Work around Cygwin CI failure #2004, except for test_installation
#2007
Conversation
In case somehow another one leaked in.
Since we don't need pinning, and this avoids installing and using the chocolatey package manager, which we're not using for anything else.
|
Thanks so much for your help with this! Despite the test failure, I think this already is a great improvement to what was there before so I'd merge it on that merit alone. |
My thinking is that it could be marked conditionally xfail when the platform is Cygwin and the |
The previous commit pinned the python39-pip package, but it may still be upgraded when installed automatically in the `python -m venv` step, since that step runs something like `python3.9 -Im ensurepip --upgrade --default-pip`. This replays 4605dd6 to install pip in a separate step. This differs from the solution in gitpython-developers#2007 in that this does not use the bootstrap script.
Together with gitpython-developers#2007, this works around gitpython-developers#2004, allowing all tests to pass on Cygwin CI. In gitpython-developers#2007, installation of the environment in which tests run was fixed by downloading and running the `get-pip.py` bootstrap script. If we were to modify our helper that sets up the (separate) virtual environment in `test_installation` so that it does the same thing (or conditionally does so on CI, since the problem does not seem to happen in local installations), that would likely "fix" this more thoroughly, allowing the test to pass. But part of the goal of the installation test is to test that installation works in a typical environment on the platform it runs on. So it is not obivous that making it pass in that way would be an improvement compared to marking it `xfail` with the exception type that occurs due to gitpython-developers#2004. So this just does that, for now.
|
I tried a few more things, but I did not find a good fix or better workaround, so I've opened #2009 to mark that one test conditionally |
I don't have a complete fix for #2004 yet, but since it's confirmed to fail on the main branch of this upstream repository (mentioned in #1988 (comment), though not due to #1988), I figured my partial fix might be worth putting into place. With these changes, the problem in the Cygwin test job turns from being unable to run any tests to being able to run all the tests but with one new failure compared to the way it was before #2004 (i.e. with one non-xfail failure).
This uses the official
pipbootstrap script https://bootstrap.pypa.io/get-pip.py to installpipin the virtual environment. This allows the actual installation of GitPython and its dependencies, including all test dependencies, to complete without problems. Aside from tests that are expected to fail and are already marked xfail, all tests pass except fortest_installation, which fails for the same reason. It presumably would also pass if the bootstrap script were used there too, but I believe the current intent of that test is to verify that installation succeeds when done in an ordinary way.I am likewise reluctant to mark
test_installationas xfail on Cygwin, unless that turns out to be truly necessary, because none of this happens on my local Cygwin environment on a Windows 10 development machine. It seems like it is triggered by some combination of changing packages and something wrong or otherwise specific about the Cygwin workflow.This PR includes some changes to the workflow that are not actually necessary for the workaround it confers. The changes that are needed are:
venvautomatically attempt to installpip.pipin the virtual environment explicitly using the bootstrap script.wgetCygwin package to download the bootstrap script without extra complexity or risking encoding/line-ending mismatches.All the rest is to make the workflow run a little faster, and maybe be slightly more robust in ways that are not relevant to #2004 and not very important. The most significant of them is switching back to the Cygwin installation action we had previously been using, which is less configurable (it does not support installing Cygwin packages at specific versions, which we had needed to do at one time), but which is faster because it does not first have to install Chocolatey to use it as an intermediary for installing Cygwin.
I kept those changes here because I think it's easier to iterate on this if it is left in. But if you'd like I'd be pleased to undo the nonessential changes, which can always be brought back later and separately. I have tested in 7fd5f15 to make sure that this really works, so undoing those changes could be as simple as adding that one commit.