How to handle client disconnection?

[vikramarsid]

Hello,

Using Nginx + Gunicorn(gthreads) + Falcon. How should I detect if a client has disconnected from the API server? Is there a way to handle client-side disconnection to cleanup upstream database jobs?

Thanks for the help.

Cheers.

CC: @vytas7

@vytas7 says: adding a checklist.

• [ ] Explain when this is relevant (i.e. when the application server gets a chance to handle this) at all -- streaming vs fully buffering requests by a front-end proxy (such as Nginx)
• [ ] How does Gunicorn behave upon disconnect? On timeout?
• [ ] How does uWSGI behave upon disconnect? On timeout?
• [ ] Check if anything needs to be clarified for ASGI.

[kgriffs]

Hi @vikramarsid, you could see if gunicorn triggers any of its server hooks in that case, although that would depend on NGINX prematurely disconnecting from gunicorn rather than eating the response, which AFAICT is what happens by default (see also: http://nginx.org/en/docs/http/ngx_http_proxy_module.html#proxy_ignore_client_abort).

@jmvrbanac has an example of using hooks here: https://github.com/jmvrbanac/falcon-example/blob/master/example/app.py

[vytas7]

See also #1800

[vytas7]

I believe we should add some documentation to our FAQ (or elsewhere) how to handle this in popular WSGI servers (at least Gunicorn and uWSGI).

As per the discussion above, it's also worth adding a note that this only applies if Nginx is configured to stream requests to the application server. While this is not a common setting for JSON APIs (typically set to buffer + small max body size), it is essential for file uploads, proxying large requests etc.

Uvicorn (and ASGI in general) handles this more explicitly out of the box, but we could check if it deserves a mention too.

[vytas7]

FWIW, this is what uWSGI seems to be doing:

2020-12-14 18:28:00 [FALCON] [ERROR] POST / => Traceback (most recent call last):
  File "falcon/app.py", line 331, in falcon.app.App.__call__
  File "test.py", line 10, in on_post
    chunk = req.bounded_stream.read(4096)
  File "falcon/request_helpers.py", line 264, in falcon.request_helpers.BoundedStream.read
  File "falcon/request_helpers.py", line 238, in falcon.request_helpers.BoundedStream._read
OSError: error during read(4096) on wsgi.input

And Gunicorn simply returns less data when expected, when doing req.stream.read() (or bounded_stream.read()). Which is ugly IMHO!

I tried skimming once again through PEP 3333, and the spec is woefully silent on these issues... So I guess there's no right and wrong.

[metallerok]

Hi @vikramarsid, you could see if gunicorn triggers any of its server hooks in that case, although that would depend on NGINX prematurely disconnecting from gunicorn rather than eating the response, which AFAICT is what happens by default (see also: http://nginx.org/en/docs/http/ngx_http_proxy_module.html#proxy_ignore_client_abort).

@jmvrbanac has an example of using hooks here: https://github.com/jmvrbanac/falcon-example/blob/master/example/app.py

This example gunicorn hooks doesn't work