Async wrapper for requests / aiohttp, and some crawler toolkits. Let synchronization code enjoy the performance of asynchronous programming.
torequests 
Briefly speaking, requests & aiohttp wrapper for asynchronous programming rookie, to shorten the code quantity.
Install:
pip install torequests -U
requirements
python3.7+ (follow requests settings)
| requests # python
| aiohttp >= 3.6.2 # python3
Features
Inspired by tomorrow, to make async-coding brief & smooth, compatible for win32 / python 2&3.
- Convert any funtions into async-mode with concurrent.futures.
- Wrap requests lib with concurrent.futures to enjoy the concurrent performance.
- Simplify aiohttp, make it
requests-like. - Add FailureException to check the context of request failure.
- Add frequency control, prevent from anti-spider based on frequency check.
- Add retry for request. (support
response_validator) - Alenty of common crawler utils.
- Compatible with uniparser
Getting started
1. Async, threads - make functions asynchronous
from torequests import threads, Async
import time
@threads(5)
def test1(n):
time.sleep(n)
return 'test1 ok'
def test2(n):
time.sleep(n)
return 'test1 ok'
start = int(time.time())
# here async_test2 is same as test1
async_test2 = Async(test2)
future = test1(1)
# future run in non blocking thread pool
print(future, ', %s s passed' % (int(time.time() - start)))
# call future.x will block main thread and get the future.result()
print(future.x, ', %s s passed' % (int(time.time() - start)))
# output:
# <NewFuture at 0x34b1d30 state=running> , 0 s passed
# test1 ok , 1 s passed
2. tPool - thread pool for async-requests
from torequests.main import tPool
from torequests.logs import print_info
from torequests.utils import ttime
req = tPool()
test_url = 'http://p.3.cn'
def callback(task):
return (len(task.content),
print_info(
ttime(task.task_start_time), '-> ', ttime(task.task_end_time),
',', round(task.task_cost_time * 1000), 'ms'))
ss = [req.get(test_url, retry=2, callback=callback) for i in range(3)]
# or [i.x for i in ss]
req.x
# task.cx returns the callback_result
ss = [i.cx for i in ss]
print_info(ss)
# [2020-01-21 16:56:23] temp_code.py(11): 2020-01-21 16:56:23 -> 2020-01-21 16:56:23 , 54 ms
# [2020-01-21 16:56:23] temp_code.py(11): 2020-01-21 16:56:23 -> 2020-01-21 16:56:23 , 55 ms
# [2020-01-21 16:56:23] temp_code.py(11): 2020-01-21 16:56:23 -> 2020-01-21 16:56:23 , 57 ms
# [2020-01-21 16:56:23] temp_code.py(18): [(612, None), (612, None), (612, None)]
2.1 Performance.
[3.7.1 (v3.7.1:260ec2c36a, Oct 20 2018, 14:57:15) [MSC v.1915 64 bit (AMD64)]]: 2000 / 2000, 100.0%, cost 4.2121 seconds, 475.0 qps.
import timeit
from torequests import tPool
import sys
req = tPool()
start_time = timeit.default_timer()
oks = 0
total = 2000
# concurrent all the tasks
tasks = [req.get('http://127.0.0.1:8080') for num in range(total)]
for task in tasks:
r = task.x
if r.text == 'ok':
oks += 1
end_time = timeit.default_timer()
succ_rate = oks * 100 / total
cost_time = round(end_time - start_time, 4)
version = sys.version
qps = round(total / cost_time, 0)
print(
'[{version}]: {oks} / {total}, {succ_rate}%, cost {cost_time} seconds, {qps} qps.'
