Synnefo

psomas used cstavr's method of reproducing the elusive AssertionErrors
and tracked the cause in a call to self.close() inside httplib itself.

1. get_http_connection() patches httplib's HTTPConnection's .close()
to a version that would just return the connection to the pool.

2. resp = conn.getresponse() calls self.close() which puts the
connection back in the pool.

3. body = resp.read() uses the now pooled connection

4. Another greenlet gets the connection from the pool and fails to use it
because sockets cannot be shared among greenlets

Notes:

• httplib closes calls self.close() because sometimes the
  HTTP Protocol says it should (e.g. Connection: close header)

• The choice to patch .close() instead of providing a distinct
  method .pool_release() was done so that no change would be made
  to the connection API.
  Clearly this was a wrong choice, not only because of this problem,
  but also since the API has been already severely breached since
  all pooled connection users are required to never let the connection
  objects fall out of scope without calling .close() within appropriate
  try statements.

• Since our real pooling target is sockets, not httplib connections,
  it would be better if we pooled sockets instead.
  However, since we don't directly handle sockets we must rely
  on the libraries themselves for socket pooling (httplib3?)

• It seems that renaming the patched method from .close()
  to another name, and using it as such will solve the problem.