如何异步映射/过滤异步迭代？

Question

假设我有一个可以使用async for 传递的异步迭代器，那么我如何将它映射并过滤到一个新的异步迭代器？以下代码改编自我如何使用同步迭代器执行相同操作的方式不起作用，因为 yield 不允许在 async defs 中使用。

async def mapfilter(aiterable, p, func):
    async for payload in aiterable:
        if p(payload):

            # This part isn't allowed, but hopefully it should be clear
            # what I'm trying to accomplish.
            yield func(payload)

Answer 1

一个recently published PEP draft (PEP 525)，其支持是scheduled for Python 3.6，建议允许异步生成器具有与您想出的相同的语法。

同时，如果您不想处理异步迭代器样板，也可以使用CryingCyclops 在其注释中提到的asyncio_extras 库。

来自the docs：

@async_generator
async def mygenerator(websites):
    for website in websites:
        page = await http_fetch(website)
        await yield_async(page)

async def fetch_pages():
    websites = ('http://foo.bar', 'http://example.org')
    async for sanitized_page in mygenerator(websites):
        print(sanitized_page)

还有支持yield from 构造的async_generator library。

Answer 2

你 can't 在协程中使用 yield。为了实现你的想法，我看到的唯一方法是实现Asynchronous Iterator。如果我是对的，那就是这样的：

class MapFilter:
    def __init__(self, aiterable, p, func):
        self.aiterable = aiterable
        self.p = p
        self.func = func

    async def __aiter__(self):
        return self

    async def __anext__(self):
        while True:
            payload = await self.aiterable.__anext__()  # StopAsyncIteration would be raise here on no new values
            if self.p(payload):
                return self.func(payload)

让我们测试一下。这是带有帮助器arange 类的完整示例（我从here 获取）：

import asyncio


class arange:
    def __init__(self, n):
        self.n = n
        self.i = 0

    async def __aiter__(self):
        return self

    async def __anext__(self):
        i = self.i
        self.i += 1
        if self.i <= self.n:
            await asyncio.sleep(0)  # insert yield point
            return i
        else:
            raise StopAsyncIteration


class MapFilter:
    def __init__(self, aiterable, p, func):
        self.aiterable = aiterable
        self.p = p
        self.func = func

    async def __aiter__(self):
        return self

    async def __anext__(self):
        while True:
            payload = await self.aiterable.__anext__()
            if self.p(payload):
                return self.func(payload)


async def main():
    aiterable = arange(5)
    p = lambda x: bool(x>2)
    func = lambda x: x*2

    async for i in MapFilter(aiterable, p, func):
        print(i)

if __name__ == "__main__":
    loop = asyncio.get_event_loop()
    loop.run_until_complete(main())

输出：

6
8

Answer 3

https://gist.github.com/dvdotsenko/d8e0068775ac04b58993f604f122284f

用于 Python 3.6+ 的异步 map 和 filter 实现，专门用于无序返回子任务，以先完成者为准。

from collections import deque
from typing import Any, Callable, Collection, AsyncIterator, Iterator, Union


async def _next(gg):
    # repackaging non-asyncio next() as async-like anext()
    try:
        return next(gg)
    except StopIteration:
        raise StopAsyncIteration


async def _aionext(gg):
    # there is no anext() :(
    return await gg.__anext__()


async def map_unordered(fn:Callable, args:Union[Iterator,Collection,AsyncIterator], maxsize=None):
    """
    Async generator yielding return values of resolved invocations
    of `fn` against arg in args list

    Arguments are consumed and fed to callable in the order they are presented in args.
    Results are yielded NOT in order of args. Earliest done is yielded.

    If `size` is specified, worker tasks pool is constrained to that size.

    This is asyncio equivalent of Gevent's `imap_unordered(fn, args_iterable, pool_size)`
    http://www.gevent.org/api/gevent.pool.html#gevent.pool.Group.imap_unordered

    `args` may be Async Iterator or regular Iterator. 
     Thus, you can chain `map_unordered` as `args` for another `map_unordered`

    Because this is an async generator, cannot consume it as regular iterable.
    Must use `async for`.

