Treat asyncio resume_reading error as a closed connection

[cdeler]

Closes encode/httpx#1213

There is a gist in the issue, which can help you reproduce the error: link . You can also switch the async backend and check, that the error appears only when we asyncio backend.

The introduced fix mends the error, so that the error message caused by the exception looks like:

httpcore.RemoteProtocolError: peer closed connection without sending complete message body (received 917504 bytes, expected 1000000)

instead of the original stacktrace from the issue

There is a problem:
I cannot figure out how to make a unit test for this error

[florimondmanca]

Hm, okay, sounds like a sensible ugly fix to an asyncio weakness.

Note: have we tried to see if this reproduces on anyio (rubbing in asyncio)?

[cdeler]

@florimondmanca
Having updated the gist, I tried to reproduce the error versus "anyio" backend, but couldn't. I'll try to figure out if we can do something with it no way, from my point of view the problem is inside asyncio.sslproto. Technically we can try to replace TLS implementation for asyncio backend

PS
Thats miracle, thank you @agronholm for "anyio"

[florimondmanca]

Hello! I've got some nitpicks on the code comment… Other than that, I'm not sure how confident we are merging this change? Can you point here to a comment with a test setup that allows reproducing this 100% of the time, to verify this is fixed with this PR?

[florimondmanca]

(Nit) Tweaking the leading spaces here…

Suggested change

	# We have to return empty string there,
	# due to one ugly bug in asyncio
	# More information you can find in this issue:
	# https://github.com/encode/httpx/issues/1213
	#
	# The empty byte-string returned from there caused
	# httpcore.RemoteProtocolError at the top level
	# We have to return empty string there,
	# due to one ugly bug in asyncio
	# More information you can find in this issue:
	# https://github.com/encode/httpx/issues/1213
	#
	# The empty byte-string returned from there caused
	# httpcore.RemoteProtocolError at the top level

[cdeler]

Fixed

[cdeler]

@florimondmanca

Can you point here to a comment with a test setup that allows reproducing this 100% of the time, to verify this is fixed with this PR?

I created a gist with the test file

https://gist.github.com/cdeler/86d99a42c5bfb522bc15cac8d27824aa

Just run the script using python3, also you can check that the bug is not reproducible under other backends (just run path_to_script --help to check other switches)

[cdeler]

@florimondmanca

Probably I can add the test from the gist to our CI

[tomchristie]

Lovely, thanks!