Azure Pipelines运行matplotlib subplots测试随机报Tcl错误求助

问题表现 #

在Azure Pipelines托管Windows Python 3.9 CI环境运行Python单元测试时，调用plt.subplots会随机抛出_tkinter.TclError: invalid command name "tcl_findLibrary"错误，本地开发环境无法稳定复现。
报错位置前4行存在完全相同的plt.subplots调用可正常执行，错误触发点标记如下：

def test_plot_epoch_segmentation():
    """Test segmentation plots for epochs."""
    n_clusters = 4
    labels = np.random.choice(
        [-1, 0, 1, 2, 3], (len(epochs), epochs.times.size)
    )

    plot_epoch_segmentation(labels, epochs, n_clusters)
    plt.close("all")

    # provide ax
    f, ax = plt.subplots(1, 1)
    plot_epoch_segmentation(labels, epochs, n_clusters, axes=ax)
    plt.close("all")

    # provide cbar_ax
>>  f, cbar_ax = plt.subplots(1, 1)

错误栈信息 #

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
C:\hostedtoolcache\windows\Python\3.9.13\x64\lib\site-packages\matplotlib\pyplot.py:1455: in subplots
    fig = figure(**fig_kw)
C:\hostedtoolcache\windows\Python\3.9.13\x64\lib\site-packages\matplotlib\pyplot.py:808: in figure
C:\hostedtoolcache\windows\Python\3.9.13\x64\lib\site-packages\matplotlib\backends\_backend_tk.py:949: in new_figure_manager_given_figure
    window = tk.Tk(className="matplotlib")
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <tkinter.Tk object .>, screenName = None, baseName = 'pytest'
className = 'matplotlib', useTk = True, sync = False, use = None

    def __init__(self, screenName=None, baseName=None, className='Tk',
                 useTk=True, sync=False, use=None):
        """Return a new Toplevel widget on screen SCREENNAME. A new Tcl interpreter will
        be created. BASENAME will be used for the identification of the profile file (see
        readprofile).
        It is constructed from sys.argv[0] without extensions if None is given. CLASSNAME
        is the name of the widget class."""
        self.master = None
        self.children = {}
        self._tkloaded = False
        # to avoid recursions in the getattr code in case of failure, we
        # ensure that self.tk is always _something_.
        self.tk = None
        if baseName is None:
            import os
            baseName = os.path.basename(sys.argv[0])
            baseName, ext = os.path.splitext(baseName)
            if ext not in ('.py', '.pyc'):
                baseName = baseName + ext
        interactive = False
>       self.tk = _tkinter.create(screenName, baseName, className, interactive, wantobjects, useTk, sync, use)
E       _tkinter.TclError: invalid command name "tcl_findLibrary"

C:\hostedtoolcache\windows\Python\3.9.13\x64\lib\tkinter\__init__.py:2270: TclError

根因 #

这个错误的核心是无GUI（headless）CI环境使用了需要桌面会话的TkAgg绘图后端导致的：

1. Azure托管Windows runner没有常驻的交互桌面会话，Tkinter依赖的Tcl/Tk运行环境初始化本身就存在不稳定问题
2. 第一次plt.subplots调用时Tk解释器首次加载可以正常工作，执行plt.close("all")后Tk实例被销毁，第二次重新初始化Tk解释器时，受CI环境资源回收时序、会话权限影响，会随机出现找不到Tcl库入口的报错，这也是本地有桌面的开发环境无法复现的核心原因

修复方案 #

• 优先方案：全局强制matplotlib使用无GUI的Agg后端，这是CI环境跑绘图类测试的标准配置
  可以在pytest配置文件pytest.ini中添加全局配置，无需修改测试代码：

  [pytest]
  matplotlib_backend = Agg
  

• 如果不想修改pytest配置，也可以在所有测试文件的最开头（导入matplotlib/pyplot之前）添加如下代码手动指定后端：

  import matplotlib
  matplotlib.use("Agg")
  

• 不要在CI测试流程中使用TkAgg、QtAgg这类依赖本地GUI组件的后端，这类后端在无桌面环境下的实例销毁、重建逻辑本身就不稳定，容易出现随机报错。

内容的提问来源于stack exchange，提问作者Mathieu