homepage

platform.libc_ver() is trivially broken as it uses string comparison internally to determine the maximum libc version number (which is obviously broken as "2.9" > "2.10").

True. At the time the code was written, this was not an issue :-)
Is the libc version information documented somewhere ? If so, we could probably add a better parser for it.

Adding other Python versions as well, since this is a bug.

We just ran into this in pip -- https://github.com/pypa/pip/pull/3836
I'd really recommend dropping the current "grovel through the binary doing a regex search" strategy -- it's incredibly error prone, and AFAICT doesn't really give any value. This code OTOH reliably lets you detect glibc and gives the exact version number with no fuss:
 https://github.com/pypa/pip/blob/master/pip/utils/glibc.py#L9
What about non-glibc systems? Unfortunately the current libc_ver() turns out not to work well for those either. Attached is a CSV file showing the return value of ~1.2 billion calls to platform.libc_ver() by the last 6 months of pip users. You can see that the current code basically never returns anything useful for non-glibc platforms. (The one exception is that it seems to be able to detect uclibc 0.9.32 and label it as "libc 0.9.32".)
Don't get me wrong: it'd be really really useful if there were some way to detect and distinguish between the common non-glibc libcs like musl/bionic/uclibc/..., but I'm not sure how to do that -- and unfortunately the current code definitely doesn't do the job :-(.

At the time the code was written, libc and glibc were in wide spread use, so it's not surprising that it doesn't work well for other C libs.
Note that the routine returns the highest libc version number used and required by the executable (usually the Python interpreter). This does not necessarily correspond to the version installed on the system. The purpose of the function was to determine the minimum libc compatibility requirements of the executable.
The routine you quote uses ctypes and only works for glibc, so parsing needs to be kept around as fallback solution. It also returns the libc version that is currently used on the system; not necessarily the minimum version required, so semantics are different.

> The purpose of the function was to determine the minimum libc compatibility requirements of the executable.
For what it's worth, I didn't know this, the pip authors obviously didn't know this, and after rereading the docs just now I still can't quite tell that this was intended. I'm not sure why one would want these semantics either, but at a minimum it would help to document the intended semantics much more clearly.
> parsing needs to be kept around as fallback solution
The point of the data I attached was that AFAICT there don't exist any currently-in-use, non-glibc systems where "parsing" returns a meaningful result.

I need the glibc version for skipping a test if run with glibc containing a bug (see issue31630 and issue33630). platform.libc_ver() is not usable, it always returns '2.9' (the version that has bugs), while the actual version on my computer is '2.25' (the version that doesn't have these bugs).

I agree that the strategy used for platform.libc_ver() is not perfect. But the implementation has bugs that make it useless. The following PR fixes two bugs in the implementation:
1) Version numbers compared as strings.
2) Versions that are located on the border of 16 KiB blocks were not recognized or were recognized incorrectly.

New changeset 2a9b8babf0d09946ebebfdb2931cc0d3db5a1d3d by Serhiy Storchaka in branch 'master':
bpo-26544: Fixed implementation of platform.libc_ver(). (GH-7684)
https://github.com/python/cpython/commit/2a9b8babf0d09946ebebfdb2931cc0d3db5a1d3d

New changeset 7c43b801503c802ed6ea4b811f5bc73791249d94 by Serhiy Storchaka in branch '3.7':
[3.7] bpo-26544: Fixed implementation of platform.libc_ver(). (GH-7684). (GH-8193)
https://github.com/python/cpython/commit/7c43b801503c802ed6ea4b811f5bc73791249d94

New changeset d73497ba52171bc8f786a70ecf50d3104b596221 by Serhiy Storchaka (Miss Islington (bot)) in branch '3.6':
bpo-26544: Fixed implementation of platform.libc_ver(). (GH-7684). (GH-8193) (GH-8195)
https://github.com/python/cpython/commit/d73497ba52171bc8f786a70ecf50d3104b596221

New changeset b1e6e5615a8e82fcf569368fac5c5b0385929855 by Serhiy Storchaka in branch '2.7':
bpo-26544: Fixed implementation of platform.libc_ver(). (GH-7684). (GH-8193) (GH-8196)
https://github.com/python/cpython/commit/b1e6e5615a8e82fcf569368fac5c5b0385929855

