Fastapi python代码执行速度受uvicorn vs gunicorn部署的影响

Question

我写了一个 fastapi 应用程序。现在我正在考虑部署它，但是我似乎遇到了奇怪的意外性能问题，这似乎取决于我使用 uvicorn 还是 gunicorn。特别是如果我使用 gunicorn，所有代码（甚至标准库纯 python 代码）似乎都会变慢。为了进行性能调试，我编写了一个小应用程序来演示这一点：

import asyncio, time
from fastapi import FastAPI, Path
from datetime import datetime

app = FastAPI()

@app.get("/delay/{delay1}/{delay2}")
async def get_delay(
    delay1: float = Path(..., title="Nonblocking time taken to respond"),
    delay2: float = Path(..., title="Blocking time taken to respond"),
):
    total_start_time = datetime.now()
    times = []
    for i in range(100):
        start_time = datetime.now()
        await asyncio.sleep(delay1)
        time.sleep(delay2)
        times.append(str(datetime.now()-start_time))
    return {"delays":[delay1,delay2],"total_time_taken":str(datetime.now()-total_start_time),"times":times}

运行 fastapi 应用程序：

gunicorn api.performance_test:app -b localhost:8001 -k uvicorn.workers.UvicornWorker --workers 1

到达http://localhost:8001/delay/0.0/0.0 的共鸣体始终是这样的：

{
  "delays": [
    0.0,
    0.0
  ],
  "total_time_taken": "0:00:00.057946",
  "times": [
    "0:00:00.000323",
    ...smilar values omitted for brevity...
    "0:00:00.000274"
  ]
}

但是使用：

uvicorn api.performance_test:app --port 8001 

我经常得到这样的时间

{
  "delays": [
    0.0,
    0.0
  ],
  "total_time_taken": "0:00:00.002630",
  "times": [
    "0:00:00.000037",
    ...snip...
    "0:00:00.000020"
  ]
}

当我取消注释 await asyncio.sleep(delay1) 语句时，差异变得更加明显。

所以我想知道 gunicorn/uvicorn 对 python/fastapi 运行时做了什么来使代码执行速度产生 10 倍的差异。

我在 OS X 11.2.3 和英特尔 I7 处理器上使用 Python 3.8.2 执行了这些测试。

这些是我的pip freeze 输出的相关部分

fastapi==0.65.1
gunicorn==20.1.0
uvicorn==0.13.4

Answer 1

我无法重现您的结果。

我的环境： Windows 10 上 WSL2 上的 ubuntu

我的pip freeze 输出的相关部分：

fastapi==0.65.1
gunicorn==20.1.0
uvicorn==0.14.0

---

我稍微修改了代码：

import asyncio, time
from fastapi import FastAPI, Path
from datetime import datetime
import statistics

app = FastAPI()

@app.get("/delay/{delay1}/{delay2}")
async def get_delay(
    delay1: float = Path(..., title="Nonblocking time taken to respond"),
    delay2: float = Path(..., title="Blocking time taken to respond"),
):
    total_start_time = datetime.now()
    times = []
    for i in range(100):
        start_time = datetime.now()
        await asyncio.sleep(delay1)
        time.sleep(delay2)
        time_delta= (datetime.now()-start_time).microseconds
        times.append(time_delta)

    times_average = statistics.mean(times)

    return {"delays":[delay1,delay2],"total_time_taken":(datetime.now()-total_start_time).microseconds,"times_avarage":times_average,"times":times}

除了第一次加载网站外，两种方法的结果几乎相同。

这两种方法的大部分时间都在0:00:00.000530 和0:00:00.000620 之间。

每个的第一次尝试需要更长的时间：大约0:00:00.003000。 但是，在我重新启动 Windows 并再次尝试这些测试后，我注意到服务器启动后首次请求的时间不再增加（我认为这要归功于重新启动后有大量可用 RAM）

