[python]Gunicorn加持下的Flask性能测试

前言

之前学习和实际生产环境的flask都是用 app.run() 的默认方式启动的,因为只是公司内部服务,请求量不高,一直也没出过什么性能问题。最近接管其它小组的服务时,发现他们的服务使用Gunicorn + Flask的方式运行的,本地开发用的gevent的WSGIServer。对于Gunicorn之前只是耳闻,没实际用过,正好捣鼓下看看到底能有多少性能提升。本文简单记录flask在各种配置参数和运行方式的性能,后面也会跟其他语言和框架做个对比。

  • python版本:3.11
  • flask版本:3.0.3
  • Gunicorn:23.0.0
  • wrk作为性能测试工具
  • 运行环境:vbox虚拟机,debian 12, 4C4G的硬件配置

wrk测试脚本

wrk支持用lua脚本对请求的响应结果进行验证,以下脚本对响应码和响应内容进行校验

wrk.method = "GET"
wrk.host = "127.0.0.1:8080"
wrk.path = "/health"
wrk.timeout = 1.0

response = function(status, headers, body)
    if status ~= 200 then
        print("Error: expected 200 but got " .. status)
    end

    if not body:find("ok") then
        print("Error: response does not contain expected content.")
    end
end

Flask框架的测试记录

  1. 先测试默认运行方式,且没有sleep的情况下的并发性能。
from flask import Flask

app = Flask(__name__)

@app.get("/health")
def health():
    return "ok"

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=8080)

使用命令 nohup python demo.py > /dev/null 启动,以下为wrk测试结果,可以看到已经出现超时请求。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    65.47ms   81.07ms   1.99s    97.69%
    Req/Sec   575.99    107.20     1.08k    66.71%
  137538 requests in 1.00m, 22.82MB read
  Socket errors: connect 0, read 114, write 0, timeout 144
Requests/sec:   2288.49
Transfer/sec:    388.86KB
  1. 还是默认启动方式,增加等待时间,模拟处理任务的时间消耗。后续测试都会增加等待时间。
from flask import Flask
from time import sleep

app = Flask(__name__)

@app.get("/health")
def health():
    sleep(0.1)
    return "ok"

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=8080)

wrk测试结果,不出所料性能会有所下降。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   201.90ms  239.99ms   2.00s    95.05%
    Req/Sec   479.79    185.87     1.82k    68.19%
  114440 requests in 1.00m, 18.99MB read
  Socket errors: connect 0, read 2, write 0, timeout 1833
Requests/sec:   1904.45
Transfer/sec:    323.62KB
  1. flask 更新到版本2后支持使用异步函数(需要安装异步相关依赖 python -m pip install -U flask[async]
from flask import Flask
import asyncio

app = Flask(__name__)

@app.route('/health')
async def health():
    await asyncio.sleep(0.1)
    return "ok"

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=8080)

wrk测试结果,性能相较于同步函数甚至还下降了,QPS几乎砍半,看来异步版Flask还有待增强。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   275.49ms  190.22ms   2.00s    95.46%
    Req/Sec   242.86    104.25   720.00     65.91%
  57896 requests in 1.00m, 9.61MB read
  Socket errors: connect 0, read 48, write 0, timeout 611
Requests/sec:    964.31
Transfer/sec:    163.86KB
  1. 接管的Flask应用在本地使用gevent的WSGIServer运行,所以也来试试。
from gevent.pywsgi import WSGIServer
from flask import Flask
from time import sleep

app = Flask(__name__)

@app.route('/health')
def health():
    sleep(0.1)
    return "ok"

if __name__ == "__main__":
    http_server = WSGIServer(('0.0.0.0', 8080), app)
    http_server.serve_forever()

wrk测试结果,惨不忍睹,像是单线程在挨个处理请求,每个请求都会阻塞住。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   185.36ms  221.17ms   1.93s    96.30%
    Req/Sec     5.54      3.39    10.00     46.55%
  592 requests in 1.00m, 67.64KB read
  Socket errors: connect 0, read 0, write 0, timeout 322
Requests/sec:      9.85
Transfer/sec:      1.13KB
  1. 按网上搜的结果加上了monkey patch
from gevent.pywsgi import WSGIServer
from gevent import monkey
from flask import Flask
from time import sleep

monkey.patch_all()

app = Flask(__name__)

