下面是我的代码，用于捕获子进程ASAP的每个输出，包括部分行。它同时泵浦，stdout和stderr的顺序几乎是正确的。

在Python 2.7 linux和windows上测试并正确工作。

#!/usr/bin/python
#
# Runner with stdout/stderr catcher
#
from sys import argv
from subprocess import Popen, PIPE
import os, io
from threading import Thread
import Queue
def __main__():
    if (len(argv) > 1) and (argv[-1] == "-sub-"):
        import time, sys
        print "Application runned!"
        time.sleep(2)
        print "Slept 2 second"
        time.sleep(1)
        print "Slept 1 additional second",
        time.sleep(2)
        sys.stderr.write("Stderr output after 5 seconds")
        print "Eol on stdin"
        sys.stderr.write("Eol on stderr\n")
        time.sleep(1)
        print "Wow, we have end of work!",
    else:
        os.environ["PYTHONUNBUFFERED"]="1"
        try:
            p = Popen( argv + ["-sub-"],
                       bufsize=0, # line-buffered
                       stdin=PIPE, stdout=PIPE, stderr=PIPE )
        except WindowsError, W:
            if W.winerror==193:
                p = Popen( argv + ["-sub-"],
                           shell=True, # Try to run via shell
                           bufsize=0, # line-buffered
                           stdin=PIPE, stdout=PIPE, stderr=PIPE )
            else:
                raise
        inp = Queue.Queue()
        sout = io.open(p.stdout.fileno(), 'rb', closefd=False)
        serr = io.open(p.stderr.fileno(), 'rb', closefd=False)
        def Pump(stream, category):
            queue = Queue.Queue()
            def rdr():
                while True:
                    buf = stream.read1(8192)
                    if len(buf)>0:
                        queue.put( buf )
                    else:
                        queue.put( None )
                        return
            def clct():
                active = True
                while active:
                    r = queue.get()
                    try:
                        while True:
                            r1 = queue.get(timeout=0.005)
                            if r1 is None:
                                active = False
                                break
                            else:
                                r += r1
                    except Queue.Empty:
                        pass
                    inp.put( (category, r) )
            for tgt in [rdr, clct]:
                th = Thread(target=tgt)
                th.setDaemon(True)
                th.start()
        Pump(sout, 'stdout')
        Pump(serr, 'stderr')

        while p.poll() is None:
            # App still working
            try:
                chan,line = inp.get(timeout = 1.0)
                if chan=='stdout':
                    print "STDOUT>>", line, "<?<"
                elif chan=='stderr':
                    print " ERROR==", line, "=?="
            except Queue.Empty:
                pass
        print "Finish"

if __name__ == '__main__':
    __main__()

Python 3.4为异步IO引入了新的临时API——asyncio模块。

该方法类似于@Bryan Ward的twisted-based回答——定义一个协议，一旦数据准备好，就调用它的方法:

#!/usr/bin/env python3
import asyncio
import os

class SubprocessProtocol(asyncio.SubprocessProtocol):
    def pipe_data_received(self, fd, data):
        if fd == 1: # got stdout data (bytes)
            print(data)

    def connection_lost(self, exc):
        loop.stop() # end loop.run_forever()

if os.name == 'nt':
    loop = asyncio.ProactorEventLoop() # for subprocess' pipes on Windows
    asyncio.set_event_loop(loop)
else:
    loop = asyncio.get_event_loop()
try:
    loop.run_until_complete(loop.subprocess_exec(SubprocessProtocol, 
        "myprogram.exe", "arg1", "arg2"))
    loop.run_forever()
finally:
    loop.close()

请参阅文档中的“Subprocess”。

有一个高级接口asyncio.create_subprocess_exec()，它返回允许使用StreamReader.readline()协程异步读取一行的Process对象 (使用async/await Python 3.5+语法):

#!/usr/bin/env python3.5
import asyncio
import locale
import sys
from asyncio.subprocess import PIPE
from contextlib import closing

async def readline_and_kill(*args):
    # start child process
    process = await asyncio.create_subprocess_exec(*args, stdout=PIPE)

    # read line (sequence of bytes ending with b'\n') asynchronously
    async for line in process.stdout:
        print("got line:", line.decode(locale.getpreferredencoding(False)))
        break
    process.kill()
    return await process.wait() # wait for the child process to exit


if sys.platform == "win32":
    loop = asyncio.ProactorEventLoop()
    asyncio.set_event_loop(loop)
else:
    loop = asyncio.get_event_loop()

with closing(loop):
    sys.exit(loop.run_until_complete(readline_and_kill(
        "myprogram.exe", "arg1", "arg2")))

Readline_and_kill()执行以下任务:

启动子进程，将其标准输出重定向到管道 异步从子进程的stdout中读取一行 杀子流程 等待它退出

如果需要，每个步骤都可以被超时秒限制。

选择模块帮助您确定下一个有用的输入在哪里。

但是，您几乎总是更喜欢使用单独的线程。一个是阻塞读取stdin，另一个是阻塞你不想阻塞的地方。

现有的解决方案不适合我(详情见下文)。最后成功的是使用read(1)实现readline(基于这个答案)。后者不阻塞:

from subprocess import Popen, PIPE
from threading import Thread
def process_output(myprocess): #output-consuming thread
    nextline = None
    buf = ''
    while True:
        #--- extract line using read(1)
        out = myprocess.stdout.read(1)
        if out == '' and myprocess.poll() != None: break
        if out != '':
            buf += out
            if out == '\n':
                nextline = buf
                buf = ''
        if not nextline: continue
        line = nextline
        nextline = None

