Python异步框架大战：FastAPI、Sanic、Tornado VS Go 的 Gin

一、前言

异步编程在构建高性能 Web 应用中起着关键作用，而 FastAPI、Sanic、Tornado 都声称具有卓越的性能。本文将通过性能压测对这些框架与Go的Gin框架进行全面对比，揭示它们之间的差异。

原文：Python异步框架大战：FastAPI、Sanic、Tornado VS Go 的 Gin

二、环境准备

系统环境配置

编程语言

语言	版本	官网/Github
Python	3.10.12	https://www.python.org/
Go	1.20.5	https://go.dev/

压测工具

工具	介绍	官网/Github
ab	Apache的压力测试工具,使用简单	https://httpd.apache.org/docs/2.4/programs/ab.html
wrk	高性能多线程压力测试工具	https://github.com/wg/wrk
JMeter	功能强大的压力/负载测试工具	https://github.com/apache/jmeter

这里选择 wrk 工具进行压测，mac 安装直接通过brew快速安装

brew install wrk

window安装可能要依赖它的子系统才方便安装，或者换成其他的压测工具例如JMeter。

web框架

框架	介绍	压测版本	官网/Github
FastAPI	基于Python的高性能web框架	0.103.1	https://fastapi.tiangolo.com/
Sanic	Python的异步web服务器框架	23.6.0	https://sanic.dev/zh/
Tornado	Python的非阻塞式web框架	6.3.3	https://www.tornadoweb.org/en/stable/
Gin	Go语言的web框架	1.9.1	https://gin-gonic.com/
Fiber	todo	todo	https://gofiber.io/
Flask	todo	todo	https://github.com/pallets/flask
Django	todo	todo	https://www.djangoproject.com/

数据库配置

数据库名	介绍	压测版本	依赖库
MySQL	关系型数据库	8.0	sqlalchemy+aiomysql
Redis	NoSQL数据库	7.2	aioredis

三、wrk 工具 http压测

FastAPI

普通http请求压测

依赖安装

pip install fastapi==0.103.1
pip install uvicorn==0.23.2

编写测试路由

from fastapi import FastAPIapp = FastAPI(summary="fastapi性能测试")@app.get(path="/http/fastapi/test")
async def fastapi_test():return {"code": 0, "message": "fastapi_http_test", "data": {}}

Uvicorn 运行，这里是起四个进程运行部署

uvicorn fastapi_test:app --log-level critical --port 8000 --workers 4

wrk压测

开20个线程，建立500个连接，持续请求30s

wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/test

压测结果

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/testRunning 30s test @ http://127.0.0.1:8000/http/fastapi/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     3.06ms    2.89ms  36.65ms   85.34%Req/Sec     3.85k     3.15k   41.59k    70.05%2298746 requests in 30.11s, 383.64MB readSocket errors: connect 267, read 100, write 0, timeout 0Requests/sec:  76357.51
Transfer/sec:     12.74MB

Thread Stats 这里是 20、30个压测线程的平均结果指标

• 平均延迟(Avg Latency):每个线程的平均响应延迟
• 标准差(Stdev Latency):每个线程延迟的标准差
• 最大延迟(Max Latency):每个线程遇到的最大延迟
• 延迟分布(+/- Stdev Latency):每个线程延迟分布情况
• 每秒请求数(Req/Sec):每个线程每秒完成的请求数
• 请求数分布(+/- Stdev Req/Sec):每个线程请求数的分布情况

Socket errors: connect 267, read 100, write 0, timeout 0，是压测过程中socket的错误统计

• connect:连接错误,表示在压测过程中,总共有 267 次连接异常
• read:读取错误,表示有 100 次读取数据异常
• write:写入错误,表示有0次写入异常
• timeout:超时错误,表示有0次超时