@app.route('/health')
def health():
    sleep(0.1)
    return "ok"

if __name__ == "__main__":
    http_server = WSGIServer(('0.0.0.0', 8080), app)
    http_server.serve_forever()

wrk测试结果,加上monkey patch后似乎也没什么作用。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   182.89ms  209.48ms   1.82s    96.07%
    Req/Sec     5.55      3.50    10.00     47.89%
  592 requests in 1.00m, 67.64KB read
  Socket errors: connect 0, read 0, write 0, timeout 312
Requests/sec:      9.85
Transfer/sec:      1.13KB
  1. 正式上gunicorn,代码没有任何改动,也不需要引用gevent的WSGServer。
from flask import Flask
from time import sleep

app = Flask(__name__)

@app.get("/health")
def health():
    sleep(0.1)
    return "ok"

if __name__ == "__main__":
    app.run(host="0.0.0.0", port=8080)

运行命令。指定 -k gevent demo:app 中的 demo 是代码文件名。 --worker-connections 默认为1000

gunicorn demo:app -b 0.0.0.0:8080 -w 4 -k gevent --worker-connections 2000

wrk测试结果。性能相较于默认启动方式有了接近10倍的提升,请求响应时间也很稳定,最大响应时间也只有310.48。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   126.18ms    9.52ms 310.48ms   84.68%
    Req/Sec     3.98k   165.70     4.53k    77.34%
  948506 requests in 1.00m, 143.83MB read
Requests/sec:  15799.31
Transfer/sec:      2.40MB

其它框架和语言

在t4c2000的wrk配置下,flask+unicorn的每个进程基本都占用了85+%的CPU,再提高就得加CPU核心数了,不过这样的性能已经能满足公司内部服务的需求了,而且实际业务中,短板更可能是网络IO。

这里也测测其它语言和框架,看看Flask在Gunicorn的加持下能否打出python的牌面。

Golang

上来先试试最熟悉的Go, version: 1.22.4,使用标准库。(编译打包出来就能直接运行,不需要jvm这样的虚拟机,也不需要python这样的解释器,更不需要docker这样的容器运行时,特喜欢Go这一点)

package main

import (
	"fmt"
	"net/http"
	"time"
)

func MyHandler(w http.ResponseWriter, r *http.Request) {
	time.Sleep(time.Millisecond * 100)
	w.Write([]byte("ok"))
}

func main() {
	http.HandleFunc("/health", MyHandler)
	err := http.ListenAndServe("0.0.0.0:8080", nil)
	if err != nil {
		fmt.Println(err)
	}
}

wrk结果如下,请求量是目前测试以来第一个突破百万,而且也没有timeout的出现。使用top观察资源消耗,CPU只占用了约30%,而且还只有一个进程。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   101.56ms    1.48ms 121.62ms   77.58%
    Req/Sec     4.94k   191.07     5.05k    93.49%
  1180108 requests in 1.00m, 132.80MB read
Requests/sec:  19643.94
Transfer/sec:      2.21MB

不断加大连接,直到系统平均负载到达4(虚拟机CPU核心数为4)。连接数加了10倍,QPS差不多也是10倍于Flask + Gunicorn。这时候实际上wrk也占用了不少CPU资源,服务端的性能并没到瓶颈。

$ wrk -s bm.lua -t 4 -c20000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 20000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   167.56ms   43.45ms 397.95ms   60.37%
    Req/Sec    29.48k     8.84k   48.26k    63.05%
  6867733 requests in 1.00m, 772.85MB read
Requests/sec: 114258.70
Transfer/sec:     12.86MB

FastAPI

Go的性能已经很不错了,就性能来说还不是Flask+Gunicorn能媲美的。再来试试号称性能并肩Go的FastAPI(官网features里面写的)。FastAPI版本:0.115.4

纯uvicorn启动,用的是同步函数。

from fastapi import FastAPI
import uvicorn
from fastapi.responses import PlainTextResponse
from time import sleep

app = FastAPI()

