbolic links to their targets (tested under Windows 7).

Note: I tried to make the equivalent patch for Python 2.7, but it didn't work because the "from nt import _getfinalpathname" fails.

It seems that the nt module is implemented in the posixmodule.c source file, and the Python 3 version contains the posix__getfinalpathname entry point, but the Python 2 version does not.

I presume that PyWin32 could also be used to work around this.  Too bad it isn't automatically included with Python:
http://sourceforge.net/projects/pywin32/files/pywin32/Build%20217/

file, and the Python 3 version contains the posix__getfinalpathname entry
point, but the Python 2 version does not.
>
> I presume that PyWin32 could also be used to work around this.  Too bad
it isn't automatically included with Python:
> http://sourceforge.net/projects/pywin32/files/pywin32/Build%20217/

We can't use external packages in the standard library. If you're looking
for a Python which includes pywin32, ActiveState provides this.

I think we can just implement the function we need.

Excellent!

The ntpath.py change is nearly identical in Python 2.7 to the change for Python 3.2.  The only difference is that instead of:

+        elif isinstance(path, bytes):
+            path = os.getcwdb()

It is:

+        elif isinstance(path, unicode):
+            path = os.getcwdu()

I have attached a series of patches with (hopefully) provide more robust fix for this issue, against the Python 3.3 branch.  It handles both bytes and str objects, paths that do not actually exist on the filesystem, and removal of the '\\?\' prefix returned by _getfinalpathname, unless the supplied path had that prefix already.  A couple of the patches contain some separate infrastructure that could hopefully be used to simplify some of the other Windows code in posixmodule.c (One immediate possibility would be to combine the code provided in these patches with the symbolic-link resolution code in the Windows stat functions - there is quite a bit of duplication there that could be eliminated.)

One thing these patches do not address is resolving a broken symbolic link.  The Windows API function GetFinalPathNameByHandle does not handle this case (because CreateFile cannot be used to get a handle if the symbolic link is broken).  This functionality could be implemented by manually following the reparse points, but that would basically require reimplementing GetFinalPathNameByHandle.

Finally, this patch could be fairly easily backported to Python 3.2, but that shouldn't be done without careful consideration.  It changes the return value from os.path.realpath on Windows even when there are no symbolic links in the path (the returned value will have the actual casing as stored on the filesystem, instead of the casing supplied by the user).  I don't think it should be backported to Python 2.7, because that version, like all Python versions before Python 3.2 are unaware of symbolic links on Windows (e.g. lexists is the same function as exists).

The patches were generated using git (I use git-hg) - if that format is a problem, let me know and I can regenerate them.

Uploading a new series of patches - they are all the same as the first set, except for 0006-Make-realpath-follow-symbolic-links-on-Windows.patch.  I realized that I could use os.readlink to handle broken symbolic links, so I changed the logic of os.path.realpath slightly to do that (including recursive symlink detection).

The previous version of this patch did not handle bytes arguments correctly and could fail in conjunction with a non-ASCII compatible encoding.  Also, if the result was a UNC path, it was not being handled correctly (the returned value would have started with a single backslash, not two).  Version 3 of the patch fixes these issues.  I am also attaching a single, condensed patch as requested by the Lifecycle of a Patch devguide.  I can also provide a patch series as before if desired.

Yet another patch to support symlink on Windows.