    Usage example:

            # note NO await in this assignment
            gen = map_unordered(fn, arguments_iter, maxsize=3)
            async for returned_value in gen:
                yield returned_value

    """
    if maxsize == 0:
        raise ValueError(
            'Argument `maxsize` cannot be set to zero. '
            'Use `None` to indicate no limit.'
        )

    # Make args list consumable like a generator
    # so repeated islice(args, size) calls against `args` move slice down the list.

    if hasattr(args, '__anext__'):
        n = _aionext
    elif hasattr(args, '__next__'):
        n = _next
    else:
        args = iter(args)
        n = _next

    have_args = True  # assumed. Don't len(args).
    pending_tasks = deque()

    while have_args or len(pending_tasks):
        try:
            while len(pending_tasks) != maxsize:
                arg = await n(args)
                pending_tasks.append(
                    asyncio.Task(fn(arg))
                )
        except StopAsyncIteration:
            have_args = False

        if not len(pending_tasks):
            return

        done, pending_tasks = await asyncio.wait(pending_tasks, return_when=asyncio.FIRST_COMPLETED)
        pending_tasks = deque(pending_tasks)

        for task in done:
            yield await task  # await converts task object into its return value


async def _filter_wrapper(fn, arg):
    return (await fn(arg)), arg

async def _filter_none(arg):
    return not (arg is None)

async def filter_unordered(fn:Union[Callable,None], args:Union[Iterator,Collection,AsyncIterator], maxsize=None):
    """
    Async filter generator yielding values of `args` collection that match filter condition.
    Like python's native `filter([Callable|None], iterable)` but:
    - allows iterable to be async iterator
    - allows callable to be async callable
    - returns results OUT OF ORDER - whichever passes filter test first.

    Arguments are consumed and fed to callable in the order they are presented in args.
    Results are yielded NOT in order of args. Earliest done and passing the filter condition is yielded.

    If `maxsize` is specified, worker tasks pool is constrained to that size.

    This is inspired by Gevent's `imap_unordered(fn, args_iterable, pool_size)`
    http://www.gevent.org/api/gevent.pool.html#gevent.pool.Group.imap_unordered

    Because this is an async generator, cannot consume it as regular iterable.
    Must use `async for`.

    Usage example:

            # note NO await in this assignment
            gen = filter_unordered(fn, arguments_iter, maxsize=3)
            async for returned_value in gen:
                yield returned_value

    """
    if maxsize == 0:
        raise ValueError(
            'Argument `maxsize` cannot be set to zero. '
            'Use `None` to indicate no limit.'
        )

    if hasattr(args, '__anext__'):
        n = _aionext
    elif hasattr(args, '__next__'):
        n = _next
    else:
        args = iter(args)
        n = _next

    if fn is None:
        fn = _filter_none

    have_args = True  # assumed. Don't len(args).
    pending_tasks = deque()

    while have_args or len(pending_tasks):
        try:
            while len(pending_tasks) != maxsize:
                arg = await n(args)
                pending_tasks.append(
                    asyncio.Task(_filter_wrapper(fn,arg))
                )
        except StopAsyncIteration:
            have_args = False

        if not len(pending_tasks):
            return

        done, pending_tasks = await asyncio.wait(pending_tasks, return_when=asyncio.FIRST_COMPLETED)
        pending_tasks = deque(pending_tasks)

        for task in done:
            filter_match, arg = await task
            if filter_match:
                yield arg

像 Gevent 的 imap_unordered 一样工作，但与 Gevent 的版本不同的是，它还允许 args 可迭代成为异步值生成器。意味着您可以将这些链接起来。

给定：

async def worker(seconds):
    print('> Start wait', seconds)
    await asyncio.sleep(seconds)
    print('< End wait', seconds)
    return seconds


async def to_aio_gen(ll):
    for e in ll:
        yield e

async def test_map(ll, size=None):
    t = time.time()
    async for v in map_unordered(worker, ll, maxsize=size):
        print('-- elapsed second', round(time.time() - t, 1), ' received value', v)


ll = [
    0.2,
    0.4,
    0.8,
    1.2,
    1.1,
    0.3,
    0.6,
    0.9,
]

测试输出：

非异步可迭代，池大小 = 3

>>> asyncio.run(test_map(ll, 3))
> Start wait 0.2
> Start wait 0.4
> Start wait 0.8
< End wait 0.2
-- elapsed second 0.2  received value 0.2
> Start wait 1.2
< End wait 0.4
-- elapsed second 0.4  received value 0.4
> Start wait 1.1
< End wait 0.8
-- elapsed second 0.8  received value 0.8
> Start wait 0.3
< End wait 0.3
-- elapsed second 1.1  received value 0.3
> Start wait 0.6
< End wait 1.2
-- elapsed second 1.4  received value 1.2
> Start wait 0.9
< End wait 1.1
-- elapsed second 1.5  received value 1.1
< End wait 0.6
-- elapsed second 1.7  received value 0.6
< End wait 0.9
-- elapsed second 2.3  received value 0.9

异步迭代器作为 arg 列表，池大小 = 3，过滤器

async def more_than_half(v):
    await asyncio.sleep(v)
    return v > 0.5

>>> asyncio.run(filter_unordered(more_than_half, aio_gen(ll), 3))
-- elapsed second 0.8  received value 0.8
-- elapsed second 1.4  received value 1.2
-- elapsed second 1.5  received value 1.1
-- elapsed second 1.7  received value 0.6
-- elapsed second 2.3  received value 0.9