MySQL数据查询请求压测

这里在简单试下数据库查询时候的情况

首先先补充下项目依赖

pip install hui-tools[db-orm, db-redis]==0.2.0

hui-tools是我自己开发的一个工具库，欢迎大家一起来贡献。https://github.com/HuiDBK/py-tools

#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Author: Hui
# @Desc: { fastapi性能测试 }
# @Date: 2023/09/10 12:24
import uvicorn
from fastapi import FastAPI
from py_tools.connections.db.mysql import SQLAlchemyManager, DBManagerapp = FastAPI(summary="fastapi性能测试")async def init_orm():db_client = SQLAlchemyManager(host="127.0.0.1",port=3306,user="root",password="123456",db_name="house_rental")db_client.init_mysql_engine()DBManager.init_db_client(db_client)@app.on_event("startup")
async def startup_event():"""项目启动时准备环境"""await init_orm()@app.get(path="/http/fastapi/mysql/test")
async def fastapi_mysql_query_test():sql = "select id, username, role from user_basic where username='hui'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))return {"code": 0, "message": "fastapi_http_test", "data": {**user_info}}

wrk压测

wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/mysql/test

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/mysql/testRunning 30s test @ http://127.0.0.1:8000/http/fastapi/mysql/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency    38.81ms   19.35ms 226.42ms   76.86%Req/Sec   317.65    227.19   848.00     57.21%180255 requests in 30.09s, 36.95MB readSocket errors: connect 267, read 239, write 0, timeout 0Non-2xx or 3xx responses: 140Requests/sec:   5989.59
Transfer/sec:      1.23MB

可以发现就加入一个简单的数据库查询，QPS从 76357.51 降到 5989.59 足足降了有10倍多，其实是单机数据库处理不过来太多请求，并发的瓶颈是在数据库，可以尝试加个redis缓存对比MySQL来说并发提升了多少。

Redis缓存查询压测

#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Author: Hui
# @Desc: { fastapi性能测试 }
# @Date: 2023/09/10 12:24
import json
from datetime import timedeltaimport uvicorn
from fastapi import FastAPI
from py_tools.connections.db.mysql import SQLAlchemyManager, DBManager
from py_tools.connections.db.redis_client import RedisManagerapp = FastAPI(summary="fastapi性能测试")async def init_orm():db_client = SQLAlchemyManager(host="127.0.0.1",port=3306,user="root",password="123456",db_name="house_rental")db_client.init_mysql_engine()DBManager.init_db_client(db_client)async def init_redis():RedisManager.init_redis_client(async_client=True,host="127.0.0.1",port=6379,db=0,)@app.on_event("startup")
async def startup_event():"""项目启动时准备环境"""await init_orm()await init_redis()@app.get(path="/http/fastapi/redis/{username}")
async def fastapi_redis_query_test(username: str):# 先判断缓存有没有user_info = await RedisManager.client.get(name=username)if user_info:user_info = json.loads(user_info)return {"code": 0, "message": "fastapi_redis_test", "data": {**user_info}}sql = f"select id, username, role from user_basic where username='{username}'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))# 存入redis缓存中, 3minawait RedisManager.client.set(name=user_info.get("username"),value=json.dumps(user_info),ex=timedelta(minutes=3))return {"code": 0, "message": "fastapi_redis_test", "data": {**user_info}}if __name__ == '__main__':uvicorn.run(app)

运行

wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/redis/hui

结果

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8000/http/fastapi/redis/huiRunning 30s test @ http://127.0.0.1:8000/http/fastapi/redis/hui20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     9.60ms    5.59ms 126.63ms   88.41%Req/Sec     1.22k     0.91k    3.45k    57.54%730083 requests in 30.10s, 149.70MB readSocket errors: connect 267, read 101, write 0, timeout 0Requests/sec:  24257.09
Transfer/sec:      4.97MB

缓存信息

添加了redis缓存，并发能力也提升了不少，因此在业务开发中一些查多改少的数据可以适当的做缓存。

压测结论

压测类型	测试时长	线程数	连接数	请求总数	QPS	平均延迟	最大延迟	总流量	吞吐量/s
普通请求	30s	20	500	2298746	76357.51	3.06ms	36.65ms	383.64MB	12.74MB
MySQL查询	30s	20	500	730083	5989.59	38.81ms	226.42ms	36.95MB	1.23MB
Redis缓存	30s	20	500	730083	24257.09	9.60ms	126.63ms	149.70MB	4.97MB

给 mysql 查询加了个 redis 缓存 qps 提升了 3倍多，对于一些查多改少的数据，根据业务设置适当的缓存可以大大提升系统的吞吐能力。其他框架我就直接上代码测，就不一一赘述了，直接看结果指标。