---

非首次运行示例（3 次尝试）：

# `uvicorn performance_test:app --port 8083`

{"delays":[0.0,0.0],"total_time_taken":553,"times_avarage":4.4,"times":[15,7,5,4,4,4,4,5,5,4,4,5,4,4,5,4,4,5,4,4,5,4,4,5,4,4,4,5,4,4,5,4,4,5,4,4,4,4,4,5,4,5,5,4,4,4,4,4,4,5,4,4,4,5,4,4,4,4,4,4,5,4,4,5,4,4,4,4,5,4,4,5,4,4,4,4,4,5,4,4,5,4,4,5,4,4,5,4,4,4,4,4,4,4,5,4,4,4,5,4]}
{"delays":[0.0,0.0],"total_time_taken":575,"times_avarage":4.61,"times":[15,6,5,5,5,5,5,5,5,5,5,4,5,5,5,5,4,4,4,4,4,5,5,5,4,5,4,4,4,5,5,5,4,5,5,4,4,4,4,5,5,5,5,4,4,4,4,5,5,4,4,4,4,4,4,4,4,5,5,4,4,4,4,5,5,5,5,5,5,5,4,4,4,4,5,5,4,5,5,4,4,4,4,4,4,5,5,5,4,4,4,4,5,5,5,5,4,4,4,4]}
{"delays":[0.0,0.0],"total_time_taken":548,"times_avarage":4.31,"times":[14,6,5,4,4,4,4,4,4,4,5,4,4,4,4,4,4,5,4,4,5,4,4,4,4,4,4,4,5,4,4,4,5,4,4,4,4,4,4,4,4,5,4,4,4,4,4,4,5,4,4,4,4,4,5,5,4,4,4,4,4,4,4,5,4,4,4,4,4,5,4,4,5,4,4,5,4,4,5,4,4,4,4,4,4,4,5,4,4,5,4,4,5,4,4,5,4,4,4,4]}


# `gunicorn performance_test:app -b localhost:8084 -k uvicorn.workers.UvicornWorker --workers 1`

{"delays":[0.0,0.0],"total_time_taken":551,"times_avarage":4.34,"times":[13,6,5,5,5,5,5,4,4,4,5,4,4,4,4,4,5,4,4,5,4,4,5,4,4,4,4,4,5,4,4,4,4,4,5,4,4,4,4,4,4,4,5,4,4,5,4,4,4,4,4,4,4,4,5,4,4,4,4,4,4,4,5,4,4,4,4,4,4,4,4,4,5,4,4,5,4,5,4,4,5,4,4,4,4,5,4,4,5,4,4,4,4,4,4,4,5,4,4,5]}
{"delays":[0.0,0.0],"total_time_taken":558,"times_avarage":4.48,"times":[14,7,5,5,5,5,5,5,4,4,4,4,4,4,5,5,4,4,4,4,5,4,4,4,5,5,4,4,4,5,5,4,4,4,5,4,4,4,5,5,4,4,4,4,5,5,4,4,5,5,4,4,5,5,4,4,4,5,4,4,5,4,4,5,5,4,4,4,5,4,4,4,5,4,4,4,5,4,5,4,4,4,5,4,4,4,5,4,4,4,5,4,4,4,5,4,4,4,5,4]}
{"delays":[0.0,0.0],"total_time_taken":550,"times_avarage":4.34,"times":[15,6,5,4,4,4,4,4,4,5,4,4,4,4,4,5,4,4,5,4,4,5,4,4,4,4,4,5,4,4,4,4,5,5,4,4,4,4,5,4,4,4,4,4,5,4,4,5,4,4,5,4,4,5,4,4,5,4,4,5,4,4,4,4,4,4,5,4,4,5,4,4,4,4,4,4,4,4,4,5,4,4,5,4,4,4,4,4,4,4,4,5,4,4,5,4,4,4,4,4]}

