This is the most minimal change for what might be relatively minimal
edge case. Alternative is replacing reload(), but let's not do that yet.
Closes#555
The idea behind transport=smart is to select between paramiko and
OpenSSH given the availability of connection multiplexing and/or OSX
kernel bugs. We need to make no such choice.
There has always been a race in PushFileService since given a parent
asked to forward modules to two children via some intermediary:
interm = router.local()
c1 = router.local(via=interm)
c2 = router.local(via=interm)
service.propagate_to(c1, 'foo/bar.py')
service.propagate_to(c2, 'foo/bar.py')
Two calls will be emitted to 'interm':
PushFileService.store_and_forward(c1, 'foo/bar.py', [blob])
PushFileService.store(c2, 'foo/bar.py')
Which will be processed in-order up to the point where service pool
threads in 'interm' are woken to process the message.
While it is guaranteed store_and_forward() will be processed first, no
guarantee existed that its assigned pool thread would wake and take
_lock first, thus it was possible for forward() to win the race, and for
a request to arrive to forward a file that had not been placed in local
cache yet.
Here we get rid of SerializedInvoker entirely, as it is partially to
blame for hiding the race: SerializedInvoker can only ensure no two
messages are processed simultaneously, it cannot ensure the messages are
processed in their intended order.
Instead, teach forward() that it may be called before
store_and_forward(), and if that is the case, to place the forward
request on to _waiters alongside any local threads blocked in get().
This was needed at some point in the past, but the tests don't seem to
care about it any more. We'll fix any CI breakage by changing the tests,
since verifying implicit localhost behaviour is important.
David Wilson