前一篇blog说明了怎样把采用线程所实现的项目逐步迁移到asyncio方案上面。迁移后的run_tasks协程,可以将多份输入文件通过tail_async协程正确地合并成一份输出文件。
import asyncio
# On Windows, a ProactorEventLoop can't be created within
# threads because it tries to register signal handlers. This
# is a work-around to always use the SelectorEventLoop policy
# instead. See: https://bugs.python.org/issue33792
policy = asyncio.get_event_loop_policy()
policy._loop_factory = asyncio.SelectorEventLoop
async def run_tasks(handles, interval, output_path):
with open(output_path, 'wb') as output:
async def write_async(data):
output.write(data)
tasks = []
for handle in handles:
coro = tail_async(handle, interval, write_async)
task = asyncio.create_task(coro)
tasks.append(task)
await asyncio.gather(*tasks)
但这样写有个大问题,就是针对输出文件所做的open、close以及write操作,全都要放在主线程中执行,而这些操作又需要在程序所处的操作系统执行系统调用,这些调用可能会让事件循环阻塞很长一段时间,导致其他协程没办法推进。这会降低程序的总体响应能力,而且会增加延迟,对于高并发服务器来说,这个问题尤其严重。
调用asyncio.run函数时,把debug参数设为True,可以帮助我们发现这种问题。例如,下面这种写法就能显示出,slow_coroutine协程所执行的系统调用耗时比较长,这可以提醒我们注意,要读取的文件是否已经损坏,或者其中某一行是否读不出来。
import time
async def slow_coroutine():
time.sleep(0.5) # Simulating slow I/O
asyncio.run(slow_coroutine(), debug=True)
为了进一步提升程序的响应能力,我们可以想办法把那些有可能会执行系统调用的操作从程序本身的事件循环里面拿走。例如,新建这样一个Thread子类,让它把那种给输出文件写入数据的操作封装到自己的事件循环里面,这样就不会阻塞程序本身的事件循环了。
其他线程中的协程,可以直接调用这个线程类的write方法,并对该方法做await。其实这个write方法,只不过是把真正负责执行写入操作的那个real_write封装了起来。这种封装方式能够确保线程安全,因此不需要再通过Lock加锁.
然后,我们按照相似的思路,编写真正负责停止本线程的real_stop方法,并把它封装到stop里面,这样的话,其他协程就可以通过stop方法告知本线程应该结束工作。这项操作同样是线程安全的。
另外,还可以定义__aenter__与__aexit__方法,让我们的线程能够用在异步版本的with语句之中,以确保该线程的启动与关闭会安排在适当的时机执行,而不拖慢主事件循环所在的那条线程。
from threading import Thread
class WriteThread(Thread):
def __init__(self, output_path):
super().__init__()
self.output_path = output_path
self.output = None
self.loop = asyncio.new_event_loop()
def run(self):
asyncio.set_event_loop(self.loop)
with open(self.output_path, 'wb') as self.output:
self.loop.run_forever()
# Run one final round of callbacks so the await on
# stop() in another event loop will be resolved.
self.loop.run_until_complete(asyncio.sleep(0))
# Example 4
async def real_write(self, data):
self.output.write(data)
async def write(self, data):
coro = self.real_write(data)
future = asyncio.run_coroutine_threadsafe(
coro, self.loop)
await asyncio.wrap_future(future)
# Example 5
async def real_stop(self):
self.loop.stop()
async def stop(self):
coro = self.real_stop()
future = asyncio.run_coroutine_threadsafe(
coro, self.loop)
await asyncio.wrap_future(future)
# Example 6
async def __aenter__(self):
loop = asyncio.get_event_loop()
await loop.run_in_executor(None, self.start)
return self
async def __aexit__(self, *_):
await self.stop()
这段代码演示了如何结合使用 Python 的 asyncio
模块和线程 (threading.Thread
) 来创建一个在单独线程中运行的异步事件循环。这在需要在异步环境中执行 I/O 操作,但又不想阻塞主事件循环时非常有用。下面是对代码中关键部分的解释:
WriteThread
类 (继承自 Thread
)初始化 (__init__
方法):
output_path
,这是输出文件的路径。self.loop
创建了一个新的异步事件循环。run
方法:
self.loop.run_forever()
使事件循环持续运行,直到调用 stop
方法。self.loop.run_until_complete(asyncio.sleep(0))
确保事件循环能够完成所有挂起的协程。real_write
协程:
write
协程:
asyncio.run_coroutine_threadsafe
安全地从另一个线程运行 real_write
协程。real_write
