如何修复“致命的 Python 错误：_enter_buffered_busy：无法在解释器关闭时获取 <_io.BufferedWriter name='<stdout>'> 的锁”错误？

Question

我目前正在用 python 编写一个应用程序，它使用多线程同时下载多个文件。我对 Python 中的多线程并不完全熟悉或了解，所以我不确定应该提供多少或哪些信息。

当下载程序几乎完成文件下载时，它崩溃了，我收到此错误消息。

fatal Python error: _enter_buffered_busy: could not aquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads
Python runtime state: finalizing (tstate=000001EF6024D440)

Current thread 0x00002cf8 (most recent call first):
<no Python frame>

我对此进行了多次测试，并注意到以下几点：

• 此错误并不总是出现。大多数时候，我对应用程序没有任何问题
• 我已经尝试过并收到错误
  • 运行 Python3.6 的 Linux Ubuntu 18.04（笔记本电脑）
  • Linux Ubuntu 20.04（桌面）运行Python3.6
  • Windows 10（桌面版）
  • 运行 Ubuntu / python3.6-slim-buster 基础镜像的 Docker 容器

很抱歉，我必须将所有这些代码转储到 SO 上；我真的不知道这个错误意味着什么，也不知道从哪里开始排除故障。我们的存储库位于这里

这是处理多线程的代码：

打打

import datetime
import json
import multiprocessing
import os
from pathlib import Path
from queue import Queue
from threading import Thread
from typing import List, Dict

try:
    from dozent.downloader_tools import DownloaderTools
    from dozent.progress_tracker import ProgressTracker
except ModuleNotFoundError:
    from downloader_tools import DownloaderTools
    from progress_tracker import ProgressTracker

CURRENT_FILE_PATH = Path(__file__)
DEFAULT_DATA_DIRECTORY = CURRENT_FILE_PATH.parent.parent / 'data'
TWITTER_ARCHIVE_STREAM_LINKS_PATH = CURRENT_FILE_PATH.parent / 'twitter-archive-stream-links.json'

FIRST_DAY_OF_SUPPORT = datetime.date(2017, 6, 1)
LAST_DAY_OF_SUPPORT = datetime.date(2020, 6, 30)

class _DownloadWorker(Thread):  # skip_tests

    def __init__(self, queue: Queue, download_dir: Path, tracker: ProgressTracker = None):
        Thread.__init__(self)
        self.queue = queue
        self.download_dir = download_dir
        self.tracker = tracker

    def run(self):
        while True:
            # Get the work from the queue and expand the tuple
            link = self.queue.get()
            try:
                DownloaderTools.download_with_pysmartdl(link, str(self.download_dir), tracker=self.tracker)
            finally:
                self.queue.task_done()

#
# Skipping non-multithreading code
#

    def download_timeframe(self, start_date: datetime.date, end_date: datetime.date, verbose: bool = True,
                           download_dir: Path = DEFAULT_DATA_DIRECTORY):
        """
        Download all tweet archives from self.start_date to self.end_date
        :return: None
        """

        # Create a queue to communicate with the worker threads
        queue = Queue()
        if verbose:
            tracker = ProgressTracker()
            tracker.daemon = True
            tracker.start()
        else:
            tracker = None

        os.makedirs(download_dir, exist_ok=True)

        for x in range(multiprocessing.cpu_count()):
            worker = _DownloadWorker(queue, download_dir, tracker=tracker)
            # worker.set_verbosity(verbose=verbosity)
            # Setting daemon to True will let the main thread exit even though the workers are blocking
            worker.daemon = True
            worker.start()

        for sample_date in self.get_links_for_days(start_date=start_date, end_date=end_date):
            print(f"Queueing tweets download for {sample_date['month']}-{sample_date['year']}")
            queue.put(sample_date['link'])

        queue.join()
        tracker.join()

catch_thread_exections.py

import threading
import sys
import traceback
import os
import signal

"""
usage: install()
Once installed, all exceptions caught from within threads will cause the program to end.
"""


def sendKillSignal(etype, value, tb):
    lines = traceback.format_exception(etype, value, tb)
    for line in lines:
        print(line, flush=True)

    os.kill(os.getpid(), signal.SIGTERM)


original_init = threading.Thread.__init__


def patched_init(self, *args, **kwargs):
    original_init(self, *args, **kwargs)
    original_run = self.run

    def patched_run(*args, **kw):
        try:
            original_run(*args, **kw)
        except Exception:
            sys.excepthook(*sys.exc_info())
    self.run = patched_run


def install():
    sys.excepthook = sendKillSignal
    threading.Thread.__init__ = patched_init

