fix(h2_connection_reset): add stream reset when client cancel the req…

[MarkLux]

Summary

add the missing connection reset of h2 connection when client cancel the request.

fix#943

in discussion with #941

[lovelydinosaur]

Thanks, seems like a good start. Make sure to run scripts/unasync.py to also apply the changes to the httpcore/_sync/http2.py module.

https://github.com/encode/httpcore/actions/runs/10464774808/job/28978912844#step:5:26

(Related... maybe we could change the build logs when unasync --check fails, so that there's clearer feedback to the developer.)

[MarkLux]

sorry to bother but we are really in a hurry to fix this one, could you have a review again? @tomchristie

[MarkLux]

langchain-ai#5

approved by langchain community now, shall we have any further progress on this one? @tomchristie

[lovelydinosaur]

need send cancel frame when the exception is not from remote peer.

It's not obvious to me that not isinstance(exc, RemoteProtocolError) is correct/sufficient here.

Should this instead be something like just wrapping the yield?...

try:
yieldchunkexceptExceptionasexc:
self._connection._reset_steam(
stream_id=self._stream_id,
error_code=h2.settings.ErrorCodes.CANCEL,
)
raiseexc

[mabrowning]

BaseException covers asyncio.CancelledError as well, which should also reset the stream.

[lovelydinosaur]

approved by langchain community now, shall we have any further progress on this one?

Thanks @MarkLux - apologies for the delay it's a bit of a complex one to review.

Thoughts on my review comment?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[mabrowning]

This PR or something like it is still needed. Even with the latest version of httpcore, if the client closes the response before the server is done, the client does not send any RST_STREAM frames; the server will continue to generate and stream data and the client will ignore it.

[mabrowning]

After updating the local h2_state, we should write the updated bytes on the wire, otherwise its reliant on some (possibly much) later write (like initiating a new request or a keep-alive frame) sending the reset.