Hi,
On Wed, May 26, 2021 at 2:05 AM Neil Schemenauer
- Cython doesn't work because of _PyGen_Send change [1]
My team fixed the Python 3.10 compatibility in the Cython 0.29.x branch, but the latest 0.29.x release (0.29.23, April 14, 2021) doesn't include these fixes yet. A new Cython release is needed. I'm not sure how to help Cython to release a new version. In Fedora, we have downstream patches until a new Cython is released: https://src.fedoraproject.org/rpms/Cython/pull-request/28
- numpy cannot be installed because of _Py_HashDouble() change [2]
"pip install numpy" also fails with an error about "incompatible tag". It's an issue in packaging which is vendored in wheel which is used (as a vendored copy?) by pip to build a local wheel package of numpy. numpy doesn't ship wheel packages for Python 3.10 beta releases which is a good idea, the ABI is not stable yet.
Can we do any things to improve the situation? Perhaps using the pre-release functionality of PyPI would help. We would have to somehow encourage 3rd party packages to upload pre-releases that are compatible with our beta/RC releases.
In Fedora, we use a "COPR build": we partially rebuild the OS with Python 3.10, and then we check for package build failures. We are working on switching to Python 3.10 by default: * https://fedoraproject.org/wiki/Changes/Python3.10 * https://bugzilla.redhat.com/show_bug.cgi?id=1890881 For example, my colleague Lumir submitted the trio bug report ;-) https://github.com/python-trio/trio/issues/1899 We are doing our best to report issues to projects. Sometimes we help to fix them, but we cannot fix all compatibility issues! In the past, I tried to work on a small Python script to run the test suite of some projects with the main branch of Python. I had practical issues: * There is no simple way to get install test dependencies of a Python project, especially non-Python dependencies. * When a dependency was not compatible with the next Python, I had to maintain a patch until my fix is accepted upstream. * Sometimes, I had to pull the code from Git rather than using a tarball/ZIP (to not have to wait for a new release). At the end, all these practical issues need solutions which look very closely to a recipe to build a package for a Linux distribution: * Recipe to pull the source. * Recipe to pull build dependencies. * Downstream patches (until a fix is merged upstream). * Recipe to run tests: we try to always run tests when we build a package in Fedora. Advantages of using Fedora COPR: * Reuse Fedora infrastructure (building packages need physical servers to build packages and run tests) * Reuse the public Fedora bug tracker to coordinate the work. * Reuse Fedora "specfiles" (recipe to build a RPM package). * Prepare Fedora to be updated to the next Python. Sadly, Fedora COPR might be a little bit too Fedora specific: require to understand specfiles, how to submit a PR on a specfile, use the Fedora bug tracker, etc. Fedora Rawhide was just updated to switch Python 3.10 by default! /usr/bin/python3 is now Python 3.10 beta2. In COPR we could skip tests to build dependencies and discover more compatibiliy issues. In Rawhide, tests cannot be skipped and so some packages cannot be built yet. For example, astropy has a few failing tests: https://github.com/astropy/astropy/issues/11821 Again, the practical problem is that numpy cannot easily be installed on Python 3.10 (pip compatibiltiy tag issue), and so astropy may wait until Python 3.10 final is released to look into these tests failures. Victor -- Night gathers, and now my watch begins. It shall not end until my death.