https://github.com/eventlet/eventlet/issues/94
Because of the way paramiko utilises a client thread to manage its
communication, it's not been compatible with eventlet when run in
parallel.
It's not the only place these problems would arise.
This stemmed from the reuse of a fileno by the underlying OS.
Because listeners are registered against this descriptor, it would
be possible for old listeners to receive events destined for newer
descriptors; occasionally code would attempt to utilise the new
descriptor from a different greenlet, giving rise to the 'second
simultaneous read' problem.
Whenever a Python object is created to wrap one of these filenos,
we now signal the hub in order that it can correctly obsolete
extant listeners against that fileno. This is a fairly tricky
operation, due to the way that listeners' threads are interleaved
with the hub's operation - there are a number of small fixes here
to defend against one listener from effectively obsoleting another
when an event is pending against it.