【问题标题】：Why do UVICORN/Starlette/FastAPI spawn more threads when not using "ASYNC" and don't when using "ASYNC"?为什么 UVICORN/Starlette/FastAPI 在不使用 \"ASYNC\" 时产生更多线程而在使用 \"ASYNC\" 时不产生更多线程？
【发布时间】：2023-02-16 19:12:52
【问题描述】：

好的，所以我正在对 using ASYNC vs without using ASYNC in FastAPI 进行比较研究。但是，我得到了一些意想不到的结果，不明白为什么。

这是设置 1：

不使用异步的 API

import uvicorn
from fastapi import FastAPI
import PIL.Image as Image
import requests
from loguru import logger
import sys

log_format = "{level} {process}-{thread} {time} {name}:{line} - {message}"
logger.remove()
logger.add(sys.stderr, format=log_format, backtrace=True, diagnose=True)
logger.add("logs/" + "t_{time}.log", format=log_format, colorize=True, backtrace=True, diagnose=True)

Image.MAX_IMAGE_PIXELS = None

def get_the_image_from_net():
    a = requests.get("https://eoimages.gsfc.nasa.gov/images/imagerecords/73000/73751/world.topo.bathy.200407.3x21600x21600.A1.jpg")
    return True


app = FastAPI()

@app.get("/expectoPatronum")
def get_image_of_the_new_world():
    """
    Gets Image of the World
    """
    logger.info("Received request for getting Image of the world")
    image_bytes: bytes = get_the_image_from_net()
    logger.info("Image has been downloaded and processed as bytes array")
    return True

uvicorn.run(app, host="0.0.0.0", port=10009)

我如何调用 ABOVE API（通过使用多处理池）

import time
from multiprocessing import Pool, Manager
import requests


def test_function(index_iterator: int):
    start_time = time.time()
    response = requests.get("http://localhost:10009/expectoPatronum")
    print(f"response.content: {str(response.content)}")
    if response.status_code != 200:
        print("----------------------NOT 200")
        print(f"response.content: {str(response.content)}")
    end_time = time.time()
    elapsed_time = end_time - start_time


pool = Pool(5)
pool.map(test_function, [i for i in range(1,6)])
pool.close

这是 LOGS 在调用它们时从我的 API 端读取的内容：

INFO 9408-140671786272512 2022-07-13T01:32:37.498465+0530 __main__:27 - Received request for getting Image of the world
INFO 9408-140671777879808 2022-07-13T01:32:37.501623+0530 __main__:27 - Received request for getting Image of the world
INFO 9408-140671769487104 2022-07-13T01:32:37.504744+0530 __main__:27 - Received request for getting Image of the world
INFO 9408-140671760897792 2022-07-13T01:32:37.504929+0530 __main__:27 - Received request for getting Image of the world
INFO 9408-140671752242944 2022-07-13T01:32:37.505638+0530 __main__:27 - Received request for getting Image of the world
INFO 9408-140671786272512 2022-07-13T01:33:04.845982+0530 __main__:29 - Image has been downloaded and processed as bytes array
INFO 9408-140671777879808 2022-07-13T01:33:16.167435+0530 __main__:29 - Image has been downloaded and processed as bytes array
INFO 9408-140671769487104 2022-07-13T01:33:17.284000+0530 __main__:29 - Image has been downloaded and processed as bytes array
INFO 9408-140671752242944 2022-07-13T01:33:37.771086+0530 __main__:29 - Image has been downloaded and processed as bytes array
INFO 9408-140671760897792 2022-07-13T01:33:38.016435+0530 __main__:29 - Image has been downloaded and processed as bytes array

请注意：创建了多个线程（140671786272512、140671777879808 ..等等）

这是设置 2：

使用异步的 API

import uvicorn
from fastapi import FastAPI, Response, APIRouter
import httpx
from loguru import logger
import sys

log_format = "{level} {process}-{thread} {time} {name}:{line} - {message}"
logger.remove()
logger.add(sys.stderr, format=log_format, backtrace=True, diagnose=True)
logger.add("logs/" + "t_{time}.log", format=log_format, colorize=True, backtrace=True, diagnose=True)