@app.get("/health")
def index():
    sleep(0.1)
    return PlainTextResponse(status_code=200,content="ok")

if __name__ == '__main__':
    uvicorn.run(app, host="127.0.0.1", port=8080, access_log=False)

wrk测试结果,可以看到相当低下,甚至还不如flask的默认运行方式,超时请求数都过2w了。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.22s   438.35ms   1.95s    60.00%
    Req/Sec   115.03     57.31   410.00     86.69%
  23494 requests in 1.00m, 3.02MB read
  Socket errors: connect 0, read 0, write 0, timeout 22894
Requests/sec:    390.92
Transfer/sec:     51.54KB

改用异步函数再试试。

from fastapi import FastAPI
import uvicorn
from fastapi.responses import PlainTextResponse
from time import sleep
import asyncio

app = FastAPI()

@app.get("/health")
async def health():
    await asyncio.sleep(0.1)
    return PlainTextResponse(status_code=200,content="ok")

if __name__ == '__main__':
    uvicorn.run(app, host="127.0.0.1", port=8080, access_log=False)

wrk测试结果,可以看到性能好很多了,而且没有timeout。QPS是Flask默认启动方式的2倍,但实际性能应该不止2倍。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   484.53ms   63.18ms 654.48ms   61.26%
    Req/Sec     1.09k   698.29     3.13k    63.78%
  246744 requests in 1.00m, 31.77MB read
Requests/sec:   4106.80
Transfer/sec:    541.43KB

uvicorn支持指定worker数,这里设置为CPU核心数。

from fastapi import FastAPI
import uvicorn
from fastapi.responses import PlainTextResponse
from time import sleep
import asyncio

app = FastAPI()

@app.get("/health")
async def health():
    await asyncio.sleep(0.1)
    return PlainTextResponse(status_code=200,content="ok")

if __name__ == '__main__':
    uvicorn.run(app="demo2:app", host="127.0.0.1", port=8080, access_log=False, workers=4)

wrk测试结果,响应时间还是非常稳的,完全没有timeout的情况,延迟还更低。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   164.52ms   13.57ms 273.27ms   73.49%
    Req/Sec     3.05k   544.20     4.64k    68.67%
  727517 requests in 1.00m, 93.73MB read
Requests/sec:  12123.17
Transfer/sec:      1.56MB

gunicorn也支持uvicorn,看看fastapi在gunicorn的加持下会有怎样的性能表现。

gunicorn demo2:app -b 127.0.0.1:8080 -w 4 -k uvicorn.workers.UvicornWorker --worker-connections 2000

wrk测试结果,相较于unicorn运行方式,性能提升并不多。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   146.43ms   21.21ms 263.13ms   69.52%
    Req/Sec     3.43k   508.71     4.88k    71.40%
  818281 requests in 1.00m, 105.35MB read
Requests/sec:  13620.16
Transfer/sec:      1.75MB

Sanic

之前用过一段时间Sanic,也是个python异步框架,版本:24.6.0

from sanic import Sanic
from sanic.response import text
import asyncio

app = Sanic("HelloWorld")

@app.get("/health")
async def hello_world(request):
    await asyncio.sleep(0.1)
    return text("ok")

if __name__ == "__main__":
    app.run(host="127.0.0.1", port=8080, fast=True, debug=False, access_log=False)

wrk测试结果。虽然QPS比FastAPI高,但是有timeout的情况,不是很稳定。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   104.05ms   45.30ms   1.82s    99.59%
    Req/Sec     4.84k   538.65     5.07k    95.98%
  1154792 requests in 1.00m, 115.64MB read
  Socket errors: connect 0, read 0, write 0, timeout 88
Requests/sec:  19218.08
Transfer/sec:      1.92MB

