[ci] Appveyor builds failing

[jameslamb]

Description

For the last 2 days, Appveyor CI jobs have been failing, blocking development on this project.
I've seen these changes on several, totally-unrelated PRs (e.g. #5503, #5504, #5505, #5506), so I'm fairly sure it's related to that job's configuration and not breaking changes on PRs.

Reproducible example

See any recent appveyor jobs, for example this one from #5506: https://ci.appveyor.com/project/guolinke/lightgbm/builds/44872180/job/sxsvo19536lcd058.

Additional Comments

LightGBM is compiling successfully, but with several errors in the Python unit tests.

C:\Miniconda3-x64\envs\test-env\lib_pypy\_functools.py:80: in __call__
    return self._func(*(self._args + fargs), **fkeywords)
C:\Miniconda3-x64\envs\test-env\lib\site-packages\sklearn\utils\estimator_checks.py:3489: in check_fit_idempotent
    estimator.fit(X_train, y_train)
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\sklearn.py:1011: in fit
    init_model=init_model
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\sklearn.py:806: in fit
    callbacks=callbacks
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\engine.py:223: in train
    booster = Booster(params=params, train_set=train_set)
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\basic.py:2775: in __init__
    train_set.construct()
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\basic.py:1929: in construct
    self.feature_name = self.get_feature_name()
C:\Miniconda3-x64\envs\test-env\lib\site-packages\lightgbm\basic.py:2380: in get_feature_name
    ptr_string_buffers = (ctypes.c_char_p * num_feature)(*map(ctypes.addressof, string_buffers))
C:\Miniconda3-x64\envs\test-env\lib_pypy\_ctypes\array.py:203: in __init__
    self[i] = arg
C:\Miniconda3-x64\envs\test-env\lib_pypy\_ctypes\array.py:229: in __setitem__
    cobj = self._type_.from_param(value)
C:\Miniconda3-x64\envs\test-env\lib_pypy\_ctypes\primitive.py:382: in from_param
    return super(SimpleType, self).from_param(value)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
self = <class 'ctypes.c_char_p'>, value = 547343581888
    def from_param(self, value):
        if isinstance(value, self):
            return value
        try:
            as_parameter = value._as_parameter_
        except AttributeError:
            raise TypeError("expected %s instance instead of %s" % (
>               self.__name__, type(value).__name__))
E           TypeError: expected c_char_p instance instead of int
C:\Miniconda3-x64\envs\test-env\lib_pypy\_ctypes\basics.py:59: TypeError
_____ test_sklearn_integration[LGBMRegressor()-check_fit_check_is_fitted] _____
self = <class 'ctypes.c_char_p'>, value = 547343579440
    def from_param(self, value):
        if isinstance(value, self):
            return value
        try:
>           as_parameter = value._as_parameter_
E           AttributeError: 'int' object has no attribute '_as_parameter_'
C:\Miniconda3-x64\envs\test-env\lib_pypy\_ctypes\basics.py:56: AttributeError

Haven't investigated much yet, just opening this to track.

[jameslamb]

I suspect this is relevant... it looks like these builds are now getting a PyPy version of Python, instead of CPython.

From the full logs (link), it seems that the step installing graphviz and matplotlib

LightGBM/.ci/test_windows.ps1

Lines 54 to 55 in dc4794b

	# matplotlib and python-graphviz have to be installed separately to prevent conda from downgrading to pypy
	conda install -q -y -n $env:CONDA_ENV matplotlib python-graphviz ; Check-Output $?

is once again causing Python to be "downgraded" from a CPython-based build to a PyPy one 🙃

## Package Plan ##
  environment location: C:\Miniconda3-x64\envs\test-env
  added / updated specs:
    - matplotlib
    - python-graphviz
The following packages will be downloaded:
    package                    |            build
    ---------------------------|-----------------
    ...
    pypy3.7-7.3.7              |       h4094faa_3        26.2 MB  conda-forge
    pytest-7.1.1               |   py37h4c0cbd9_0         501 KB  conda-forge
    python-3.7.12              |        0_73_pypy           5 KB  conda-forge
    python-graphviz-0.20.1     |     pyh22cad53_0          35 KB  conda-forge
    python_abi-3.7             |    2_pypy37_pp73           4 KB  conda-forge
    ...
    ------------------------------------------------------------
                                           Total:       179.8 MB
The following NEW packages will be INSTALLED:
  ...
  pypy3.7            conda-forge/win-64::pypy3.7-7.3.7-h4094faa_3
  ...
The following packages will be DOWNGRADED:
  ...
  python                        3.7.12-h900ac77_100_cpython --> 3.7.12-0_73_pypy
  python_abi                                    3.7-2_cp37m --> 3.7-2_pypy37_pp73
  ...

This is not the first time this has happened. See #5197, #5198.

[shiyu1994]

@jameslamb Thanks for opening this issue. Is it possible to resolve this issue for example by installing these two packages with pip instead of conda?

[jameslamb]

Maybe! Although that might bring its own environment-resolution issues. I'm not certain how to solve this, haven't had time to investigate it yet.

[jameslamb]

This was fixed in #5510.

I just accidentally put (fixes #5509) (wrong issue number) in the title.

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.