.format(**vars()))
3. Requests - aiohttp-wrapper
from torequests.dummy import Requests
from torequests.logs import print_info
from torequests.utils import ttime
req = Requests(frequencies={'p.3.cn': (2, 1)})
def callback(task):
return (len(task.content),
print_info(
ttime(task.task_start_time), '->', ttime(task.task_end_time),
',', round(task.task_cost_time * 1000), 'ms'))
ss = [
req.get('http://p.3.cn', retry=1, timeout=5, callback=callback)
for i in range(4)
]
req.x # this line can be removed
ss = [i.cx for i in ss]
print_info(ss)
# [2020-01-21 18:15:33] temp_code.py(11): 2020-01-21 18:15:32 -> 2020-01-21 18:15:33 , 1060 ms
# [2020-01-21 18:15:33] temp_code.py(11): 2020-01-21 18:15:32 -> 2020-01-21 18:15:33 , 1061 ms
# [2020-01-21 18:15:34] temp_code.py(11): 2020-01-21 18:15:32 -> 2020-01-21 18:15:34 , 2081 ms
# [2020-01-21 18:15:34] temp_code.py(11): 2020-01-21 18:15:32 -> 2020-01-21 18:15:34 , 2081 ms
# [2020-01-21 18:15:34] temp_code.py(20): [(612, None), (612, None), (612, None), (612, None)]
3.1 Performance.
aiohttp is almostly 3 times faster than requests + ThreadPoolExecutor, even without uvloop on windows10.
3.7.1 (v3.7.1:260ec2c36a, Oct 20 2018, 14:57:15) [MSC v.1915 64 bit (AMD64)]
sync_test: 2000 / 2000, 100.0%, cost 1.2965 seconds, 1543.0 qps.
async_test: 2000 / 2000, 100.0%, cost 1.2834 seconds, 1558.0 qps.
Sync usage has little performance lost.
import asyncio
import sys
import timeit
from torequests.dummy import Requests
def sync_test():
with Requests() as req:
start_time = timeit.default_timer()
oks = 0
total = 2000
# concurrent all the tasks
tasks = [req.get('http://127.0.0.1:8080') for num in range(total)]
req.x
for task in tasks:
if task.text == 'ok':
oks += 1
end_time = timeit.default_timer()
succ_rate = oks * 100 / total
cost_time = round(end_time - start_time, 4)
qps = round(total / cost_time, 0)
print(
f'sync_test: {oks} / {total}, {succ_rate}%, cost {cost_time} seconds, {qps} qps.'
)
async def async_test():
req = Requests()
async with Requests() as req:
start_time = timeit.default_timer()
oks = 0
total = 2000
# concurrent all the tasks
tasks = [req.get('http://127.0.0.1:8080') for num in range(total)]
for task in tasks:
r = await task
if r.text == 'ok':
oks += 1
end_time = timeit.default_timer()
succ_rate = oks * 100 / total
cost_time = round(end_time - start_time, 4)
qps = round(total / cost_time, 0)
print(
f'async_test: {oks} / {total}, {succ_rate}%, cost {cost_time} seconds, {qps} qps.'
)
if __name__ == "__main__":
print(sys.version)
sync_test()
loop = asyncio.get_event_loop()
loop.run_until_complete(async_test())
3.2 using torequests.dummy.Requests in async environment.
import asyncio
import uvicorn
from starlette.applications import Starlette
from starlette.responses import PlainTextResponse
from torequests.dummy import Requests
api = Starlette()
api.req = Requests()
@api.route('/')
async def index(req):
# await for Response or FailureException
r = await api.req.get('http://p.3.cn', timeout=(1, 1))
return PlainTextResponse(r.content)
if __name__ == "__main__":
uvicorn.run(api)
3.3 using torequests.aiohttp_dummy.Requests for better performance.
Removes the frequency_controller & sync usage (task.x) & compatible args of requests for good performance, but remains retry / callback / referer_info.
# -*- coding: utf-8 -*-
from torequests.aiohttp_dummy import Requests
from asyncio import get_event_loop
async def main():
url = 'http://httpbin.org/json'
# simple usage, catch_exception defaults to True
async with Requests() as req:
r = await req.get(url, retry=1, encoding='u8', timeout=3)
print(r.ok)
print(r.url)
print(r.content[:5])
print(r.status_code)
print(r.text[:5])
print(r.json()['slideshow']['title'])
print(r.is_redirect)
print(r.encoding)
# or use `req = Requests()` without context
req = Requests()
r = await req.get(
url,
referer_info=123,
callback=lambda r: r.referer_info + 1,
)
print(r) # 124
get_event_loop().run_until_complete(main())
4. utils: some useful crawler toolkits
ClipboardWatcher: watch your clipboard changing.
Counts: counter while every time being called.
