usedevelop without setup.py failure

[AntoineD]

When usedevelop is enabled with a project that only has a setup.cfg but no setup.py, on the second tox execution with no changes at all, tox crashes because VirtualEnv._needs_reinstall relies on setup.py to get the name of the project's package.

A PR is coming next.

[gaborbernat]

I don't think it's possible to use usedevelop without a setup.py. How does this even passes at first run?

[AntoineD]

I don't think it's possible to use usedevelop without a setup.py. How does this even passes at first run?

This happens with an env that does not run the tests, I'll see what happens otherwise.

[AntoineD]

This works fine when running pytest in the commands of the such an env.

@gaborbernat What do you mean by not possible?

[gaborbernat]

usedevelop implies the installation of the package via setup.py. The installation will fail without setup.py.

[AntoineD]

From what I can see, the installation works fine with only setup.cfg and no setup.py when usedevelop = True.

Have you tried?

[gaborbernat]

Hmm, did setuptools and pip change this? Can you dig up from pip+setuptools changelog when this was added? I know a while back you could not install in develop mode a package, because the interface was the setup.py file itself. It might work now but we need to figure out what versions of pip+setuptools require that 🤔

[AntoineD]

The failures of the PR's CI's jobs seems to be related to what you are saying. Things behave differently on my PC, I'll dig up.

[AntoineD]

This has been fixed in pip 21.1.

[posita]

Just curious, but what's the fix here? Tox should require pip>=21.1?

[AntoineD]

@posita Yes (not considering tox).

[posita]

A PR is coming next.

Is this still true, or should this be closed without a fix?

[AntoineD]

The PR has been closed without merge.

[posita]

I don't think pip fixes this. I still see this with pip==21.2.4.

Digging into this a bit, this works (including on the second and subsequent runs):

[tox]
# …
skipsdist = true

[testenv]
# …
deps =
    --editable .  # <-- this works

This fails on the second run (per the OP):

[tox]
# …
skipsdist = true

[testenv]
# …
usedevelop = True  # <-- this works once, then fails on subsequent runs

[posita]

Here's a test case: posita/test-case-for-tox-2197.

$ tox -e py39,py39-workaround  # first run
# …
  py39: commands succeeded
  py39-workaround: commands succeeded
  congratulations :)
$ tox -e py39,py39-workaround  # second run
# …
ERROR:   py39: InvocationError for command /private/tmp/foo/.tox/py39/bin/python /private/tmp/foo/setup.py --name (exited with code 2)
  py39-workaround: commands succeeded

[ssbarnea]

I am quite interested about fixing this because I already started to remove the historical setup.py from most projects since pip started to allow editable installs without it.

[webknjaz]

I don't think pip fixes this. I still see this with pip==21.2.4.

Pip does fix pip install -e . but it's not what tox uses IIRC. I think tox attempts invoking the deprecated¹ setup.py develop

Footnotes

1. See https://blog.ganssle.io/articles/2021/10/setup-py-deprecated.html for the details. ↩

[ssbarnea]

I spent the last couple of hours trying to keep coverage working while dropping the broken usedevelop and switching to --editable . but without any luck so far.

Apparently almost all repo setups I found implement coverage using tox editable installs. Shortly I cannot drop setup.py without breaking coverage. Did anyone managed to find a configuration where this still works?

[gaborbernat]

https:/tox-dev/tox/blob/rewrite/tox.ini#L29 should work

[ssbarnea]

In the end I was able to make it work with ansible/ansible-compat#96 I no longer use pytest-cov as running coverage ... is more reliable.

I am really glad we have a workaround.