完成,并使用 await asyncio.wrap_future(future)
将 Future
对象包装为协程。real_stop
协程:
self.loop.stop()
来停止事件循环。stop
协程:
write
方法,它使用 asyncio.run_coroutine_threadsafe
来安排 real_stop
在事件循环中运行。__aenter__
协程:
__enter__
方法,允许使用 with
语句来管理 WriteThread
对象的生命周期。await
来启动线程。__aexit__
协程:
__exit__
方法,用于清理操作,比如停止事件循环。Thread
的子类中创建和运行自己的事件循环,可以在不阻塞主线程的情况下执行异步操作。asyncio.run_coroutine_threadsafe
可以在不同的线程中安排协程的执行。__aenter__
和 __aexit__
方法,WriteThread
对象可以在 with
语句中使用,这提供了一种优雅的资源管理方式。这种模式允许你将异步 I/O 操作与线程结合使用,充分利用 asyncio
的优势,同时避免在 I/O 密集型操作中阻塞主事件循环。
写好了新的线程类之后,我们可以重构run_tasks,把它变成纯粹的异步版本。这个版本更易读懂,而且完全避免了那些耗时较长的系统调用把主事件循环所在的线程拖慢。
class NoNewData(Exception):
pass
def readline(handle):
offset = handle.tell()
handle.seek(0, 2)
length = handle.tell()
if length == offset:
raise NoNewData
handle.seek(offset, 0)
return handle.readline()
async def tail_async(handle, interval, write_func):
loop = asyncio.get_event_loop()
while not handle.closed:
try:
line = await loop.run_in_executor(
None, readline, handle)
except NoNewData:
await asyncio.sleep(interval)
else:
await write_func(line)
async def run_fully_async(handles, interval, output_path):
async with WriteThread(output_path) as output:
tasks = []
for handle in handles:
coro = tail_async(handle, interval, output.write)
task = asyncio.create_task(coro)
tasks.append(task)
await asyncio.gather(*tasks)
现在验证这样写是否正确。我们把一批输入文件所对应的句柄放在handles里面,交给run_fully_async去合并,然后调用confirm_merge函数,以确认这些文件之中的内容,已经合并到了输出文件里面。
import collections
import os
import random
import string
from tempfile import TemporaryDirectory
def write_random_data(path, write_count, interval):
with open(path, 'wb') as f:
for i in range(write_count):
time.sleep(random.random() * interval)
letters = random.choices(
string.ascii_lowercase, k=10)
data = f'{path}-{i:02}-{"".join(letters)}\n'
f.write(data.encode())
f.flush()
def start_write_threads(directory, file_count):
paths = []
for i in range(file_count):
path = os.path.join(directory, str(i))
with open(path, 'w'):
# Make sure the file at this path will exist when
# the reading thread tries to poll it.
pass
paths.append(path)
args = (path, 10, 0.1)
thread = Thread(target=write_random_data, args=args)
thread.start()
return paths
def close_all(handles):
time.sleep(1)
for handle in handles:
handle.close()
def setup():
tmpdir = TemporaryDirectory()
input_paths = start_write_threads(tmpdir.name, 5)
handles = []
for path in input_paths:
handle = open(path, 'rb')
handles.append(handle)
Thread(target=close_all, args=(handles,)).start()
output_path = os.path.join(tmpdir.name, 'merged')
return tmpdir, input_paths, handles, output_path
# Example 9
def confirm_merge(input_paths, output_path):
found = collections.defaultdict(list)
with open(output_path, 'rb') as f:
for line in f:
for path in input_paths:
if line.find(path.encode()) == 0:
found[path].append(line)
expected = collections.defaultdict(list)
for path in input_paths:
with open(path, 'rb') as f:
expected[path].extend(f.readlines())
for key, expected_lines in expected.items():
found_lines = found[key]
assert expected_lines == found_lines
input_paths = ...
handles = ...
output_path = ...