---

带有注释await asyncio.sleep(delay1) 的非首次运行示例（3 次尝试）：

# `uvicorn performance_test:app --port 8083`

{"delays":[0.0,0.0],"total_time_taken":159,"times_avarage":0.6,"times":[3,1,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,0,0,1,1,0,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,0,0,1,0,0,0,0,0,1,1,1,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,0]}
{"delays":[0.0,0.0],"total_time_taken":162,"times_avarage":0.49,"times":[3,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,0,1,0,0,0,0,1,1,1,1,1,0,0,0,0,1,1,1,1,0,0,1,0,0,0,0,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,1,0,0,0,0,1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0,1,1]}
{"delays":[0.0,0.0],"total_time_taken":156,"times_avarage":0.61,"times":[3,1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,1,0,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1]}


# `gunicorn performance_test:app -b localhost:8084 -k uvicorn.workers.UvicornWorker --workers 1`

{"delays":[0.0,0.0],"total_time_taken":159,"times_avarage":0.59,"times":[2,0,0,0,0,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,0,0,0,0,1,0,1,1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,0,1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,1,1,1,1,1,0,0,0,0,1,1,1,1,1,0,0]}
{"delays":[0.0,0.0],"total_time_taken":165,"times_avarage":0.62,"times":[3,1,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,1,0,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0,1,1,1,1,1]}
{"delays":[0.0,0.0],"total_time_taken":164,"times_avarage":0.54,"times":[2,0,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0,0,1,1,1,1,1,0,0,0,1,1,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1,1]}

我制作了一个 Python 脚本来更精确地对这些时间进行基准测试：

import statistics
import requests
from time import sleep

number_of_tests=1000

sites_to_test=[
    {
        'name':'only uvicorn    ',
        'url':'http://127.0.0.1:8083/delay/0.0/0.0'
    },
    {
        'name':'gunicorn+uvicorn',
        'url':'http://127.0.0.1:8084/delay/0.0/0.0'
    }]


for test in sites_to_test:

    total_time_taken_list=[]
    times_avarage_list=[]

    requests.get(test['url']) # first request may be slower, so better to not measure it

    for a in range(number_of_tests):
        r = requests.get(test['url'])
        json= r.json()

        total_time_taken_list.append(json['total_time_taken'])
        times_avarage_list.append(json['times_avarage'])
        # sleep(1) # results are slightly different with sleep between requests

    total_time_taken_avarage=statistics.mean(total_time_taken_list)
    times_avarage_avarage=statistics.mean(times_avarage_list)

    print({'name':test['name'], 'number_of_tests':number_of_tests, 'total_time_taken_avarage':total_time_taken_avarage, 'times_avarage_avarage':times_avarage_avarage})

结果：

{'name': 'only uvicorn    ', 'number_of_tests': 2000, 'total_time_taken_avarage': 586.5985, 'times_avarage_avarage': 4.820865}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 2000, 'total_time_taken_avarage': 571.8415, 'times_avarage_avarage': 4.719035}

带有注释await asyncio.sleep(delay1)的结果

{'name': 'only uvicorn    ', 'number_of_tests': 2000, 'total_time_taken_avarage': 151.301, 'times_avarage_avarage': 0.602495}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 2000, 'total_time_taken_avarage': 144.4655, 'times_avarage_avarage': 0.59196}

我还制作了上述脚本的另一个版本，它每 1 个请求更改 url（它给出的时间稍长）：

import statistics
import requests
from time import sleep

number_of_tests=1000

sites_to_test=[
    {
        'name':'only uvicorn    ',
        'url':'http://127.0.0.1:8083/delay/0.0/0.0',
        'total_time_taken_list':[],
        'times_avarage_list':[]
    },
    {
        'name':'gunicorn+uvicorn',
        'url':'http://127.0.0.1:8084/delay/0.0/0.0',
        'total_time_taken_list':[],
        'times_avarage_list':[]
    }]


for test in sites_to_test:
    requests.get(test['url']) # first request may be slower, so better to not measure it

for a in range(number_of_tests):

    for test in sites_to_test:
        r = requests.get(test['url'])
        json= r.json()

        test['total_time_taken_list'].append(json['total_time_taken'])
        test['times_avarage_list'].append(json['times_avarage'])
        # sleep(1) # results are slightly different with sleep between requests


for test in sites_to_test:
    total_time_taken_avarage=statistics.mean(test['total_time_taken_list'])
    times_avarage_avarage=statistics.mean(test['times_avarage_list'])

    print({'name':test['name'], 'number_of_tests':number_of_tests, 'total_time_taken_avarage':total_time_taken_avarage, 'times_avarage_avarage':times_avarage_avarage})