Sanic

压测方式都是一样的我就不像fastapi一样的一个一个写了，直接写全部压测然后看结果

环境安装

pip install sanic==23.6.0
pip install hui-tools'[db-orm, db-redis]'==0.2.0

编写测试路由

#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Author: Hui
# @Desc: { sanic性能测试 }
# @Date: 2023/09/10 12:24
import json
from datetime import timedeltafrom py_tools.connections.db.mysql import SQLAlchemyManager, DBManager
from py_tools.connections.db.redis_client import RedisManager
from sanic import Sanic
from sanic.response import json as sanic_jsonapp = Sanic("sanic_test")async def init_orm():db_client = SQLAlchemyManager(host="127.0.0.1",port=3306,user="root",password="123456",db_name="house_rental")db_client.init_mysql_engine()DBManager.init_db_client(db_client)async def init_redis():RedisManager.init_redis_client(async_client=True,host="127.0.0.1",port=6379,db=0,)@app.listener('before_server_start')
async def server_start_event(app, loop):await init_orm()await init_redis()@app.get(uri="/http/sanic/test")
async def fastapi_test(req):return sanic_json({"code": 0, "message": "sanic_http_test", "data": {}})@app.get(uri="/http/sanic/mysql/test")
async def sanic_myql_query_test(req):sql = "select id, username, role from user_basic where username='hui'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))return sanic_json({"code": 0, "message": "sanic_mysql_test", "data": {**user_info}})@app.get(uri="/http/sanic/redis/<username>")
async def sanic_redis_query_test(req, username: str):# 先判断缓存有没有user_info = await RedisManager.client.get(name=username)if user_info:user_info = json.loads(user_info)return sanic_json({"code": 0, "message": "sanic_redis_test", "data": {**user_info}})sql = f"select id, username, role from user_basic where username='{username}'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))# 存入redis缓存中, 3minawait RedisManager.client.set(name=user_info.get("username"),value=json.dumps(user_info),ex=timedelta(minutes=3))return sanic_json({"code": 0, "message": "sanic_redis_test", "data": {**user_info}})def main():app.run()if __name__ == '__main__':# sanic sanic_test.app -p 8001 -w 4 --access-log=Falsemain()

运行

Sanic 内置了一个生产web服务器，可以直接使用

sanic python.sanic_test.app -p 8001 -w 4 --access-log=False

普通http请求压测

同样是起了四个进程看看性能如何

wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/test

压测结果

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/testRunning 30s test @ http://127.0.0.1:8001/http/sanic/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     1.93ms    2.20ms  61.89ms   91.96%Req/Sec     6.10k     3.80k   27.08k    69.37%3651099 requests in 30.10s, 497.92MB readSocket errors: connect 267, read 163, write 0, timeout 0Requests/sec: 121286.47
Transfer/sec:     16.54MB

Sanic 果然性能很强，在python中估计数一数二了。

mysql数据查询请求压测

运行

wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/mysql/test

结果

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/mysql/testRunning 30s test @ http://127.0.0.1:8001/http/sanic/mysql/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency    35.22ms   21.75ms 264.37ms   78.52%Req/Sec   333.14    230.95     1.05k    68.99%198925 requests in 30.10s, 34.72MB readSocket errors: connect 267, read 146, write 0, timeout 0Requests/sec:   6609.65
Transfer/sec:      1.15MB

Redis缓存查询压测

运行

wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/redis/hui

结果

➜  ~ wrk -t20 -d30s -c500 http://127.0.0.1:8001/http/sanic/redis/huiRunning 30s test @ http://127.0.0.1:8001/http/sanic/redis/hui20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     6.91ms    4.13ms 217.47ms   95.62%Req/Sec     1.71k     0.88k    4.28k    68.05%1022884 requests in 30.09s, 178.52MB readSocket errors: connect 267, read 163, write 0, timeout 0Requests/sec:  33997.96
Transfer/sec:      5.93MB

压测结论

压测类型	测试时长	线程数	连接数	请求总数	QPS	平均延迟	最大延迟	总流量	吞吐量/s
普通请求	30s	20	500	3651099	121286.47	1.93ms	61.89ms	497.92MB	16.54MB
MySQL查询	30s	20	500	198925	6609.65	35.22ms	264.37ms	34.72MB	1.15MB
Redis缓存	30s	20	500	1022884	33997.96	6.91ms	217.47ms	178.52MB	5.93MB

