Lost/hanging requests with 3.6.2

[pvizeli]

Long story short

Based on: home-assistant/supervisor#1342

We run into an issue that requests produce timeouts and were never logged by our middleware they protocol/monitor requests. All instances with this issue running with an nginx proxy manager in front. Do we modify the URL with a parameter, it worked. It is very random. The only way to fix this issue was to downgrade aiohttp.

We can see the request is going to instance but we can't see it on the application layer. Maybe interesting ist that all the requests use cookies for session handling.

Expected behaviour

Everything work like 3.6.1

Actual behaviour

Random connections run into a timeouts.

Steps to reproduce

Run an nginx proxymanager in front of aiohttp 3.6.2 and use a site with cookies.

Your environment

Python 3.7 with uvloop and without. We use aioHTTP as server application in this case.

[webknjaz]

without uvloop doesn't have this issue

did you report this to uvloop?

[pvizeli]

EDIT: It look like without uvloop it doesn't have this issue.

I would do this after someone can point me to the part of code they changed on 3.6.2 and can affect this. Maybe @1st1 can assist me later.

EDIT: we had now the issue without uvloop @webknjaz. It is very random. So with 30 requests, 1-2 requests hangs.

[webknjaz]

@pvizeli any chance you could git bisect this? Also, please post the log of pip install, we need to know if you use C-extensions or not and if you're installing from wheels or building stuff during install time.

[webknjaz]

Source comparison: v3.6.1...v3.6.2#files_bucket

[webknjaz]

Changelog: https://docs.aiohttp.org/en/stable/changes.html#id1

[webknjaz]

Looking at https://pypi.org/project/aiohttp/3.6.1/#files vs. https://pypi.org/project/aiohttp/3.6.2/#files: you might've hit a wheel with py3-none-any marker which would mean that you're running a pure-python version so pip install is really important here.
Until you post this info, we have no idea what you're really running and consequently how to reproduce it.

[pvizeli]

We build our own wheel (on Alpine) and we use the C extension. I can get a freeze from that docker if you want? The affected container is homeassistant/amd64-hassio-supervisor:190

[pvizeli]

Not sure, but: https://github.com/mjpieters/aiohttp/blob/96e441a435a216a15e0b64ff140d01eb075761ec/aiohttp/client_proto.py#L179-L189

They now init a timeout handler on receive data, also if data are empty. I think that needs to be after None check of data.

[webknjaz]

Have you seen this? https://twitter.com/1st1/status/1187214384625541121?s=19

[webknjaz]

cc @mjpieters ^

[webknjaz]

@pvizeli if you could come up with a PR that works for you, that would be super helpful. This also needs regression tests so that it won't happen again.