        #--- do whatever you want with line here
        print 'Line is:', line
    myprocess.stdout.close()

myprocess = Popen('myprogram.exe', stdout=PIPE) #output-producing process
p1 = Thread(target=process_output, args=(myprocess,)) #output-consuming thread
p1.daemon = True
p1.start()

#--- do whatever here and then kill process and thread if needed
if myprocess.poll() == None: #kill process; will automatically stop thread
    myprocess.kill()
    myprocess.wait()
if p1 and p1.is_alive(): #wait for thread to finish
    p1.join()

为什么现有的解决方案不起作用:

Solutions that require readline (including the Queue based ones) always block. It is difficult (impossible?) to kill the thread that executes readline. It only gets killed when the process that created it finishes, but not when the output-producing process is killed. Mixing low-level fcntl with high-level readline calls may not work properly as anonnn has pointed out. Using select.poll() is neat, but doesn't work on Windows according to python docs. Using third-party libraries seems overkill for this task and adds additional dependencies.

一种解决方案是让另一个进程执行对该进程的读取，或者创建一个带有超时的进程线程。

这是一个超时函数的线程版本:

http://code.activestate.com/recipes/473878/

但是，是否需要在stdout传入时读取它? 另一种解决方案可能是将输出转储到一个文件中，然后使用p.t wait()等待进程完成。

f = open('myprogram_output.txt','w')
p = subprocess.Popen('myprogram.exe', stdout=f)
p.wait()
f.close()


str = open('myprogram_output.txt','r').read()

在这里添加这个答案，因为它提供了在Windows和Unix上设置非阻塞管道的能力。

所有的ctypes细节都要感谢@techtonik的回答。

有一个稍微修改过的版本，可以在Unix和Windows系统上使用。

兼容Python3(只需要一些小的修改)。 包括posix版本，并定义了用于这两个版本的异常。

这样你就可以在Unix和Windows代码中使用相同的函数和异常。

# pipe_non_blocking.py (module)
"""
Example use:

    p = subprocess.Popen(
            command,
            stdout=subprocess.PIPE,
            )

    pipe_non_blocking_set(p.stdout.fileno())

    try:
        data = os.read(p.stdout.fileno(), 1)
    except PortableBlockingIOError as ex:
        if not pipe_non_blocking_is_error_blocking(ex):
            raise ex
"""


__all__ = (
    "pipe_non_blocking_set",
    "pipe_non_blocking_is_error_blocking",
    "PortableBlockingIOError",
    )

import os


if os.name == "nt":
    def pipe_non_blocking_set(fd):
        # Constant could define globally but avoid polluting the name-space
        # thanks to: https://stackoverflow.com/questions/34504970
        import msvcrt

        from ctypes import windll, byref, wintypes, WinError, POINTER
        from ctypes.wintypes import HANDLE, DWORD, BOOL

        LPDWORD = POINTER(DWORD)

        PIPE_NOWAIT = wintypes.DWORD(0x00000001)

        def pipe_no_wait(pipefd):
            SetNamedPipeHandleState = windll.kernel32.SetNamedPipeHandleState
            SetNamedPipeHandleState.argtypes = [HANDLE, LPDWORD, LPDWORD, LPDWORD]
            SetNamedPipeHandleState.restype = BOOL

            h = msvcrt.get_osfhandle(pipefd)

            res = windll.kernel32.SetNamedPipeHandleState(h, byref(PIPE_NOWAIT), None, None)
            if res == 0:
                print(WinError())
                return False
            return True

        return pipe_no_wait(fd)

    def pipe_non_blocking_is_error_blocking(ex):
        if not isinstance(ex, PortableBlockingIOError):
            return False
        from ctypes import GetLastError
        ERROR_NO_DATA = 232

        return (GetLastError() == ERROR_NO_DATA)

    PortableBlockingIOError = OSError
else:
    def pipe_non_blocking_set(fd):
        import fcntl
        fl = fcntl.fcntl(fd, fcntl.F_GETFL)
        fcntl.fcntl(fd, fcntl.F_SETFL, fl | os.O_NONBLOCK)
        return True

    def pipe_non_blocking_is_error_blocking(ex):
        if not isinstance(ex, PortableBlockingIOError):
            return False
        return True

    PortableBlockingIOError = BlockingIOError

为了避免读取不完整的数据，我最终编写了自己的readline生成器(它为每一行返回字节字符串)。

它是一个发电机，所以你可以例如…

def non_blocking_readlines(f, chunk=1024):
    """
    Iterate over lines, yielding b'' when nothings left
    or when new data is not yet available.

    stdout_iter = iter(non_blocking_readlines(process.stdout))

    line = next(stdout_iter)  # will be a line or b''.
    """
    import os

    from .pipe_non_blocking import (
            pipe_non_blocking_set,
            pipe_non_blocking_is_error_blocking,
            PortableBlockingIOError,
            )

    fd = f.fileno()
    pipe_non_blocking_set(fd)

    blocks = []

    while True:
        try:
            data = os.read(fd, chunk)
            if not data:
                # case were reading finishes with no trailing newline
                yield b''.join(blocks)
                blocks.clear()
        except PortableBlockingIOError as ex:
            if not pipe_non_blocking_is_error_blocking(ex):
                raise ex

            yield b''
            continue

        while True:
            n = data.find(b'\n')
            if n == -1:
                break

            yield b''.join(blocks) + data[:n + 1]
            data = data[n + 1:]
            blocks.clear()
        blocks.append(data)