How to use `async for` in Python?

Question

I mean what do I get from using async for. Here is the code I write with async for, AIter(10) could be replaced with get_range().

But the code runs like sync not async.

import asyncio

async def get_range():
    for i in range(10):
        print(f"start {i}")
        await asyncio.sleep(1)
        print(f"end {i}")
        yield i

class AIter:
    def __init__(self, N):
        self.i = 0
        self.N = N

    def __aiter__(self):
        return self

    async def __anext__(self):
        i = self.i
        print(f"start {i}")
        await asyncio.sleep(1)
        print(f"end {i}")
        if i >= self.N:
            raise StopAsyncIteration
        self.i += 1
        return i

async def main():
    async for p in AIter(10):
        print(f"finally {p}")

if __name__ == "__main__":
    asyncio.run(main())

The result I excepted should be :

start 1
start 2
start 3
...
end 1
end 2
...
finally 1
finally 2
...

However, the real result is:

start 0
end 0
finally 0
start 1
end 1
finally 1
start 2
end 2

I know I could get the excepted result by using asyncio.gather or asyncio.wait.

But it is hard for me to understand what I got by use async for here instead of simple for.

What is the right way to use async for if I want to loop over several Feature object and use them as soon as one is finished. For example:

async for f in feature_objects:
    data = await f
    with open("file", "w") as fi:
        fi.write()

Answer 1

But it is hard for me to understand what I got by use async for here instead of simple for.

The underlying misunderstanding is expecting async for to automatically parallelize the iteration. It doesn't do that, it simply allows sequential iteration over an async source. For example, you can use async for to iterate over lines coming from a TCP stream, messages from a websocket, or database records from an async DB driver.

None of the above would work with an ordinary for, at least not without blocking the event loop. This is because for calls __next__ as a blocking function and doesn't await its result. You cannot manually await elements obtained by for because for expects __next__ to signal the end of iteration by raising StopIteration. If __next__ is a coroutine, the StopIteration exception won't be visible before awaiting it. This is why async for was introduced, not just in Python, but also in other languages with async/await and generalized for.

If you want to run the loop iterations in parallel, you need to start them as parallel coroutines and use asyncio.as_completed or equivalent to retrieve their results as they come:

async def x(i):
    print(f"start {i}")
    await asyncio.sleep(1)
    print(f"end {i}")
    return i

# run x(0)..x(10) concurrently and process results as they arrive
for f in asyncio.as_completed([x(i) for i in range(10)]):
    result = await f
    # ... do something with the result ...

If you don't care about reacting to results immediately as they arrive, but you need them all, you can make it even simpler by using asyncio.gather:

# run x(0)..x(10) concurrently and process results when all are done
results = await asyncio.gather(*[x(i) for i in range(10)])

Answer 2

(Adding on the accepted answer - for Charlie's bounty).

Assuming you want to consume each yielded value concurrently, a straightforward way would be:

import asyncio

async def process_all():
    tasks = []

    async for obj in my_async_generator:
        # Python 3.7+. Use ensure_future for older versions.
        task = asyncio.create_task(process_obj(obj))
        tasks.append(task)
    
    await asyncio.gather(*tasks)


async def process_obj(obj):
    ...

Explanation:

Consider the following code, without create_task:

async def process_all():
    async for obj in my_async_generator:
        await process_obj(obj))

This is roughly equivalent to:

async def process_all():
    obj1 = await my_async_generator.__anext__():
    await process_obj(obj1))

    obj2 = await my_async_generator.__anext__():
    await process_obj(obj1))
    
    ...

Basically, the loop cannot continue because its body is blocking. The way to go is to delegate the processing of each iteration to a new asyncio task which will start without blocking the loop. The, gather wait for all of the tasks - which means, for every iteration to be processed.

Answer 3

As others have pointed out, async for doesn't create tasks to be run concurrently. It's used to allows sequential iteration over an async source.

As an example, in aiokafka, you could do async for msg in consumer. The __anext__ method in consumer is called in each iteration. This method is defined as async def __anext__ , allowing a call await self.get_one() inside it.

In comparision, when you use a normal for loop, it internally invokes the __next__ special method. However, the regular __next__ method does not support waiting for an async source, such as using await get_one().

Answer 4

Code based on fantastic answer from @matan129, just missing the async generator to make it runnable, once I have that (or if someone wants to contributed it) will finilize this:

import time

import asyncio


async def process_all():
    """
    Example where the async for loop allows to loop through concurrently many things without blocking on each individual
    iteration but blocks (waits) for all tasks to run.
    ref:
    - https://stackoverflow.com/questions/56161595/how-to-use-async-for-in-python/72758067#72758067
    """
    tasks = []

    async for obj in my_async_generator:
        # Python 3.7+. Use ensure_future for older versions.
        task = asyncio.create_task(process_obj(obj))  # concurrently dispatches a coroutine to be executed.
        tasks.append(task)

    await asyncio.gather(*tasks)


async def process_obj(obj):
    await asyncio.sleep(5)  # expensive IO


if __name__ == '__main__':
    # - test asyncio
    s = time.perf_counter()
    asyncio.run(process_all())
    # - print stats
    elapsed = time.perf_counter() - s
    print(f"{__file__} executed in {elapsed:0.2f} seconds.")
    print('Success, done!\a')