Asynchronous Requests with Python requests

Question

I tried the sample provided within the documentation of the requests library for python.

With async.map(rs), I get the response codes, but I want to get the content of each page requested. This, for example, does not work:

out = async.map(rs)
print out[0].content

Solution 1

Note

The below answer is not applicable to requests v0.13.0+. The asynchronous functionality was moved to grequests after this question was written. However, you could just replace requests with grequests below and it should work.

I've left this answer as is to reflect the original question which was about using requests < v0.13.0.

---

To do multiple tasks with async.map asynchronously you have to:

1. Define a function for what you want to do with each object (your task)
2. Add that function as an event hook in your request
3. Call async.map on a list of all the requests / actions

Example:

from requests import async
# If using requests > v0.13.0, use
# from grequests import async

urls = [
    'http://python-requests.org',
    'http://httpbin.org',
    'http://python-guide.org',
    'http://kennethreitz.com'
]

# A simple task to do to each response object
def do_something(response):
    print response.url

# A list to hold our things to do via async
async_list = []

for u in urls:
    # The "hooks = {..." part is where you define what you want to do
    # 
    # Note the lack of parentheses following do_something, this is
    # because the response will be used as the first argument automatically
    action_item = async.get(u, hooks = {'response' : do_something})

    # Add the task to our list of things to do via async
    async_list.append(action_item)

# Do our list of things to do via async
async.map(async_list)

Solution 2

async is now an independent module : grequests.

See here : https://github.com/kennethreitz/grequests

And there: Ideal method for sending multiple HTTP requests over Python?

installation:

$ pip install grequests

usage:

build a stack:

import grequests

urls = [
    'http://www.heroku.com',
    'http://tablib.org',
    'http://httpbin.org',
    'http://python-requests.org',
    'http://kennethreitz.com'
]

rs = (grequests.get(u) for u in urls)

send the stack

grequests.map(rs)

result looks like

[<Response [200]>, <Response [200]>, <Response [200]>, <Response [200]>, <Response [200]>]

grequests don't seem to set a limitation for concurrent requests, ie when multiple requests are sent to the same server.

Solution 3

I tested both requests-futures and grequests. Grequests is faster but brings monkey patching and additional problems with dependencies. requests-futures is several times slower than grequests. I decided to write my own and simply wrapped requests into ThreadPoolExecutor and it was almost as fast as grequests, but without external dependencies.

import requests
import concurrent.futures

def get_urls():
    return ["url1","url2"]

def load_url(url, timeout):
    return requests.get(url, timeout = timeout)

with concurrent.futures.ThreadPoolExecutor(max_workers=20) as executor:

    future_to_url = {executor.submit(load_url, url, 10): url for url in     get_urls()}
    for future in concurrent.futures.as_completed(future_to_url):
        url = future_to_url[future]
        try:
            data = future.result()
        except Exception as exc:
            resp_err = resp_err + 1
        else:
            resp_ok = resp_ok + 1

Solution 4

Unfortunately, as far as I know, the requests library is not equipped for performing asynchronous requests. You can wrap async/await syntax around requests, but that will make the underlying requests no less synchronous. If you want true async requests, you must use other tooling that provides it. One such solution is aiohttp (Python 3.5.3+). It works well in my experience using it with the Python 3.7 async/await syntax. Below I write three implementations of performing n web requests using

1. Purely synchronous requests (sync_requests_get_all) using the Python requests library
2. Synchronous requests (async_requests_get_all) using the Python requests library wrapped in Python 3.7 async/await syntax and asyncio
3. A truly asynchronous implementation (async_aiohttp_get_all) with the Python aiohttp library wrapped in Python 3.7 async/await syntax and asyncio

"""
Tested in Python 3.5.10
"""

import time
import asyncio
import requests
import aiohttp

from asgiref import sync

def timed(func):
    """
    records approximate durations of function calls
    """
    def wrapper(*args, **kwargs):
        start = time.time()
        print('{name:<30} started'.format(name=func.__name__))
        result = func(*args, **kwargs)
        duration = "{name:<30} finished in {elapsed:.2f} seconds".format(
            name=func.__name__, elapsed=time.time() - start
        )
        print(duration)
        timed.durations.append(duration)
        return result
    return wrapper

timed.durations = []