Openresty

openresty基于nginx,通过集成lua,也可以用来写api。配置如下,只是增加了一个location,稍微调整下nginx的参数

worker_processes  auto;
worker_cpu_affinity auto;

events {
    worker_connections  65535;
}

http {
    include       mime.types;
    default_type  application/octet-stream;
    access_log off;
    sendfile        on;
    keepalive_timeout  65;

    server {
        listen       8080 deferred;
        server_name  localhost;
        location /health {
            content_by_lua_block {
                ngx.sleep(0.1)
                ngx.print("ok")
            }
        }
        location / {
            root   html;
            index  index.html index.htm;
        }
        error_page   500 502 503 504  /50x.html;
        location = /50x.html {
            root   html;
        }
    }
}

wrk测试结果,和Go语言相当。

$ wrk -s bm.lua -t 4 -c2000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://192.168.0.201:8080/health
  4 threads and 2000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   101.43ms    1.62ms 136.59ms   88.97%
    Req/Sec     4.94k   213.51     5.33k    92.65%
  1178854 requests in 1.00m, 211.36MB read
Requests/sec:  19619.86
Transfer/sec:      3.52MB

top 观察openresty的cpu占用并不高,加大连接再试试。连接数达到25000后,系统平均负载已经基本满了,而且wrk也占用了不少CPU资源。和Go差不多,并没有到服务端的性能瓶颈,而是受到系统资源限制。

$ wrk -s bm.lua -t 4 -c25000 -d60s http://127.0.0.1:8080/health
Running 1m test @ http://127.0.0.1:8080/health
  4 threads and 25000 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   149.17ms   36.61ms 659.40ms   71.22%
    Req/Sec    41.00k     6.64k   59.03k    65.95%
  9330399 requests in 1.00m, 1.63GB read
Requests/sec: 155277.58
Transfer/sec:     27.84MB

小结

整理下测试数据汇总成表格如下

项目 总请求量 每秒请求量 平均响应时间 最大响应时间 备注
Flask-no sleep 137538 2288.49 65.47ms 1.99s 有响应超时情况
Flask-同步方式 114440 1904.45 201.90ms 2.00s 有响应超时情况
Flask-异步函数 57896 964.31 275.49ms 2.00s 有响应超时情况
Flask+gevent 592 9.85 185.36ms 1.93s 有响应超时情况
Flask+gevent(monkeypatch) 592 9.85 182.89ms 1.82s 有响应超时情况
Flask+gevent+unicorn 948506 15799.31 126.18ms 310.48ms
Golang 1180108 19643.94 101.56ms 121.62ms
Golang 6867733 114258.70 167.56ms 397.95ms wrk的配置为t4c20000
FastAPI-同步函数 23494 390.92 1.22s 1.95s 有响应超时情况
FastAPI-异步函数 246744 4106.80 484.53ms 654.48ms
FastAPI-多worker 727517 12123.17 164.52ms 273.27ms
FastAPI+Gunicorn 818281 13620.16 146.43ms 273.27ms
Sanic 1154792 19218.08 104.05ms 1.82s 有响应超时情况,不是很稳定
OpenResty 1178854 19619.86 101.43ms 136.59ms
OpenResty 9330399 155277.58 149.17ms 659.40ms wrk的配置为t4c25000

根据测试结果,测试的三个Python Web框架中,Flask+gevent+unicorn综合最佳,不低的QPS,而且没有请求超时的情况,也不需要将代码修改成异步方式。Sanic的QPS虽高,但是有响应超时的情况,说明并不稳定,而且代码需要是异步的。FastAPI+Gunicorn的表现也不差,在不使用Gunicorn的情况下也能提供不错的性能,但代码同样需要改成异步方式。对于Sanic和FastAPI,Gunicorn的加持并不必要,而Gunicorn对Flask的性能提升至少7倍,而且能避免请求超时的情况,生产环境下应该尽量使用Gunicorn来运行Flask。

Go比各个Python框架的性能都更好,资源占用也更低,运行方式还更简单,不需要依赖编程语言环境和其他组件,非要说缺点的话就是开发没有Python快。

OpenResty的性能在测试中是最高的,主要是nginx本身性能良好。缺点是开发更麻烦。虽然是用lua开发,但lua作为动态语言,既不如Python极其灵活,还有动态语言本身代码不够清晰的缺点。以前尝试过用openresty实现一个crud服务,后来连自己都懒得维护就放弃了,干脆只用来当网关。

鱼与熊掌不可兼得,开发速度跟运行速度往往相斥,除非代码以后都是AI来写。就公司目前这服务的使用情况来说,Flask+Gunicorn的性能已经足够,还不需要改代码,实乃社畜良伴。而且现在啥都上k8s了,服务扩展也简单,性能不够就加实例嘛 ?

标签:游戏攻略