进度跟踪器.py

 
import sys
import os

import numpy as np
import threading
import shutil
import queue

from typing import Dict, Callable, Tuple
from collections import namedtuple

Task = namedtuple('Task', 'id progress_callback progress prefix suffix line_length')


class ProgressTracker(threading.Thread):
    """It is assumed that no other printing to stdout will occur while running this ProgressTracker thread.
    """

    def __init__(self):
        threading.Thread.__init__(self)

        self.tasks: Dict[int, Task] = {}
        self.message_queue = queue.Queue()
        self.sum_progress = 0
        self.next_task = 0

        self.terminal_size = None

        resized = self.refresh_terminal_size()
        assert(resized)  # This won't work if we couldn't read the terminal size

        self.lock = threading.Lock()
        self.tasks_pending = False
        self.running = True

    def refresh_terminal_size(self) -> bool:
        """
        Records any change in the terminal size.
        :returns: True if a change is detected, else False
        """

        old_size = self.terminal_size
        try:
            self.terminal_size = os.get_terminal_size()[0]
        except OSError:
            self.terminal_size = shutil.get_terminal_size()[0]

        return old_size != self.terminal_size

    def join(self):
        self.message_queue.join()
        self.stop()

    def stop(self):
        self.running = False
        self.update(0)

    def update(self, task_id: int):
        self.message_queue.put(task_id)

    def register_task(self, name: str, progress_callback: Callable[[], Tuple[float, str, str]]) -> int:
        """Register a new task with the given progress callback.
        :param name: A unique name for this task.
        :param progress_callback: A function, which will report a tuple of:
                                    a progress percentage in [0,100],
                                    a string prefix to be output before the progress bar,
                                    and a string suffix to be output after the progress bar.
        :return: The ID of the registered task.
        """

        task_id = self.next_task
        self.next_task += 1

        # Perform operation in helper function on own thread, so that this function does not block
        thread = threading.Thread(target=self._register_task_helper, args=(name, task_id, progress_callback))
        thread.daemon = True
        thread.start()

        return task_id

    def _register_task_helper(self, name: str, task_id: int, progress_callback: Callable[[], Tuple[float, str]]):
        with self.lock:
            # Create the task, with initial progress 0
            self.tasks[task_id] = (Task(id=task_id,
                                        progress_callback=progress_callback,
                                        progress=0,
                                        prefix=None,
                                        suffix=None,
                                        line_length=0))
            self.tasks_pending = True
            self._delete_progress_bars()
            print(f"Queueing tweet download {task_id}: {name}", flush=True)

    def _delete_progress_bars(self, starting_id: int = 0):
        if self.tasks:
            sys.stdout.write('\b' * (np.sum([self.tasks[i].line_length for i in range(starting_id, len(self.tasks))])))

    def _update_task_progress(self, task_id: int, draw: bool = True):
        """Update the progress bar for the given task
        :param task_id: Identifier of the task to update.
        :param draw: Whether to draw the updated progress bar, defaults to True
        """

        # Update the progress value
        task = self.tasks[task_id]
        assert(task.id == task_id)

        progress, prefix, suffix = task.progress_callback()
        assert(0 <= progress <= 100)

        self.sum_progress += progress - task.progress
        task = Task(id=task_id,
                    progress_callback=task.progress_callback,
                    progress=progress,
                    prefix=prefix,
                    suffix=suffix,
                    line_length=task.line_length)
        self.tasks[task_id] = task

        if draw:
            # Redraw all bars if the terminal size changes
            if self.refresh_terminal_size():
                task_id = 0