@timed
def sync_requests_get_all(urls):
    """
    performs synchronous get requests
    """
    # use session to reduce network overhead
    session = requests.Session()
    return [session.get(url).json() for url in urls]


@timed
def async_requests_get_all(urls):
    """
    asynchronous wrapper around synchronous requests
    """
    session = requests.Session()
    # wrap requests.get into an async function
    def get(url):
        return session.get(url).json()
    async_get = sync.sync_to_async(get)

    async def get_all(urls):
        return await asyncio.gather(*[
            async_get(url) for url in urls
        ])
    # call get_all as a sync function to be used in a sync context
    return sync.async_to_sync(get_all)(urls)

@timed
def async_aiohttp_get_all(urls):
    """
    performs asynchronous get requests
    """
    async def get_all(urls):
        async with aiohttp.ClientSession() as session:
            async def fetch(url):
                async with session.get(url) as response:
                    return await response.json()
            return await asyncio.gather(*[
                fetch(url) for url in urls
            ])
    # call get_all as a sync function to be used in a sync context
    return sync.async_to_sync(get_all)(urls)


if __name__ == '__main__':
    # this endpoint takes ~3 seconds to respond,
    # so a purely synchronous implementation should take
    # little more than 30 seconds and a purely asynchronous
    # implementation should take little more than 3 seconds.
    urls = ['https://postman-echo.com/delay/3']*10

    async_aiohttp_get_all(urls)
    async_requests_get_all(urls)
    sync_requests_get_all(urls)
    print('----------------------')
    [print(duration) for duration in timed.durations]

On my machine, this is the output:

async_aiohttp_get_all          started
async_aiohttp_get_all          finished in 3.20 seconds
async_requests_get_all         started
async_requests_get_all         finished in 30.61 seconds
sync_requests_get_all          started
sync_requests_get_all          finished in 30.59 seconds
----------------------
async_aiohttp_get_all          finished in 3.20 seconds
async_requests_get_all         finished in 30.61 seconds
sync_requests_get_all          finished in 30.59 seconds

Solution 5

maybe requests-futures is another choice.

from requests_futures.sessions import FuturesSession

session = FuturesSession()
# first request is started in background
future_one = session.get('http://httpbin.org/get')
# second requests is started immediately
future_two = session.get('http://httpbin.org/get?foo=bar')
# wait for the first request to complete, if it hasn't already
response_one = future_one.result()
print('response one status: {0}'.format(response_one.status_code))
print(response_one.content)
# wait for the second request to complete, if it hasn't already
response_two = future_two.result()
print('response two status: {0}'.format(response_two.status_code))
print(response_two.content)

It is also recommended in the office document. If you don't want involve gevent, it's a good one.

View more solutions

Comments

[Mariusz Jamro]

Maybe the responses you're getting have empty body?

[Chewie]

Works for me. Please post the full error you're getting.

[trbck]

there is no error. it just runs forever by the provided test urls.

[trbck]

it obviously appears when I use urls over https. http is working just fine

[OrangeDog]

Looks like requests-threads exists now.

[guettli]

Most answers are outdated. In the year 2021 the current bandwagon-effect winner is: docs.aiohttp.org/en/stable

[synthesizerpatel]

With regards to the limitation on concurrent requests - you can specify a pool size when running the map()/imap(). i.e. grequests.map(rs, size=20) to have 20 concurrent grabs.

[outforawhile]

Nice idea to have left your comment : due to compatibility issues between latest requests and grequests (lack of max_retries option in requests 1.1.0) i had to downgrade requests to retrieve async and I have found that the asynchronous functionality was moved with versions 0.13+ (pypi.python.org/pypi/requests)

[Jeff]

Thanks for pointing that out! Will edit my answer to reflect this info.

[Nick]

this is "normal" requests in threads. is not bad example buy is off-topic.