Tornado

环境安装

pip install tornado==6.3.3
pip install gunicorn==21.2.0
pip install hui-tools[db-orm, db-redis]==0.2.0

编写测试路由

#!/usr/bin/python3
# -*- coding: utf-8 -*-
# @Author: Hui
# @Desc: { tornado 性能测试 }
# @Date: 2023/09/20 22:42
import asyncio
from datetime import timedelta
import json
import tornado.web
import tornado.ioloop
from tornado.httpserver import HTTPServer
from py_tools.connections.db.mysql import SQLAlchemyManager, DBManager
from py_tools.connections.db.redis_client import RedisManagerclass TornadoBaseHandler(tornado.web.RequestHandler):passclass TornadoTestHandler(TornadoBaseHandler):async def get(self):self.write({"code": 0, "message": "tornado_http_test", "data": {}})class TornadoMySQLTestHandler(TornadoBaseHandler):async def get(self):sql = "select id, username, role from user_basic where username='hui'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))self.write({"code": 0, "message": "tornado_mysql_test", "data": {**user_info}})class TornadoRedisTestHandler(TornadoBaseHandler):async def get(self, username):user_info = await RedisManager.client.get(name=username)if user_info:user_info = json.loads(user_info)self.write({"code": 0, "message": "tornado_redis_test", "data": {**user_info}})returnsql = f"select id, username, role from user_basic where username='{username}'"ret = await DBManager().run_sql(sql)column_names = [desc[0] for desc in ret.cursor.description]result_tuple = ret.fetchone()user_info = dict(zip(column_names, result_tuple))# 存入redis缓存中, 3minawait RedisManager.client.set(name=user_info.get("username"),value=json.dumps(user_info),ex=timedelta(minutes=3),)self.write({"code": 0, "message": "tornado_redis_test", "data": {**user_info}})def init_orm():db_client = SQLAlchemyManager(host="127.0.0.1",port=3306,user="root",password="123456",db_name="house_rental",)db_client.init_mysql_engine()DBManager.init_db_client(db_client)def init_redis():RedisManager.init_redis_client(async_client=True,host="127.0.0.1",port=6379,db=0,)def init_setup():init_orm()init_redis()def make_app():init_setup()return tornado.web.Application([(r"/http/tornado/test", TornadoTestHandler),(r"/http/tornado/mysql/test", TornadoMySQLTestHandler),(r"/http/tornado/redis/(.*)", TornadoRedisTestHandler),])app = make_app()async def main():# init_setup()# app = make_app()server = HTTPServer(app)server.bind(8002)# server.start(4) # start 4 worker# app.listen(8002)await asyncio.Event().wait()if __name__ == "__main__":# gunicorn -k tornado -w=4 -b=127.0.0.1:8002 python.tornado_test:appasyncio.run(main())

运行tornado服务

gunicorn -k tornado -w=4 -b=127.0.0.1:8002 python.tornado_test:app

wrk 压测

wrk -t20 -d30s -c500 http://127.0.0.1:8002/http/tornado/testwrk -t20 -d30s -c500 http://127.0.0.1:8002/http/tornado/mysql/testwrk -t20 -d30s -c500 http://127.0.0.1:8002/http/tornado/redis/hui

结果

➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8002 /http/tornado/test
Running 30s test @ http://127.0.0.1:8002/http/tornado/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     6.54ms    1.92ms  34.75ms   63.85%Req/Sec     1.79k     1.07k    3.83k    56.23%1068205 requests in 30.07s, 280.15MB readSocket errors: connect 267, read 98, write 0, timeout 0Requests/sec:  35525.38
Transfer/sec:      9.32MB➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8002 /http/tornado/mysql/test
Running 30s test @ http://127.0.0.1:8002/http/tornado/mysql/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency    41.29ms   16.51ms 250.81ms   71.45%Req/Sec   283.47    188.81     0.95k    65.31%169471 requests in 30.09s, 51.88MB readSocket errors: connect 267, read 105, write 0, timeout 0Requests/sec:   5631.76
Transfer/sec:      1.72MB➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8002 /http/tornado/redis/hui
Running 30s test @ http://127.0.0.1:8002/http/tornado/redis/hui20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency    11.69ms    3.83ms 125.75ms   78.27%Req/Sec     1.00k   537.85     2.20k    64.34%599840 requests in 30.07s, 183.63MB readSocket errors: connect 267, read 97, write 0, timeout 0Non-2xx or 3xx responses: 2Requests/sec:  19947.28
Transfer/sec:      6.11MB