Null: will return self when be called, and alway be False.
Regex: Regex Mapper for string -> regex -> object.
Saver: simple object persistent toolkit with pickle/json.
Timer: timing tool.
UA: some common User-Agents for crawler.
curlparse: translate curl-string into dict of request.
md5: str(obj) -> md5_string.
print_mem: show the proc-mem-cost with psutil, use this only for lazinesssss.
ptime: %Y-%m-%d %H:%M:%S -> timestamp.
ttime: timestamp -> %Y-%m-%d %H:%M:%S
slice_by_size: slice a sequence into chunks, return as a generation of chunks with size.
slice_into_pieces: slice a sequence into n pieces, return a generation of n pieces.
timeago: show the seconds as human-readable.
unique: unique one sequence.
find_one: use regex like Javascript to find one string with index(like [0], [1]).
find_jsons: find all json substrings from a mass string contains list / dict JSON.
...
Benchmarks
Benchmark of concurrent is not very necessary and accurate, just take a look.
Test Server: golang net/http
Source code: go_test_server.go
Client Testing Code
Source code: py_test_client.py
Test Result on Windows (without uvloop)
Intel® Core™ i7-8750H CPU @ 2.20GHz (12 CPUs), ~2.2GHz; 12 logical CPUs.
Windows-10-10.0.18362-SP0
3.7.1 (v3.7.1:260ec2c36a, Oct 20 2018, 14:57:15) [MSC v.1915 64 bit (AMD64)]
['aiohttp(3.6.2)', 'torequests(4.9.14)', 'requests(2.23.0)', 'httpx(0.12.1)']
================================================================================
test_aiohttp : 2000 / 2000 = 100.0%, cost 1.289s, 1551 qps, 100% standard.
test_dummy : 2000 / 2000 = 100.0%, cost 1.397s, 1432 qps, 92% standard.
test_aiohttp_dummy : 2000 / 2000 = 100.0%, cost 1.341s, 1491 qps, 96% standard.
test_httpx : 2000 / 2000 = 100.0%, cost 4.065s, 492 qps, 32% standard.
test_tPool : 2000 / 2000 = 100.0%, cost 5.161s, 388 qps, 25% standard.
Test Result on Linux (with uvloop)
Intel® Xeon® Platinum 8163 CPU @ 2.50GHz; 1 logical CPU.
Linux-4.15.0-13-generic-x86_64-with-Ubuntu-18.04-bionic
3.7.3 (default, Apr 3 2019, 19:16:38)
[GCC 8.0.1 20180414 (experimental) [trunk revision 259383]]
['aiohttp(3.6.2)', 'torequests(4.9.14)', 'requests(2.23.0)', 'httpx(0.12.1)']
================================================================================
test_aiohttp : 2000 / 2000 = 100.0%, cost 0.652s, 3068 qps, 100% standard.
test_dummy : 2000 / 2000 = 100.0%, cost 0.802s, 2495 qps, 81% standard.
test_aiohttp_dummy : 2000 / 2000 = 100.0%, cost 0.726s, 2754 qps, 90% standard.
test_httpx : 2000 / 2000 = 100.0%, cost 2.349s, 852 qps, 28% standard.
test_tPool : 2000 / 2000 = 100.0%, cost 2.708s, 739 qps, 24% standard.
Conclusion
- aiohttp is the fastest, for the cython utils
- aiohttp’s qps is 3068 on 1 cpu linux with uvloop, near to golang’s 3300.
- torequests.dummy.Requests based on aiohttp.
- about 8% performance lost without uvloop.
- about 20% performance lost with uvloop.
- torequests.aiohttp_dummy.Requests based on aiohttp.
- less than 4% performance lost without uvloop.
- less than 10% performance lost with uvloop.
- httpx is faster than requests + threading, but not very obviously on linux.
PS:
golang - net/http 's performance
`2000 / 2000, 100.00 %, cost 0.26 seconds, 7567.38 qps` on windows (12 cpus)
`2000 / 2000, 100.00 %, cost 0.61 seconds, 3302.48 qps` on linux (1 cpu)
1. slower than windows, because golang benefit from multiple CPU count
2. linux 1 cpu, but windows is 12
golang http client testing code: