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。对于性能调试,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}
使用以下命令运行 fastapi appi:
gunicorn api.performance_test:app -b localhost:8001 -k uvicorn.workers.UvicornWorker --workers 1
get to 的响应体http://localhost:8001/delay/0.0/0.0
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
我不断地得到这样的时间
{
"delays": [
0.0,
0.0
],
"total_time_taken": "0:00:00.002630",
"times": [
"0:00:00.000037",
...snip...
"0:00:00.000020"
]
}
当我取消注释该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,以创建代码执行速度的 10 倍差异。
值得一提的是,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pip freeze
输出的相关部分
fastapi==0.65.1
gunicorn==20.1.0
uvicorn==0.13.4
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
fastapi==0.65.1
gunicorn==20.1.0
uvicorn==0.14.0
我稍微修改了代码:
除了第一次加载网站外,我对两种方法的结果几乎相同。
这两种方法的时间介于两者之间,0:00:00.000530
并且0:00:00.000620
大部分时间都是如此。
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):
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}
带注释的非首次运行示例await asyncio.sleep(delay1)
(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]}
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:
# `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]}
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}
带注释的结果 我还制作了上述脚本的另一个版本,它每 1 个请求更改 url(它给出的时间略高): 结果: 带注释的结果 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。 另外请重新 启动您的计算机/服务器,它可能会产生影响。 更新 1: 我看到我使用的 uvicorn 版本 更新 2: 我运行了更多的基准测试,我注意到一些有趣的事情: 整个需求.txt: 结果: 整个需求.txt: 结果: 更新 3: 我只{'name': 'only uvicorn ', 'number_of_tests': 2000, 'to
{'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}{'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}await asyncio.sleep(delay1)
await asyncio.sleep(delay1)
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}){'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}{'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}0.14.0
比问题中所述的要新0.13.4
。我也用旧版本进行了测试,0.13.4
但结果相似,我仍然无法重现您的结果。在requirements.txt中使用uvloop:
在requirements.txt中没有uvloop:
Python 3.9.5
在这个答案中使用。uvicorn==0.14.0
fastapi==0.65.1
gunicorn==20.1.0
uvloop==0.15.2
uvicorn==0.14.0
fastapi==0.65.1
gunicorn==20.1.0