[Pedro]

Dumb question: What is the speed increase of using grequests as opposed to simply requests? What limits are there regarding requests? e.g. would putting 3500 requests in async.map be OK?

[saarp]

As of now this is not python3-capable (gevent fails to build v2.6 on py3.4).

[Allan Ruin]

I not quite understand the async part. if I let results = grequests.map(rs) the the code after this line is block, I can see the async effect?

[Allan Ruin]

from grequests import async do not work.. and this definition of dosomething work for me def do_something(response, **kwargs):, I find it from stackoverflow.com/questions/15594015/…

[Monkey Boson]

@YSY Feel free to post an issue: github.com/ctheiss/simple-requests/issues; I literally use this library thousands of times a day.

[YSY]

Boston, how do you handle 404/500 errors? what about https urls? will appreciate a snipping that supports thousands of urls. can you please paste an example? thanks

[Monkey Boson]

@YSY By default 404/500 errors raise an exception. This behaviour can be overridden (see pythonhosted.org/simple-requests/…). HTTPS urls are tricky due to the reliance on gevent, which currently has an outstanding bug on this (github.com/gevent/gevent/issues/477). There is a shim in the ticket you can run, but it will still throw warnings for SNI servers (but it will work). As for snipping, I'm afraid all my usages are at my company and closed. But I assure you we execute thousands of requests over tens of jobs.

[bryanph]

if the async.map call still blocks, then how is this asynchronous? Besides the requests themselves being sent asynchronously, the retrieval is still synchronous?

[Jose Cherian]

One of the easiest solutions. Number of concurrent requests can be increased by defining max_workers parameter

[Slow Harry]

What type of exception is possible here?

[Hodza]

requests.exceptions.Timeout

[Slow Harry]

Thanks, but I can't understand why you create a timeout for request?

[Hodza]

In my company we handle thouthands of urls per second and some sites can block thread for too long. So timouts increase our overall speed. Also we use max request size. stackoverflow.com/questions/23514256/…

[Slow Harry]

Thanks a lot, for open answer, could you also explain me, is there motivation to handle one request with multiple threads, or it is impractical ?

[Hodza]

Sorry I dont understand your question. Use only single url in multiple threads? Only one case DDoS attacks ))

[user3125280]

the downvote was accidental and now its locked! sorry, this was actually a very useful answer for me

[Martin Thoma]

Replacing from requests import async by import grequests as async worked for me.

[Ja8zyjits]

i second @AllanRuin, may be Jeff should add that line in the answer, that shall help future users.

[user]

@user3125280, I've edited the answer, feel free to recast your vote.

[GrayedFox]

Thanks a bunch for this!

[Macumbaomuerte]

using grequests on python3 on windows, without the **kwargs @AllanRuin mentioned won't work, thanks for pointing out, i was gonna smash my computer before i found this

[user1717828]

It'd be nice to see an example of this scaled so we're not using one variable name per item to loop over.

[Bunyk]

Oh, thank you very much for "The asynchronous functionality was moved to grequests after this question was written". I was looking for it everywhere in docs, and was not able to found. :)

[Corneliu Maftuleac]

having one thread per request is a hell waste of resources! it is not possible to do for example 500 requests simultaneously, it will kill your cpu. this should never be considered a good solution.

[Dreampuf]

@CorneliuMaftuleac good point. Regarding the thread usage, you definitely need to care about it and the library provide an option to enable the threading pool or processing pool. ThreadPoolExecutor(max_workers=10)

[Corneliu Maftuleac]

@Dreampuf processing pool I believe is even worse?

[Dreampuf]

@CorneliuMaftuleac I have to say it depends on your case. It has always been a trade-off about I/O bound or CPU bound.

[Jon]

This feels like it was copied and pasted from the documentation page. As @user1717828, it'd be more beneficial to see a scaled up example.

[OrangeDog]

grequests now recommends requests-threads or requests-futures

[nagylzs]

I don't understand why this answer got so many upvotes. The OP question was about async requests. ThreadPoolExecutor runs threads. Yes, you can make requests in multiple threads, but that will never be an async program, so I how could it be an answer for the original question?