Can we close this issue? It seems like the bug has been fixed, no?

please don't close this yet. the patch introduces a dependency on the distutils package. I don't see any reason to compare the libc version as a python egg version, so please avoid that.
if we need to have a module to compare arbitrary version strings, then we should think about adding one in the standard library, but I think that's what the packaging module is for.

a tuple comparison should be good enough

no, it's not fixed at all. Setting as a release blocker for 3.6 and 3.7.

proposed patch attached

platform-no-distutils.diff restores the original bug: "2.9" > "2.10".
PR 8356 removes the dependency from distutils and use a sophisticated function for parsing versions.

fine! what prevents merging and backporting this issue?

I'll merge it if it looks good to you.

I added a comment to the PR, but other than that I think it's good to go.

+1

New changeset 7d81e8f5995df6980a1a02923e224a481375f130 by Serhiy Storchaka in branch 'master':
bpo-26544: Get rid of dependence from distutils in platform. (GH-8356)
https://github.com/python/cpython/commit/7d81e8f5995df6980a1a02923e224a481375f130

Would it be possible to add a few tests on platform._comparable_version()? It would make me more confortable for backports.

New changeset 7917aadb3edb7616d6164c5eaba24df6ac0a5fc6 by Serhiy Storchaka in branch 'master':
bpo-26544: Add test for platform._comparable_version(). (GH-8973)
https://github.com/python/cpython/commit/7917aadb3edb7616d6164c5eaba24df6ac0a5fc6

Thanks for adding tests ;-) It now looks better to me ;-)

New changeset 20a8392cec2967f15ae81633c1775645b3ca40da by Serhiy Storchaka in branch '3.7':
[3.7] bpo-26544: Get rid of dependence from distutils in platform. (GH-8356). (GH-8970)
https://github.com/python/cpython/commit/20a8392cec2967f15ae81633c1775645b3ca40da

New changeset 1a3eb125dc07a28a5af62446778ed7cca95ed3da by Victor Stinner (Miss Islington (bot)) in branch '3.6':
[3.7] bpo-26544: Get rid of dependence from distutils in platform. (GH-8356) (GH-8970) (GH-9061)
https://github.com/python/cpython/commit/1a3eb125dc07a28a5af62446778ed7cca95ed3da

New changeset 9734024ec65311e33936faa83fb1cb249ef0de9d by Victor Stinner (Miss Islington (bot)) in branch '2.7':
bpo-26544: Get rid of dependence from distutils in platform. (GH-8356) (GH-8952)
https://github.com/python/cpython/commit/9734024ec65311e33936faa83fb1cb249ef0de9d

Thanks everybody! The issue should now be fixed in all supported branches ;-)

Thank you for merging PRs Victor, but seems you have merged wrong PR for 2.7.

> Thank you for merging PRs Victor, but seems you have merged wrong PR for 2.7.
Oh. I didn't notice that there were two PRs for 2.7. Do you want to rebase your PR 9062 on 2.7, or revert the commit that I merged? It's up to you ;-)
I reopen the issue.
It seems like PR 8971 is also open.

This is now only open and a release blocker for 2.7, correct?

Seems the only difference between PR 8952 and PR 9062 is additional tests. This is not important.

New changeset 8687bd86e6f138ef0699a1e9f3f9555765949b51 by Victor Stinner in branch '2.7':
bpo-26544: Make platform.libc_ver() less slow (GH-10868)
https://github.com/python/cpython/commit/8687bd86e6f138ef0699a1e9f3f9555765949b51

> bpo-26544: Make platform.libc_ver() less slow (GH-10868)
> https://github.com/python/cpython/commit/8687bd86e6f138ef0699a1e9f3f9555765949b51
"Coarse benchmark on Fedora 29: 1.6 sec => 0.1 sec."
Oops, my benchmark in the commit message was wrong, it included the startup time. Correct benchmark says 44x faster, it's way better!
Python 2.7: [old_py2] 1.51 sec +- 0.03 sec -> [if_in] 34.6 ms +- 0.4 ms: 43.61x faster (-98%)