tmpdir, input_paths, handles, output_path = setup()
asyncio.run(run_fully_async(handles, 0.1, output_path))
confirm_merge(input_paths, output_path)
tmpdir.cleanup()
完整代码:
# Example 1
import asyncio
# On Windows, a ProactorEventLoop can't be created within
# threads because it tries to register signal handlers. This
# is a work-around to always use the SelectorEventLoop policy
# instead. See: https://bugs.python.org/issue33792
policy = asyncio.get_event_loop_policy()
policy._loop_factory = asyncio.SelectorEventLoop
async def run_tasks(handles, interval, output_path):
with open(output_path, 'wb') as output:
async def write_async(data):
output.write(data)
tasks = []
for handle in handles:
coro = tail_async(handle, interval, write_async)
task = asyncio.create_task(coro)
tasks.append(task)
await asyncio.gather(*tasks)
# Example 2
import time
async def slow_coroutine():
time.sleep(0.5) # Simulating slow I/O
asyncio.run(slow_coroutine(), debug=True)
# Example 3
from threading import Thread
class WriteThread(Thread):
def __init__(self, output_path):
super().__init__()
self.output_path = output_path
self.output = None
self.loop = asyncio.new_event_loop()
def run(self):
asyncio.set_event_loop(self.loop)
with open(self.output_path, 'wb') as self.output:
self.loop.run_forever()
# Run one final round of callbacks so the await on
# stop() in another event loop will be resolved.
self.loop.run_until_complete(asyncio.sleep(0))
async def real_write(self, data):
self.output.write(data)
async def write(self, data):
coro = self.real_write(data)
future = asyncio.run_coroutine_threadsafe(
coro, self.loop)
await asyncio.wrap_future(future)
async def real_stop(self):
self.loop.stop()
async def stop(self):
coro = self.real_stop()
future = asyncio.run_coroutine_threadsafe(
coro, self.loop)
await asyncio.wrap_future(future)
async def __aenter__(self):
loop = asyncio.get_event_loop()
await loop.run_in_executor(None, self.start)
return self
async def __aexit__(self, *_):
await self.stop()
class NoNewData(Exception):
pass
def readline(handle):
offset = handle.tell()
handle.seek(0, 2)
length = handle.tell()
if length == offset:
raise NoNewData
handle.seek(offset, 0)
return handle.readline()
async def tail_async(handle, interval, write_func):
loop = asyncio.get_event_loop()
while not handle.closed:
try:
line = await loop.run_in_executor(
None, readline, handle)
except NoNewData:
await asyncio.sleep(interval)
else:
await write_func(line)
async def run_fully_async(handles, interval, output_path):
async with WriteThread(output_path) as output:
tasks = []
for handle in handles:
coro = tail_async(handle, interval, output.write)
task = asyncio.create_task(coro)
tasks.append(task)
await asyncio.gather(*tasks)
# This is all code to simulate the writers to the handles
import collections
import os
import random
import string
from tempfile import TemporaryDirectory
def write_random_data(path, write_count, interval):
with open(path, 'wb') as f:
for i in range(write_count):
time.sleep(random.random() * interval)
letters = random.choices(
string.ascii_lowercase, k=10)
data = f'{path}-{i:02}-{"".join(letters)}\n'
f.write(data.encode())
f.flush()
def start_write_threads(directory, file_count):
paths = []
for i in range(file_count):
path = os.path.join(directory, str(i))
with open(path, 'w'):
# Make sure the file at this path will exist when
# the reading thread tries to poll it.
pass
paths.append(path)
args = (path, 10, 0.1)
thread = Thread(target=write_random_data, args=args)
thread.start()
return paths
def close_all(handles):
time.sleep(1)
for handle in handles:
handle.close()
def setup():
tmpdir = TemporaryDirectory()
input_paths = start_write_threads(tmpdir.name, 5)
handles = []
for path in input_paths:
handle = open(path, 'rb')
handles.append(handle)
Thread(target=close_all, args=(handles,)).start()
output_path = os.path.join(tmpdir.name, 'merged')
return tmpdir, input_paths, handles, output_path
def confirm_merge(input_paths, output_path):
found = collections.defaultdict(list)
with open(output_path, 'rb') as f:
for line in f:
for path in input_paths:
if line.find(path.encode()) == 0:
found[path].append(line)
expected = collections.defaultdict(list)
for path in input_paths:
with open(path, 'rb') as f:
expected[path].extend(f.readlines())
for key, expected_lines in expected.items():
found_lines = found[key]
assert expected_lines == found_lines
input_paths = ...
handles = ...
output_path = ...
tmpdir, input_paths, handles, output_path = setup()
asyncio.run(run_fully_async(handles, 0.1, output_path))
confirm_merge(input_paths, output_path)
tmpdir.cleanup()