[Hodza]

Actually, the question was about how to load URLs in parallel. And yes thread pool executor is not the best option, it is better to use async io, but it works well in Python. And I don't understand why threads couldn't be used for async? What if you need to run CPU bound task asynchronously?

[nurettin]

confirmed, works great. On the project page it says this work has been overtaken by the following project github.com/encode/httpx

[Isaac Philip]

Library looks sleek with respect to interaction. Is Python3+ usable? Sorry could not see any mention.

[Coder Nr 23]

If I understand correctly, this will block the event loop while doing the GRPC and HTTP call? So if these calls take seconds to complete, your entire event loop will block for seconds? To avoid this, you need to use GRPC or HTTP libraries that are async. Then you can for example do await response = requests.get(URL). No?

[Monkey Boson]

@Jethro absolutely right, the library would need a total re-write since the underlying technologies are quite different in Python 3. For right now, the library is "complete" but only works for Python 2.

[DragonBobZ]

Unfortunately, when trying this out, I found that making a wrapper around requests is barely faster (and in some cases slower) than just calling a list of URLs synchronously. E.g, requesting an endpoint that takes 3 seconds to respond 10 times using the strategy above takes about 30 seconds. If you want true async performance, you need to use something like aiohttp.

[rlat]

And respx for mocking/testing :)

[Amir Hossein Baghernezad]

Twisted is old fashioned. Use HTTPX instead.

[arshbot]

@DragonBobZ In my case I saw a ~40% reduction in time. The main benefit was to be able to perform necessary chores while waiting for the next call. In my dataset I was making hundreds of calls, so scale could also be a factor.

[arshbot]

@CoderNr23 Someone can correct me on this, but iirc even with that syntax, executing the tasks is fundamentally synchronous - the order of which tasks when is what's made async. In python, you're merely packaging a synchronous execution tree which will be executed with some starter like run_until_complete - this is unless you use the threading module, which delegates async to the OS layer. Read up on the GIL problem in python for more info

[DragonBobZ]

@arshbot Yes, if your chores are asynchronous, then you will see speed-ups, despite waiting on synchronous calls to requests.get. But the question is how to perform asynchronous requests with the python requests library. This answer does not do that, so my criticism stands.

[iedmrc]

I prefer using the event loop rather than thread pools.

[iedmrc]

I think this should be bumped. Using async event loops seems enough to fire asynchronous requests. No need to install external dependencies.

[AJ.]

Hi @Uri, I am getting below error in trying the code you mentioned in this answer. await asyncio.gather(*map(get_async, urls)) ^ SyntaxError: invalid syntax Please guide

[Uri]

Note that you need an async context to use await.

[rtaft]

This works but it generates a new thread for each request, which seems to defeat the purpose of using asyncio.

[vaskrneup]

@rtaft Thank you for the suggestion, I have corrected my words.

[Sergio Chumacero]

@iedmrc sadly, this is not the case. For a task to be non-blocking it has to be implemented using the newer async tools in Python, and this is not the case with the requests library. If you just use stick requests tasks in an async event loop, those would still be blocking. That being said, you can (as suggested in other responses) use things like gevent or threads with requests, but certainly not asyncio.

[theberzi]

On the github, repo, the author of grequests recommends using requests-threads or requests-futures instead.

[guettli]

"asnyc" is this a typo, or on purpose?

[DragonBobZ]

definitely a typo

[wescpy]

Your async_aiohttp_get_all() is a nice solution. I came up with something similar, but had an extra async def fetch_all(urls): return await asyncio.gather(*[fetch(url) for url in urls]) outside of it, which had my solution creating separate aiohttp.ClientSession() instances for each URL whereas by embedding a local function, you're able to reuse the same session... much more Pythonic IMO. Can you remind me of the benefit of using sync.async_to_sync() with the existence of get_all() vs. asyncio.run() without get_all()?

[Tuhin Mitra]

awesomely done, definitely async_aiohttp working better than all!