MAINT: Python 3.12 support improve#4300
Conversation
* Fixes MDAnalysisgh-4299, and allows the testsuite to have 12 failures instead of 56 with Python `3.12.0rc3` for me locally on x86_64 Linux; full suite still passing for Python `3.11.x` * the main idea is to circumvent more aggressive CPython blocks on serialization by overriding the highest-priority pickling method available, `__reduce_ex__`, which CPython had locked down, making it such that our old `__getstate__` pickling shims were ignored as indicators of pickling safety * there is probably a slightly simpler diff that allows this to work, but this took a few hours to draft so feel free to push in simplifications if you're sure you've tested on 3.11.x and 3.12 RC [skip cirrus]
|
Hello @tylerjereddy! Thanks for opening this PR. We checked the lines you've touched for PEP 8 issues, and found:
|
Linter Bot Results:Hi @tylerjereddy! Thanks for making this PR. We linted your code and found the following: Some issues were found with the formatting of your code.
Please have a look at the Please note: The |
|
Thanks @tylerjereddy - should we temporarily merge my Py3.12 PR knowing it fails and then just turn off the cron shim to test here? |
|
We could I suppose. Do we even need a cron guard? The final CPython |
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## develop #4300 +/- ##
===========================================
- Coverage 93.40% 93.36% -0.05%
===========================================
Files 170 184 +14
Lines 22257 23374 +1117
Branches 4071 4071
===========================================
+ Hits 20790 21823 +1033
- Misses 951 1035 +84
Partials 516 516
☔ View full report in Codecov by Sentry. |
* force the use of a relative path for `mkdtemp` in our source code because, from the Python `3.12` release notes: > `tempfile.mkdtemp() now always returns an absolute path, even if the argument provided to the dir parameter is a relative path.` * with this patch, the full testsuite now passes on Python `3.12.0` (which is now out) when combined with MDAnalysisgh-4301 and MDAnalysisgh-4300 (though see my cautions about thread safety/concurrency for the latter) * this fixes the 1 remaining failure with 3.12: `MDAnalysisTests/analysis/test_hole2.py::TestCheckAndFixLongFilename::test_symlink_dir` [skip cirrus]
* force the use of a relative path for `mkdtemp` in our source code because, from the Python `3.12` release notes: > `tempfile.mkdtemp() now always returns an absolute path, even if the argument provided to the dir parameter is a relative path.` * with this patch, the full testsuite now passes on Python `3.12.0` (which is now out) when combined with gh-4301 and gh-4300 (though see my cautions about thread safety/concurrency for the latter) * this fixes the 1 remaining failure with 3.12: `MDAnalysisTests/analysis/test_hole2.py::TestCheckAndFixLongFilename::test_symlink_dir` [skip cirrus]
4 participants
Fixes Python 3.12 cannot pickle PickleBufferIOPicklable instances #4299, and allows the testsuite to have 12 failures instead of 56 with Python
3.12.0rc3for me locally on x86_64 Linux; full suite still passing for Python3.11.xthe main idea is to circumvent more aggressive CPython blocks on serialization by overriding the highest-priority pickling method available,
__reduce_ex__, which CPython had locked down, making it such that our old__getstate__pickling shims were ignored as indicators of pickling safetythere is probably a slightly simpler diff that allows this to work, but this took a few hours to draft so feel free to push in simplifications if you're sure you've tested on 3.11.x and 3.12 RC
also, I believe this does slightly increase the chance of hanging the testsuite at very high levels of concurrency, but I needed more than 30 processes to repro that, so it may just be a feature of our testsuite in general (it was fine on 12 cores at least, but on 32 or so file locks had issues I think)
[skip cirrus]
📚 Documentation preview 📚: https://mdanalysis--4300.org.readthedocs.build/en/4300/