FastAPI高并发场景下httpx性能劣于aiohttp的原因及代码排查请求

在FastAPI框架中进行同步与异步路由测试时，发现复用同一客户端实例的httpx在高并发请求场景下性能远不如aiohttp，现针对代码问题排查如下：

原测试代码 #

import uvicorn
from fastapi import FastAPI
import requests
import httpx
import aiohttp

app = FastAPI(description="sync and async route func test")

aio_client = httpx.AsyncClient()
aio_session: aiohttp.ClientSession = None
req_session = requests.Session()


@app.on_event("startup")
async def startup_event():
    global aio_session
    aio_session = aiohttp.ClientSession()


async def async_get(url):
    # async with aio_session.get(url) as resp:
    #     return resp

    resp = await aio_client.get(url)
    return resp


@app.get("/ping")
async def ping():
    return "pong"


@app.get("/async_route_func_demo")
async def async_route_func():
    url = "https://juejin.cn/"
    resp = await async_get(url)
    # print("resp", resp)
    return "async_route_func_demo"


@app.get("/async_route_use_sync_io_demo")
async def async_route_func():
    url = "https://juejin.cn/"
    resp = req_session.get(url)
    # print("resp", resp)
    return "async_route_func_demo"


@app.get("/sync_route_func_demo")
def sync_route_func():
    url = "https://juejin.cn/"
    resp = req_session.get(url)
    # print("resp", resp)
    return "sync_route_func_demo"


def main():
    uvicorn.run(app, log_level="warning")


if __name__ == '__main__':
    main()

环境信息 #

• Python版本：3.10.12
• 操作系统：macOS M1 Pro
• 依赖版本：

fastapi==0.101.0
uvicorn==0.23.2
requests==2.32.0
httpx==0.27.0

wrk测试结果 #

高并发测试显示，使用复用httpx异步客户端实例的路由接口，QPS、延迟等核心性能指标均显著落后于使用aiohttp客户端的版本。

问题排查与修正 #

原代码核心问题 #

1. httpx客户端初始化时机错误：在模块级别直接创建httpx.AsyncClient()，未在FastAPI的异步启动上下文内完成初始化，导致客户端与事件循环绑定异常，高并发场景下易出现资源竞争或阻塞。
2. 未正确销毁httpx客户端：仅在启动事件中初始化了aiohttp会话，但未在服务关闭时关闭httpx客户端，会引发连接池资源泄漏，长期运行后性能持续下降。
3. 对比测试逻辑未启用：原代码中aiohttp的调用逻辑被注释，无法直接完成两者的性能对比验证。

修正后的代码 #

import uvicorn
from fastapi import FastAPI
import requests
import httpx
import aiohttp

app = FastAPI(description="sync and async route func test")

aio_client: httpx.AsyncClient = None
aio_session: aiohttp.ClientSession = None
req_session = requests.Session()


@app.on_event("startup")
async def startup_event():
    global aio_session, aio_client
    # 在异步启动上下文内初始化客户端，确保与事件循环正确绑定
    aio_session = aiohttp.ClientSession()
    # 对齐aiohttp默认连接池配置，缩小性能差异
    aio_client = httpx.AsyncClient(
        limits=httpx.Limits(max_keepalive_connections=100, max_connections=200)
    )


@app.on_event("shutdown")
async def shutdown_event():
    # 关闭所有客户端，释放资源
    await aio_session.close()
    await aio_client.aclose()
    req_session.close()


async def async_get_httpx(url):
    resp = await aio_client.get(url)
    return resp


async def async_get_aiohttp(url):
    async with aio_session.get(url) as resp:
        return resp


@app.get("/ping")
async def ping():
    return "pong"


@app.get("/async_route_httpx")
async def async_route_httpx():
    url = "https://juejin.cn/"
    await async_get_httpx(url)
    return "async_route_httpx"


@app.get("/async_route_aiohttp")
async def async_route_aiohttp():
    url = "https://juejin.cn/"
    await async_get_aiohttp(url)
    return "async_route_aiohttp"


@app.get("/async_route_use_sync_io")
async def async_route_use_sync_io():
    url = "https://juejin.cn/"
    resp = req_session.get(url)
    return "async_route_use_sync_io"


@app.get("/sync_route_func")
def sync_route_func():
    url = "https://juejin.cn/"
    resp = req_session.get(url)
    return "sync_route_func"


def main():
    uvicorn.run(app, log_level="warning")


if __name__ == '__main__':
    main()

额外说明 #

httpx默认连接池配置与aiohttp存在差异，初始化时显式设置limits参数对齐两者的连接池大小，可进一步缩小性能差距。

内容的提问来源于stack exchange，提问作者hui