Comment 0 for bug 1572719

You can reproduce this by issuing a GET request for a few hundred MB file and never consuming the response, but keep the client socket open. Swift will log a 499 but the socket does not always close.

ChunkWriteTimeout is meant to protect the proxy from a slow reading client:
  https://github.com/openstack/swift/blob/master/swift/proxy/controllers/base.py#L889-L905

Sometimes when this exception is thrown there is still data in the process socket buffer, so when eventlet tries to close the socket it first flushes it:
  https://github.com/eventlet/eventlet/blob/master/eventlet/wsgi.py#L631
  https://hg.python.org/cpython/file/v2.7.11/Lib/SocketServer.py#l711

The problem is that if the client is not consuming the socket buffer then that flush will wait forever; it's trying to write on a socket that just threw a timeout trying to write! The flush write is not protected by any timeout.

As far as I can tell, this WRITE_TIMEOUT does nothing:
  https://github.com/openstack/swift/blob/master/swift/common/wsgi.py#L407

wsgi.server() takes a socket_timeout that might be what we're after?
  https://github.com/openstack/swift/blob/master/swift/common/wsgi.py#L437-L440

Even with socket_timeout, eventlet needs to be patched. This should be in a finally block:
  https://github.com/eventlet/eventlet/blob/master/eventlet/wsgi.py#L636-L637

All of this is probably mitigated by most operators setting an idle timeout in their load balancers, but I wanted to report it. Going directly to a proxy I was able to hold sockets open for long periods of time.

I did the initial research on version 2.2.2 but I was able to reproduce on 2.7.0. I'm trying to translate links to master branch on github. I apologize in advance if it's not quite right.