According to the documentation https://docs.python.org/3/library/pathlib.html#pathlib.Path.resolve
If strict is False, the path is resolved as far as possible and any remainder is appended without checking whether it exists.

The current behavior is not consistent with this, and only appends the first remainder.

For example:
If we have an empty '/tmp' directory
Path('/tmp/foo').resolve() and Path('/tmp/foo/bar').resolve() both result in Path('/tmp/foo') but Path('/tmp/foo/bar').resolve() should result in Path('/tmp/foo/bar')

I can reproduce this bug. This behavior is really confusing.

Pull Request (PR 1649) treats this as a documentation problem. I would argue that the documentation is correct and this is a bug in the code.

The `strict` flag was added as a result of issue19717. The decision on what to do when strict=False seems to come in https://mail.python.org/pipermail/python-ideas/2016-September/042203.html where Guido says:

"I would prefer it if Path.resolve() resolved symlinks until it hits
something that doesn't exist and then just keep the rest of the path
unchanged."

The documented behaviour also seems much more useful than the current behaviour.

This is definitely a bug if it really behaves as described.

PRs for fixing the functionality are welcome - I'm inclined to say fix it in 3.6 as well, but it might be too big a behavioural change if people are already working around it.

In Windows/Python 3.6, the behavior matches the documented one, so this could actually be an important silent bug when executing a code wrote for Windows in another OS. I'd argue to fix it in 3.6 too.

By the way, I'm pasting what I said on the PR here: 

For what it's worth, this bug was the cause of an important downtime in our organization, because someone needed to normalize a path that was later passed to a .mkdir(parents=True), and it was, in some cases, silently doing the mkdir of a wrong path instead of creating all the parents.

This is definitely a bug and I think the fixes should be backported.

Ah, didn't catch that it doesn't occur on Windows - that explains why I've never seen it before.

Yes, definitely fix and backport. Adding the RMs in case they want to delay any upcoming releases to get the fix.

I agree bpo-30177 is not a suitable fix for this issue, as it fixes the doc instead of fixing the actual underlying bug in the function behavior.

This bug can lead to files being added to the wrong (parent) directory, which is quite critical.

We have less than 2 weeks until cutoff for 3.6.2 so it would be great if someone would take this on.

I'm on it.

The initial fix should be easy:

--- a/Lib/pathlib.py
+++ b/Lib/pathlib.py
@@ -329,8 +329,6 @@ class _PosixFlavour(_Flavour):
                     if e.errno != EINVAL:
                         if strict:
                             raise
-                        else:
-                            return newpath
                     # Not a symlink
                     path = newpath

However, the trick is going to be in the tests, which are shared between POSIX and Windows - and are apparently passing on Windows right now. I'm not entirely sure why that is, but it may not be as simple here as "works on POSIX".

Ah, looks like they require symlinks for the whole test, which means you need to be admin when running them on Windows.

Zach - do we have any buildbots running as admin for symlink tests?

I added a fix for the tests and the code.

> Ah, looks like they require symlinks for the whole test, 
> which means you need to be admin when running them on Windows.

The privilege to create symlinks isn't filtered out of a standard user's token. Are there any buildbots already running as a standard user? In that case it may be simpler to modify the user's rights via secpol.msc: Local Policies -> User Rights Assignment -> Create symbolic links.

Support could also be added for the new feature in Windows 10 to allow unprivileged creation of symlinks when the system is in developer mode and the flag SYMBOLIC_LINK_FLAG_ALLOW_UNPRIVILEGED_CREATE [1] is used. A keyword-only option to use this flag could be backported to 3.6 and enabled by default in 3.7.

[1]: https://blogs.windows.com/buildingapps/2016/12/02/symlinks-windows-10

> do we have any buildbots running as admin for symlink tests?

No, as far as I know.  I just took Eryk's suggestion and gave the buildslave user on my Windows 8.1 bot rights to create symbolic links, though.  It's now rebooting after updates, we'll see how it does shortly.

So, I asked a friend to check again with a more recent Python version on Windows and he can't reproduce the documented behavior, so the bug seems to also be present on Windows. My patch doesn't address that for now, which explains why the build fails (and why it was working before).

I updated the PR to fix the Windows part of the issue thanks to Zachary who gave me access to a Windows machine.

The code has been reviewed by (the other) Antoine, I guess there is now everything needed to merge it?

New changeset add98eb4fe41baeaa70fbd4ccc020833740609a4 by Steve Dower (Antoine Pietri) in branch 'master':
bpo-30177: pathlib: include the full path in resolve(strict=False) (#1893)
https://github.com/python/cpython/commit/add98eb4fe41baeaa70fbd4ccc020833740609a4

As usual, I can easily hit merge but may not be able to get to the backport immediately. Someone else can feel free to cherrypick and submit the PRs and I'll hit merge on them.

We should also watch the buildbots for failures though before backporting. Particularly in this area, they should have better coverage than the PR checks.

The only backport was for 3.6 as the previous versions don't have the strict= parameter. PR submitted here: https://github.com/python/cpython/pull/1985

New changeset ceabf9acf03f9bbe660d856bff90ecab475ab543 by Steve Dower (Antoine Pietri) in branch '3.6':
bpo-30177: pathlib: include the full path in resolve(strict=False) (#1893) (#1985)
https://github.com/python/cpython/commit/ceabf9acf03f9bbe660d856bff90ecab475ab543

Good point about not needing 3.5.

Buildbots were clean, so I merged it. Thanks Antoine!

New changeset 8399a177de8bfa860a66e96665488c17199cb9d2 by Ned Deily (Antoine Pietri) in branch '3.6':
[3.6] bpo-30177: add NEWS entry (#2134)
https://github.com/python/cpython/commit/8399a177de8bfa860a66e96665488c17199cb9d2

New changeset a77a35d70ba8aed047e63d4d9f7d0554e98d4c4b by Ned Deily (Antoine Pietri) in branch 'master':
bpo-30177: add NEWS entry (#2135)
https://github.com/python/cpython/commit/a77a35d70ba8aed047e63d4d9f7d0554e98d4c4b