WSGIMiddleware overhead

[gaganpreet]

I'm exploring the feasibility of moving an existing Flask app to FastAPI while running the old Flask app using the WSGIMiddleware.

I've been running some performance tests and there's a very significant overhead that WSGIMiddleware brings to the mix. I wrote a small demo here.

This is how I run the server:

uvicorn main:app --port 8000 --no-access-log  --workers 1

There are two endpoints: /users_fastapi (direct, no middleware) and /users_flask (wsgi middleware). Both endpoints query some dummy data from the in-memory db and return it as a response.

$ wrk http://127.0.0.1:8000/users_fastapi

Running 10s test @ http://127.0.0.1:8000/users_fastapi
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   568.68ms   68.02ms 637.11ms   95.32%
    Req/Sec    12.06      7.11    20.00     36.21%
  171 requests in 10.01s, 16.15MB read
Requests/sec:     17.08
Transfer/sec:      1.61MB

$ wrk http://127.0.0.1:8000/users_flask

Running 10s test @ http://127.0.0.1:8000/users_flask
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.24s   157.89ms   1.56s    77.03%
    Req/Sec     7.58      5.52    20.00     64.06%
  74 requests in 10.01s, 6.99MB read
Requests/sec:      7.39
Transfer/sec:    714.88KB

The direct route has a throughput of 17.08 reqs/s, and the one behind wsgi middleware only goes up to 7.39, which is a significant difference.

I'm looking at the middleware and trying to figure out if there's room to reduce overhead, however, any inputs or ideas are much appreciated.

[abersheeran]

Maybe you can get a better comparison by testing the performance of the original flask.

You are welcome to submit any PRs that improve performance. Here's a performance benchmark I just ran.

$ pytest ./benchmark.py -s
=================================================================================== test session starts ===================================================================================
platform win32 -- Python 3.9.6, pytest-7.0.1, pluggy-1.0.0
rootdir: C:\Users\aber\Documents\GitHub\a2wsgi
plugins: anyio-3.5.0, asyncio-0.11.0, cov-3.0.0
collected 6 items

benchmark.py
             Name              Average Time
          pure-ASGI            0.00030796286.
    a2wsgi-WSGIMiddleware      0.0008904933199999999.
    uvicorn-WSGIMiddleware     0.00062742427.
      asgiref-WsgiToAsgi       0.00141565296.
          pure-WSGI            0.00023801917.
    a2wsgi-ASGIMiddleware      0.0008830748.

[gaganpreet]

Maybe you can get a better comparison by testing the performance of the original flask.

I tried this.

gunicorn main:flask_app --bind=:8000 -w 1

$ wrk http://localhost:8000/users_flask

Running 10s test @ http://localhost:8000/users_flask
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   414.07ms   43.31ms 480.73ms   94.07%
    Req/Sec    13.61      7.62    30.00     70.07%
  236 requests in 10.01s, 22.29MB read
Requests/sec:     23.58
Transfer/sec:      2.23MB

At 23.58 reqs/s, pure Flask is even faster than the ASGI implementation. This is not a fair comparison of ASGI vs WSGI, as the ASGI code is running synchronous code.

You are welcome to submit any PRs that improve performance.

I haven't found anything yet to improve performance, I'll spend some time this weekend and see what's possible.