How to limit concurrency with Python asyncio?

Question

Let's assume we have a bunch of links to download and each of the link may take a different amount of time to download. And I'm allowed to download using utmost 3 connections only. Now, I want to ensure that I do this efficiently using asyncio.

Here's what I'm trying to achieve: At any point in time, try to ensure that I have atleast 3 downloads running.

Connection 1: 1---------7---9---
Connection 2: 2---4----6-----
Connection 3: 3-----5---8-----

The numbers represent the download links, while hyphens represent Waiting for download.

Here is the code that I'm using right now

from random import randint
import asyncio

count = 0


async def download(code, permit_download, no_concurrent, downloading_event):
    global count
    downloading_event.set()
    wait_time = randint(1, 3)
    print('downloading {} will take {} second(s)'.format(code, wait_time))
    await asyncio.sleep(wait_time)  # I/O, context will switch to main function
    print('downloaded {}'.format(code))
    count -= 1
    if count < no_concurrent and not permit_download.is_set():
        permit_download.set()


async def main(loop):
    global count
    permit_download = asyncio.Event()
    permit_download.set()
    downloading_event = asyncio.Event()
    no_concurrent = 3
    i = 0
    while i < 9:
        if permit_download.is_set():
            count += 1
            if count >= no_concurrent:
                permit_download.clear()
            loop.create_task(download(i, permit_download, no_concurrent, downloading_event))
            await downloading_event.wait()  # To force context to switch to download function
            downloading_event.clear()
            i += 1
        else:
            await permit_download.wait()
    await asyncio.sleep(9)

if __name__ == '__main__':
    loop = asyncio.get_event_loop()
    try:
        loop.run_until_complete(main(loop))
    finally:
        loop.close()

And the output is as expected:

downloading 0 will take 2 second(s)
downloading 1 will take 3 second(s)
downloading 2 will take 1 second(s)
downloaded 2
downloading 3 will take 2 second(s)
downloaded 0
downloading 4 will take 3 second(s)
downloaded 1
downloaded 3
downloading 5 will take 2 second(s)
downloading 6 will take 2 second(s)
downloaded 5
downloaded 6
downloaded 4
downloading 7 will take 1 second(s)
downloading 8 will take 1 second(s)
downloaded 7
downloaded 8

But here are my questions:

1. At the moment, I'm simply waiting for 9 seconds to keep the main function running till the downloads are complete. Is there an efficient way of waiting for the last download to complete before exiting the main function? (I know there's asyncio.wait, but I'll need to store all the task references for it to work)

2. What's a good library that does this kind of task? I know javascript has a lot of async libraries, but what about Python?

Edit: 2. What's a good library that takes care of common async patterns? (Something like async)

Answer 1

If I'm not mistaken you're searching for asyncio.Semaphore. Example of usage:

import asyncio
from random import randint


async def download(code):
    wait_time = randint(1, 3)
    print('downloading {} will take {} second(s)'.format(code, wait_time))
    await asyncio.sleep(wait_time)  # I/O, context will switch to main function
    print('downloaded {}'.format(code))


sem = asyncio.Semaphore(3)


async def safe_download(i):
    async with sem:  # semaphore limits num of simultaneous downloads
        return await download(i)


async def main():
    tasks = [
        asyncio.ensure_future(safe_download(i))  # creating task starts coroutine
        for i
        in range(9)
    ]
    await asyncio.gather(*tasks)  # await moment all downloads done


if __name__ ==  '__main__':
    loop = asyncio.get_event_loop()
    try:
        loop.run_until_complete(main())
    finally:
        loop.run_until_complete(loop.shutdown_asyncgens())
        loop.close()

Output:

downloading 0 will take 3 second(s)
downloading 1 will take 3 second(s)
downloading 2 will take 1 second(s)
downloaded 2
downloading 3 will take 3 second(s)
downloaded 1
downloaded 0
downloading 4 will take 2 second(s)
downloading 5 will take 1 second(s)
downloaded 5
downloaded 3
downloading 6 will take 3 second(s)
downloading 7 will take 1 second(s)
downloaded 4
downloading 8 will take 2 second(s)
downloaded 7
downloaded 8
downloaded 6

An example of async downloading with aiohttp can be found here. Note that aiohttp has a Semaphore equivalent built in, which you can see an example of here. It has a default limit of 100 connections.

Answer 2

I used Mikhail Gerasimov's answer and ended up with this little gem

async def gather_with_concurrency(n, *coros):
    semaphore = asyncio.Semaphore(n)

    async def sem_coro(coro):
        async with semaphore:
            return await coro
    return await asyncio.gather(*(sem_coro(c) for c in coros))

Which you would run instead of normal gather

await gather_with_concurrency(100, *my_coroutines)

Answer 3

Before reading the rest of this answer, please note that the idiomatic way of limiting the number of parallel tasks this with asyncio is using asyncio.Semaphore, as shown in Mikhail's answer and elegantly abstracted in Andrei's answer. This answer contains working, but a bit more complicated ways of achieving the same. I am leaving the answer because in some cases this approach can have advantages over a semaphore, specifically when the work to be done is very large or unbounded, and you cannot create all the coroutines in advance. In that case the second (queue-based) solution is this answer is what you want. But in most regular situations, such as parallel download through aiohttp, you should use a semaphore instead.

