Performance of the Server tanked after upgrading to Connexion 3.x (Flask)

[Aakarshit-Sharma19]

After upgrading from Connexion2.x/Flask/Gunicorn/gevent to Connexion 3.x without gevent, the performance of the service has tanked, since it cannot now handle the same number of requests event though we see the number of threads have been increased multifolds. Even with the fafiling requests, The CPU footprint has almost doubled and memory footprint increased by 15%, even though we have started seeing failed requests during our internal testing.

Is there any resolution for the same, or do we need to move to starlette/asyncio stack to maintain performance?

More likely, the problem lies in the fact that we had gevent before as worker class with gunicorn, which was lighter than threadpool.

[alanbchristie]

Not sure I am experiencing the same but having just upgraded to 3.1.0 (from 2.x) I am finding that some requests just seem to block for precisely 90 seconds before even reaching the backend. According to the debug log the request just sits somewhere and then hits the backend code.

I'm using: -

• Python 3.13.2
• connexion 3.1.0 (with 4 workers in a container with 4 cores)
• gunicorn 23.0.0 (with uvicorn 0.32.1 worker class)
• Flask 3.1.0

With an API added like this...

connexion_app.add_api(
    _OPEN_API_FILE,
    strict_validation=Config.API_STRICT_VALIDATION,
    validate_responses=Config.API_VALIDATE_RESPONSES,
    pythonic_params=True,
)

... where Config.API_STRICT_VALIDATION and Config.API_VALIDATE_RESPONSES are both True

And an app launched using: -

gunicorn app.app:connexion_app \
    --worker-class uvicorn.workers.UvicornWorker \
    --workers ${WEB_CONCURRENCY} \
    --bind 0.0.0.0:8080

...where WEB_CONCURRENCY is 4

My OpenAPI is complex at 8,519 lines but it's fine with 2.x

Because I get no evidence that the API request is hitting my backend for 90 seconds I can only assume it's sitting for an enourmous amount of time somewhere in the API parser? (

[FelixSchwarz]

I guess you need to start profiling/instrumenting your Connexion app or at least publish a minimal version so others can try to find out why it is slower.