[Python] Fastapi python code execution speed impacted by deployment with uvicorn vs gunicorn

I wrote an fastapi app. And now I am thinking about deploying it however I seem to get strange unexpected performance issues that seem to depend on wether I use uvicorn vs gunicorn. In particular all code (even standard library pure python code) seems to get slower if I use gunicorn. For performance debugging I wrote a small app that demonstrates this:

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}


Running the fastapi appi with:

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


The resonse body of a get to http://localhost:8001/delay/0.0/0.0 is consistently something like:

{
"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"
]
}


However using:

uvicorn api.performance_test:app --port 8001


I consitently get timings like these

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


The difference becomes even more prounced when I uncomment the await asyncio.sleep(delay1) statement.

So I am wondering what gunicorn/uvicorn do to the python/fastapi runtime to create this factor 10 difference in the speed of code execution.

For what is is worth I performed these tests using Python 3.8.2 on OS X 11.2.3 with an intel I7 processor.

And these are the relevant parts of my pip freeze output

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

Continue reading...