---

You basically need a fixed-size pool of download tasks. asyncio doesn't come with a pre-made task pool, but it is easy to create one: simply keep a set of tasks and don't allow it to grow past the limit. Although the question states your reluctance to go down that route, the code ends up much more elegant:

import asyncio, random

async def download(code):
    wait_time = random.randint(1, 3)
    print('downloading {} will take {} second(s)'.format(code, wait_time))
    await asyncio.sleep(wait_time)  # I/O, context will switch to main function
    print('downloaded {}'.format(code))

async def main(loop):
    no_concurrent = 3
    dltasks = set()
    i = 0
    while i < 9:
        if len(dltasks) >= no_concurrent:
            # Wait for some download to finish before adding a new one
            _done, dltasks = await asyncio.wait(
                dltasks, return_when=asyncio.FIRST_COMPLETED)
        dltasks.add(loop.create_task(download(i)))
        i += 1
    # Wait for the remaining downloads to finish
    await asyncio.wait(dltasks)

An alternative is to create a fixed number of coroutines doing the downloading, much like a fixed-size thread pool, and feed them work using an asyncio.Queue. This removes the need to manually limit the number of downloads, which will be automatically limited by the number of coroutines invoking download():

# download() defined as above

async def download_worker(q):
    while True:
        code = await q.get()
        await download(code)
        q.task_done()

async def main(loop):
    q = asyncio.Queue()
    workers = [loop.create_task(download_worker(q)) for _ in range(3)]
    i = 0
    while i < 9:
        await q.put(i)
        i += 1
    await q.join()  # wait for all tasks to be processed
    for worker in workers:
        worker.cancel()
    await asyncio.gather(*workers, return_exceptions=True)

As for your other question, the obvious choice would be aiohttp.

Answer 4

The asyncio-pool library does exactly what you need.

https://pypi.org/project/asyncio-pool/

from asyncio_pool import AioPool

LIST_OF_URLS = ("http://www.google.com", "......")

pool = AioPool(size=3)
await pool.map(your_download_coroutine, LIST_OF_URLS)

Answer 5

If you have a generator producing your tasks, there may be more tasks than you can fit in memory simultaneously.

The classic asyncio.Semaphore context-manager pattern races all tasks into memory simultaneously.

I don't like the asyncio.Queue pattern. You can prevent it preloading all the tasks into memory (by setting maxsize=1), but it still requires boilerplate to define, start up and shut down the worker coroutines (which consume from the que), and you have to ensure a worker won't fail if a task throws an exception. It feels unpythonic, as if implementing your own multiprocessing.pool.

Instead, here is an alternative:

sem = asyncio.Semaphore(n := 5) # specify maximum concurrency

async def task_wrapper(args):
    try:
        await my_task(*args)
    finally:
        sem.release()

for args in my_generator: # may yield too many to list
    await sem.acquire() 
    asyncio.create_task(task_wrapper(args))

# wait for all tasks to complete
for i in range(n):
    await sem.acquire()

This pauses the generator when there are enough active tasks, and lets the event loop clean up finished tasks. Note, for older python versions, replace create_task with ensure_future.

Answer 6

Small Update: It's no longer necessary to create the loop. I tweaked the code below. Just cleans things up slightly.

# download(code) is the same

async def main():
    no_concurrent = 3
    dltasks = set()
    for i in range(9):
        if len(dltasks) >= no_concurrent:
            # Wait for some download to finish before adding a new one
            _done, dltasks = await asyncio.wait(dltasks, return_when=asyncio.FIRST_COMPLETED)
        dltasks.add(asyncio.create_task(download(i)))
    # Wait for the remaining downloads to finish
    await asyncio.wait(dltasks)

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

Answer 7

Using semaphore, you can also create a decorator to wrap the function

import asyncio
from functools import wraps
def request_concurrency_limit_decorator(limit=3):
    # Bind the default event loop 
    sem = asyncio.Semaphore(limit)

    def executor(func):
        @wraps(func)
        async def wrapper(*args, **kwargs):
            async with sem:
                return await func(*args, **kwargs)

        return wrapper

    return executor

Then, add the decorator to the origin download function.

@request_concurrency_limit_decorator(limit=...)
async def download(...):
    ...

Now you can call the download function like before, but with Semaphore to limit the concurrency.

await download(...)