Gin

环境安装

go get "github.com/gin-gonic/gin"
go get "github.com/go-redis/redis"
go get "gorm.io/driver/mysql"
go get "gorm.io/gorm"

代码编写

package mainimport ("encoding/json""time""github.com/gin-gonic/gin""github.com/go-redis/redis""gorm.io/driver/mysql""gorm.io/gorm""gorm.io/gorm/logger"
)var (db          *gorm.DBredisClient *redis.Client
)type UserBasic struct {Id       int    `json:"id"`Username string `json:"username"`Role     string `json:"role"`
}func (UserBasic) TableName() string {return "user_basic"
}func initDB() *gorm.DB {var err errordb, err = gorm.Open(mysql.Open("root:123456@/house_rental"), &gorm.Config{// 将LogMode设置为logger.Silent以禁用日志打印Logger: logger.Default.LogMode(logger.Silent),})if err != nil {panic("failed to connect database")}sqlDB, err := db.DB()// SetMaxIdleConns sets the maximum number of connections in the idle connection pool.sqlDB.SetMaxIdleConns(10)// SetMaxOpenConns sets the maximum number of open connections to the database.sqlDB.SetMaxOpenConns(30)// SetConnMaxLifetime sets the maximum amount of time a connection may be reused.sqlDB.SetConnMaxLifetime(time.Hour)return db
}func initRedis() *redis.Client {redisClient = redis.NewClient(&redis.Options{Addr: "localhost:6379",})return redisClient
}func jsonTestHandler(c *gin.Context) {c.JSON(200, gin.H{"code": 0, "message": "gin json", "data": make(map[string]any),})
}func mysqlQueryHandler(c *gin.Context) {// 查询语句var user UserBasicdb.First(&user, "username = ?", "hui")//fmt.Println(user)// 返回响应c.JSON(200, gin.H{"code":    0,"message": "go mysql test","data":    user,})}func cacheQueryHandler(c *gin.Context) {// 从Redis中获取缓存username := "hui" // 要查询的用户名cachedUser, err := redisClient.Get(username).Result()if err == nil {// 缓存存在，将缓存结果返回给客户端var user UserBasic_ = json.Unmarshal([]byte(cachedUser), &user)c.JSON(200, gin.H{"code":    0,"message": "gin redis test","data":    user,})return}// 缓存不存在，执行数据库查询var user UserBasicdb.First(&user, "username = ?", username)// 将查询结果保存到Redis缓存userJSON, _ := json.Marshal(user)redisClient.Set(username, userJSON, time.Minute*2)// 返回响应c.JSON(200, gin.H{"code":    0,"message": "gin redis test","data":    user,})
}func initDao() {initDB()initRedis()
}func main() {//r := gin.Default()r := gin.New()gin.SetMode(gin.ReleaseMode) // 生产模式initDao()r.GET("/http/gin/test", jsonTestHandler)r.GET("/http/gin/mysql/test", mysqlQueryHandler)r.GET("/http/gin/redis/test", cacheQueryHandler)r.Run("127.0.0.1:8003")
}

wrk 压测

wrk -t20 -d30s -c500 http: //127.0.0.1:8003/http/gin/testwrk -t20 -d30s -c500 http: //127.0.0.1:8003/http/gin/mysql/testwrk -t20 -d30s -c500 http: //127.0.0.1:8003/http/gin/redis/test