            # Return to start of the progress bar for task_id
            self._delete_progress_bars(starting_id=task_id)
            self._draw_task_progress(task_id)

    def _draw_task_progress(self, task_id: int):
        """Updates the progress bar for task_id and after.
        Assumes the console begins at the end of any previously drawn progress bars.
        """

        # We re-write the progress for all tasks, in order to finish at the end of all progress bars (our assumption)
        for task in [self.tasks[i] for i in range(task_id, len(self.tasks))]:
            if task.prefix is None or task.suffix is None:
                continue

            prefix = f"{task.id:3.0f}: "

            bar = ' '*len(prefix)
            bar_width = self.terminal_size - len(prefix) - 2
            assert(bar_width > 0)

            num_dashes = int(task.progress * bar_width / 100.0)
            bar += f"[{'-' * num_dashes}{' ' * (bar_width - num_dashes)}]"

            prefix += task.prefix[:int(self.terminal_size/2)]
            suffix = task.suffix[:int(self.terminal_size/3)]
            spaces = ' ' * (self.terminal_size - len(suffix) - len(prefix))

            line = f"{prefix}{spaces}{suffix}{bar}"

            assert(self.tasks[task.id] == task)
            self.tasks[task.id] = Task(id=task.id,
                                       progress_callback=task.progress_callback,
                                       progress=task.progress,
                                       prefix=task.prefix,
                                       suffix=task.suffix,
                                       line_length=len(line))

            sys.stdout.write(line)

        sys.stdout.flush()

    def _run_task(self):
        task_id = self.message_queue.get(True)

        try:
            with self.lock:
                if self.tasks_pending:
                    self._update_task_progress(0, draw=False)
                    self.tasks_pending = False

                else:
                    self._update_task_progress(task_id, draw=True)

        finally:
            self.message_queue.task_done()

    def run(self):
        while self.running:
            self._run_task()

        self._delete_progress_bars()
        print('')

再说一遍，我不是 Python 多线程方面的专家，并且不确定从哪里开始进行故障排除。我将非常感谢任何提示！

编辑：

我提供的示例不可重现

如何重现

拉取 Docker 镜像：

docker pull socialmediapublicanalysis/dozent:latest

下载一天（2020 年 5 月 12 日）的所有推文：

docker run -it socialmediapublicanalysis/dozent:latest python -m dozent -s 2020-05-12 -e 2020-05-12

Answer 1

您的代码中发生了很多事情，这绝对不是一个最小的、可重现的示例......

也就是说，根据错误消息，您似乎正在使用守护线程，并且已知如果它们stdout在关闭时尝试打印 (to ) ，则会导致麻烦（总而言之：您可能正在尝试写入不再存在stdout守护线程）。

您应该确保.join()所有线程（并且，如果预计它们会长时间运行，请为它们配备一个关闭方法，该方法会翻转一个标志，然后由您在其中进入的任何循环进行检查；或者发送一个毒药通过工作队列的药丸）。

多线程下载的更好方法是使用文档中的ThreadPoolExecutor示例。

如果您想要进度报告，您可能会对 感兴趣tqdm。最后，您的代码将变得更加简单并且更易于维护。

下面是一个完整的示例，它使用 ThreadPoolExecutor 从 Wikipedia 下载大量（截至撰写本文时为 818 个）url，并显示进度：

import requests
from bs4 import BeautifulSoup
from urllib.parse import urljoin
from concurrent.futures import ThreadPoolExecutor
from tqdm import tqdm


tld = 'https://en.wikipedia.org'
url = '/wiki/Index_of_computing_articles'

soup = BeautifulSoup(requests.get(urljoin(tld, url)).content)
urllist = [urljoin(tld, a.get('href')) for a in soup.find_all(href=True)]

def load_url(url):
    return requests.get(url).content

def load_all(urllist):
    with ThreadPoolExecutor() as executor:
        results = list(tqdm(
            executor.map(load_url, urllist),
            total=len(urllist), unit=' pages'))
    return results

results = load_all(urllist)

运行时，它会显示一个进度条：