Broken pyperformance benchmarks

[brandtbucher]

Broken:

• sqlalchemy_declarative
• sqlalchemy_imperative

Cause: https://bugs.python.org/issue46836

Fixes:

• FIX: import internal cpython headers to fix for py311 cython/cython#4667
• Update for Python 3.11 alpha 6 python-greenlet/greenlet#294

[vstinner]

genshi

What is the issue with genshi?

[brandtbucher]

What is the issue with genshi?

Sorry, that was an unrelated issue that has since been fixed. Removed it from the list!

[gvanrossum]

See also #333 for Windows breakages.

[gvanrossum]

Seems a new Genshi issue has popped up, see #333 (comment) and ff.

[gvanrossum]

Well, it wasn't a new genshi issue -- it's fixed in master but there's no release yet.

[vstinner]

My greenlet PR is merged python-greenlet/greenlet#294 but there is no release yet.

Problem: it's unclear to me if gevent uses or can use the newer greenlet. It seems like greenlet got a new "v2" design which better hides its internals and CPython internals.

[vstinner]

Currently, on PyPI, there are only greenlet 1.x releases. The master branch of greenlet is not released yet. Maybe we change should be re-done in a greenlet "stable" branch, I didn't check if it exists.

[brandtbucher]

Currently, on PyPI, there are only greenlet 1.x releases. The master branch of greenlet is not released yet. Maybe we change should be re-done in a greenlet "stable" branch, I didn't check if it exists.

2.0.0a2 was released yesterday. I believe it includes your changes?

If so, we can just use that. pyperformance already using 2.0.0a1, currently.

[brandtbucher]

The annoying thing is that the benchmarks don't actually use greenlet (it's just a required dependency of SQLAlchemy). For example, just using an old cached 3.11 wheel works fine. 🤦🏼‍♂️

@vstinner, do you want to bump the version in pyperformance, or should I?

[gvanrossum]

Maybe involve @ericsnowcurrently too? He recently (this week) did a bunch of work, but I’m not sure he is done. The psutil thing isn’t fixed on Windows, and —python is broken.

[vstinner]

I no longer maintain pyperformance.

[gvanrossum]

Then who does?

[vstinner]

Then who does?

The latest release was done by @pablogsal: python/pyperformance@034f58b

[gvanrossum]

Interestingly, for me (on Windows) moving to greenlet==2.0.0a2 (from ...a1) the sqlalchemy benchmarks run fine. So perhaps this is done? (Then what is python-greenlet/greenlet#294 about?)

[brandtbucher]

Interestingly, for me (on Windows) moving to greenlet==2.0.0a2 (from ...a1) the sqlalchemy benchmarks run fine. So perhaps this is done? (Then what is python-greenlet/greenlet#294 about?)

python-greenlet/greenlet#294 was the fix shipped in 2.0.0a2.

[gvanrossum]

Now I'm confused. Your open checkbox links to gevent/gevent#1872 which is still open. What is that then?

Anyway, I confirmed that on macOS all BMs work (UPDATE: Also with Python 3.10 and 3.9). So maybe you can close this, or check that box?

[brandtbucher]

Now I'm confused. Your open checkbox links to gevent/gevent#1872 which is still open. What is that then?

Ah, I think that might be actually breaking the Pyston macrobenchmarks, not pyperformance. I might have added it here since the cause was the same as the others, and there was an open PR to fix it. I can track it elsewhere.

Closing, since I've confirmed locally (Linux) that the main branch of pyperformance is now working. We still need a new release, though.