结果

➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8003 /http/gin/test
Running 30s test @ http://127.0.0.1:8003/http/gin/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     2.45ms    5.68ms 186.48ms   91.70%Req/Sec     6.36k     5.62k   53.15k    83.99%3787808 requests in 30.10s, 592.42MB readSocket errors: connect 267, read 95, write 0, timeout 0Requests/sec: 125855.41
Transfer/sec:     19.68MB➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8003 /http/gin/mysql/test
Running 30s test @ http://127.0.0.1:8003/http/gin/mysql/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency    40.89ms   83.70ms   1.12s    90.99%Req/Sec   522.33    322.88     1.72k    64.84%308836 requests in 30.10s, 61.26MB readSocket errors: connect 267, read 100, write 0, timeout 0Requests/sec:  10260.63
Transfer/sec:      2.04MB
➜  ~➜  ~ wrk -t20 -d30s -c500 http:// 127.0.0.1 : 8003 /http/gin/redis/test
Running 30s test @ http://127.0.0.1:8003/http/gin/redis/test20 threads and 500 connectionsThread Stats   Avg      Stdev     Max   +/- StdevLatency     7.18ms    1.76ms  79.40ms   81.93%Req/Sec     1.63k     1.09k    4.34k    62.59%972272 requests in 30.10s, 193.79MB readSocket errors: connect 267, read 104, write 0, timeout 0Requests/sec:  32305.30
Transfer/sec:      6.44MB

四、总结

web框架	压测类型	测试时长	线程数	连接数	请求总数	QPS	平均延迟	最大延迟	总流量	吞吐量/s
FastAPI	普通请求	30s	20	500	2298746（229w）	76357.51 （76k）	3.06ms	36.65ms	383.64MB	12.74MB
	MySQL查询	30s	20	500	180255 (18w)	5989.59 (5.9k)	38.81ms	226.42ms	36.95MB	1.23MB
	Redis缓存	30s	20	500	730083 (73w)	24257.09 (24k)	9.60ms	126.63ms	149.70MB	4.97MB
Sanic	普通请求	30s	20	500	3651099（365w）	121286.47（120k）	1.93ms	61.89ms	497.92MB	16.54MB
	MySQL查询	30s	20	500	198925 （19w）	6609.65 （6k）	35.22ms	264.37ms	34.72MB	1.15MB
	Redis缓存	30s	20	500	1022884（100w）	33997.96 （33k）	6.91ms	217.47ms	178.52MB	5.93MB
Tornado	普通请求	30s	20	500	1068205（106w）	35525.38（35k）	6.54ms	34.75ms	280.15MB	9.32MB
	MySQL查询	30s	20	500	169471 （16w）	5631.76 （5.6k）	41.29ms	250.81ms	51.88MB	1.72MB
	Redis缓存	30s	20	500	599840 （59w）	19947.28 （19k）	11.69ms	125.75ms	183.63MB	6.11MB
Gin	普通请求	30s	20	500	3787808（378w）	125855.41（125k）	2.45ms	186.48ms	592.42MB	19.68MB
	MySQL查询	30s	20	500	308836 （30w）	10260.63 （10k）	40.89ms	1.12s	61.26MB	2.04MB
	Redis缓存	30s	20	500	972272 （97w）	32305.30（32k）	7.18ms	79.40ms	193.79MB	6.44MB

性能

从性能角度来看，各个Web框架的表现如下：

Gin > Sanic > FastAPI > Tornado

Gin：在普通请求方面表现最佳，具有最高的QPS和吞吐量。在MySQL查询中，性能很高，但最大延迟也相对较高。gin承受的并发请求最高有 1w qps，其他python框架都在5-6k qps，但gin的mysql查询请求最大延迟达到了1.12s， 虽然可以接受这么多并发请求，但单机mysql还是处理不过来。

还有非常重要的一点，cpython的多线程由于GIL原因不能充分利用多核CPU，故而都是通过开了四个进程来处理请求，资源开销远远大于go的gin，go底层的GMP的调度策略很强，天然支持并发。

注意：Python使用asyncio语法时切记不要使用同步IO操作不然会堵塞住主线程的事件loop，从而大大降低性能，如果没有异步库支持可以采用线程来处理同步IO。

综合评价

除了性能之外，还有其他因素需要考虑，例如框架的社区活跃性、生态系统、文档质量以及团队熟悉度等。这些因素也应该在选择Web框架时考虑。

最终的选择应该基于具体需求和项目要求。如果性能是最重要的因素之一，那么Sanic和go的一些框架可能是不错的选择。如果您更关注其他方面的因素，可以考虑框架的社区支持和适用性。我个人还是挺喜欢使用FastAPI。

五、测试源代码

https://github.com/HuiDBK/WebFrameworkPressureTest

Github上已经有其他语言的web框架的压测，感兴趣也可以去了解下： https://web-frameworks-benchmark.netlify.app/result

不知道为啥他们测试的python性能好低，可能异步没用对😄