Image.MAX_IMAGE_PIXELS = None

async def get_the_image_from_net():
    async with httpx.AsyncClient() as client:
        a = await client.get('https://eoimages.gsfc.nasa.gov/images/imagerecords/73000/73751/world.topo.bathy.200407.3x21600x21600.A1.jpg')

    return True


app = FastAPI()

@app.get("/expectoPatronum")
async def get_image_of_the_new_world():
    """
    Gets Image of the World
    """
    logger.info("Received request for getting Image of the world")
    image_bytes = await get_the_image_from_net()
    logger.info("Image has been downloaded and processed as bytes array")
    return True

uvicorn.run(app, host="0.0.0.0", port=10008)

我如何调用 ABOVE API（通过使用多处理池（与上述调用方法相同，但端口号不同）

import time
from multiprocessing import Pool, Manager
import requests


def test_function(index_iterator: int):
    start_time = time.time()
    response = requests.get("http://localhost:10008/expectoPatronum")
    print(f"response.content: {str(response.content)}")
    if response.status_code != 200:
        print("----------------------NOT 200")
        print(f"response.content: {str(response.content)}")
    end_time = time.time()
    elapsed_time = end_time - start_time


pool = Pool(5)
pool.map(test_function, [i for i in range(1,6)])
pool.close

这是 LOGS 在调用它们时从我的 API 端读取的内容：

INFO 9442-140295303571264 2022-07-13T01:36:26.762525+0530 __main__:43 - Received request for getting Image of the world
INFO 9442-140295303571264 2022-07-13T01:36:26.776561+0530 __main__:43 - Received request for getting Image of the world
INFO 9442-140295303571264 2022-07-13T01:36:26.783669+0530 __main__:43 - Received request for getting Image of the world
INFO 9442-140295303571264 2022-07-13T01:36:26.790367+0530 __main__:43 - Received request for getting Image of the world
INFO 9442-140295303571264 2022-07-13T01:36:26.796934+0530 __main__:43 - Received request for getting Image of the world
INFO 9442-140295303571264 2022-07-13T01:37:38.086156+0530 __main__:45 - Image has been downloaded and processed as bytes array
INFO 9442-140295303571264 2022-07-13T01:37:43.709798+0530 __main__:45 - Image has been downloaded and processed as bytes array
INFO 9442-140295303571264 2022-07-13T01:37:43.827959+0530 __main__:45 - Image has been downloaded and processed as bytes array
INFO 9442-140295303571264 2022-07-13T01:37:47.218717+0530 __main__:45 - Image has been downloaded and processed as bytes array
INFO 9442-140295303571264 2022-07-13T01:37:51.712889+0530 __main__:45 - Image has been downloaded and processed as bytes array

请注意：只有 1 个线程（140295303571264）

现在我无法弄清楚：为什么在 SETUP-1 中，uvicorn 会产生超过 1 个线程？我认为它是一个单线程服务器，我怎样才能限制它在进程内只产生一个线程？

两种设置的时间结果确实令人沮丧。对于不使用异步的 SETUP-1，时间为：27.36284899711609,38.68544268608093,39.79848337173462,60.28416633605957,60.529775857925415

对于使用异步的 SETUP-2，时间为：71.32960891723633,76.95274710655212,77.07157778739929,80.4650149345398,84.95978450775146

【问题讨论】：

FastAPI 在内部使用 40 个线程的线程池来处理使用非异步端点的请求。使用异步时，只使用一个线程（因为不等待的线程应该被 CPU 绑定的东西占用）。您通常可以告诉您的 ASGI 服务器运行多个 worker。
是的，“添加”async 到我的非异步 API 使其成为我想要的单线程。

标签： python python-asyncio fastapi uvicorn starlette

【解决方案1】：

当您使用普通的 def 而不是 async def 声明路径操作函数时，它会在外部 threadpool 中运行，然后等待，而不是直接调用（因为它会阻塞服务器）。你可以阅读更多 in FastAPI documentation

【讨论】：