结果：

{'name': 'only uvicorn    ', 'number_of_tests': 2000, 'total_time_taken_avarage': 589.4315, 'times_avarage_avarage': 4.789385}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 2000, 'total_time_taken_avarage': 589.0915, 'times_avarage_avarage': 4.761095}

带有注释await asyncio.sleep(delay1)的结果

{'name': 'only uvicorn    ', 'number_of_tests': 2000, 'total_time_taken_avarage': 152.8365, 'times_avarage_avarage': 0.59173}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 2000, 'total_time_taken_avarage': 154.4525, 'times_avarage_avarage': 0.59768}

这个答案应该可以帮助您更好地调试结果。

如果您分享有关您的操作系统/机器的更多详细信息，我认为这可能有助于调查您的结果。

另外请重启你的电脑/服务器，可能会有影响。

---

更新 1：

我发现我使用了更新版本的 uvicorn 0.14.0，而不是问题 0.13.4 中所述的版本。 我还用旧版本0.13.4 进行了测试，但结果相似，我仍然无法重现您的结果。

---

更新 2：

我运行了一些基准测试，发现了一些有趣的事情：

在 requirements.txt 中使用 uvloop：

整个 requirements.txt：

uvicorn==0.14.0
fastapi==0.65.1
gunicorn==20.1.0
uvloop==0.15.2

结果：

{'name': 'only uvicorn    ', 'number_of_tests': 500, 'total_time_taken_avarage': 362.038, 'times_avarage_avarage': 2.54142}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 500, 'total_time_taken_avarage': 366.814, 'times_avarage_avarage': 2.56766}

在 requirements.txt 中没有 uvloop：

整个 requirements.txt：

uvicorn==0.14.0
fastapi==0.65.1
gunicorn==20.1.0

结果：

{'name': 'only uvicorn    ', 'number_of_tests': 500, 'total_time_taken_avarage': 595.578, 'times_avarage_avarage': 4.83828}
{'name': 'gunicorn+uvicorn', 'number_of_tests': 500, 'total_time_taken_avarage': 584.64, 'times_avarage_avarage': 4.7155}

---

更新 3：

我在这个答案中只使用了Python 3.9.5。

Answer 2

差异在于您使用的底层网络服务器。

类比可以是：two cars, same brand, same options, just a different engine, what's the difference?

Web 服务器与汽车不完全一样，但我想你明白我要表达的意思。

基本上，gunicorn 是synchronous 网络服务器，而uvicorn 是asynchronous 网络服务器。由于您使用的是fastapi 和await 关键字，我想您已经知道asyncio/asynchornous programming 是什么。

我不知道代码的差异，所以我的回答要持保留态度，但uvicorn 的性能更高，因为asynchronous 部分。我对时间差异的猜测是，如果您使用async web 服务器，它已经在启动时配置为处理async 功能，而如果您使用sync web 服务器，它不是并且有某种开销以抽象该部分。

这不是一个正确的答案，但它会提示您可能存在的差异。

Answer 3

由于fastapi 是一个ASGI 框架，因此它可以通过ASGI 服务器（如uvicorn 或hypercorn）提供更好的性能。 WSGI 像 gunicorn 这样的服务器无法提供像 uvicorn 这样的性能。 ASGI 服务器针对 asynchronous 功能进行了优化。 fastapi的官方文档也鼓励使用ASGI服务器，如uvicorn或hypercorn。

https://fastapi.tiangolo.com/#installation