It should be noted that when the decorator function is executed, the created Semaphore is bound to the default event loop, so you cannot call asyncio.run to create a new loop. Instead, call asyncio.get_event_loop().run... to use the default event loop.

asyncio.Semaphore RuntimeError: Task got Future attached to a different loop

Answer 8

The accepted answer creates a task for every piece of work right away. As has been pointed out, this means creating a lot of tasks simultaneously, even if the semaphore causes the actual work to be done with less parallelism. It also doesn't fare well in situations where pieces of work are read by different processes, where one wants each process to only bite off as many items as it can currently chew.

In a situation where there's a source of work to be done by these tasks, a better approach is to block on reading from that source until there's free task slots available. This is achieved by the following code:

import asyncio


class TaskQueue:
    def __init__(self, maxsize):
        self.maxsize = maxsize
        self.sem = asyncio.BoundedSemaphore(maxsize)
        self.tasks = set()

    def __len__(self):
        return len(self.tasks)

    def _task_done(self, task):
        self.sem.release()
        self.tasks.discard(task)

    async def put(self, coroutine):
        await self.sem.acquire()
        task = asyncio.create_task(coroutine)
        self.tasks.add(task)
        task.add_done_callback(self._task_done)

    async def pending(self):
        pending = self.tasks
        while pending:
            done, pending = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
            for task in done:
                yield len(self)

The pending() method is a bonus for when one wants to track the progress of currently pending tasks. Example usage:

tasks = TaskQueue(MAX_CONCURRENT_TASKS)
async for work_item in work_iter():  # E.g., a database or a queue.
    await tasks.put(do_work(work_item))

print(f"All tasks enqueued. Currently {len(tasks)} task(s) pending.")
async for i in tasks.pending():
    print("Number of pending tasks:", i)

One could also add a join() method that just does await asyncio.gather(*self.tasks).

Answer 9

When using FastAPI on Windows, we can be limited by the number of concurrent connections, as the default is 64 (defined by the var FD_SETSIZE).

More info at https://learn.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-select?redirectedfrom=MSDN

Despite defining ProactorEventLoop (that uses IOCP), on Python versions, prior to 3.7, the select() routines are used, leading to exceptions.

One alternative is to use Andrei's answer to limit the number of concurrent connections in an ML/DL context. Using asyncio + hypercorn + FastAPI, the code is the following:

from hypercorn.config import Config
from hypercorn.asyncio import serve
from fastapi import FastAPI
import asyncio
import json
import time
import sys

app = FastAPI()
conn_limit = 10

async def gather_with_concurrency(n, *coros):
    """
    From Andrei's answer
    """
    semaphore = asyncio.Semaphore(n)

    async def sem_coro(coro):
        async with semaphore:
            return await coro
    return await asyncio.gather(*(sem_coro(c) for c in coros))


@app.get('/app/test')
def req_test():
    time.sleep(1)
    return {"test": "ok"}


if __name__ == "__main__":
    # Start the loop
    config = Config()
    config.bind = [f"0.0.0.0:12000"]
    config.workers = 1
    if sys.platform == 'win32':
        logger.info("Setting proactor event loop for Windows platform.")
        loop = asyncio.ProactorEventLoop()
        asyncio.set_event_loop(loop)
    loop = asyncio.get_event_loop()
    loop.run_until_complete(gather_with_concurrency(conn_limit, serve(app, config)))
    loop.close()

Obs: This script was tested on Python 3.7.16 against 1000 workers on Apache JMeter.

Answer 10

I like using aiometer for limiting concurrency, so maybe this helps someone stumbling across the post. Aiometer also has good documentation on github and a Wiki page. Here's a possible solution to original question, although I have provided it without an in-depth understanding of the asyncio.Event() etc. - which I am jet to learn - hence the warning at the end.

from random import randint
import asyncio
import aiometer
import functools


async def download(code, some_other_arg=None):
    wait_time = randint(1, 5)
    print('downloading {} will take {} second(s)'.format(code, wait_time))
    await asyncio.sleep(wait_time)  # I/O, context will switch to main function
    print('downloaded {}'.format(code))
    return randint(1, 100) #return some random data

tasks = [download(i) for i in range(9)]


async def main():
    async with aiometer.amap(
        functools.partial(download, some_other_arg='foo'),
        tasks,
        max_at_once=3, # Limit maximum number of concurrently running tasks.
        max_per_second=100,  # Limit request rate to not overload the server.
    ) as results:
        async for data in results:
            print("Data returned: ", data)

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

This does however produce a RuntimeWarning which I couldn't yet figure out and in a bigger project may be a dealbreaker:

sys:1: RuntimeWarning: coroutine 'download' was never awaited

If someone knows how to fix above error in this, it will help, but it still seems to run fine nonetheless. I tried this substitution in the task creation part, but it didn't run at all, with the same error:

tasks = [asyncio.create_task(download(i)) for i in range(9)]