mostly the concern here is how this is trying to achieve a consistent state.
https://codereview.appspot.com/5909056/diff/1/juju/hooks/invoker.py
File juju/hooks/invoker.py (right):
https://codereview.appspot.com/5909056/diff/1/juju/hooks/invoker.py#newcode129
juju/hooks/invoker.py:129: self._relation_contexts_flush_order = []
why are we tracking the context order here? the changes being made are to a
resource (unit-settings) that is effectively locked for the duration of the hook
execution.
https://codereview.appspot.com/5909056/diff/1/juju/hooks/invoker.py#newcode285
juju/hooks/invoker.py:285: changes = yield context.flush()
for context in self._relation_context.values() given the above comment.
https://codereview.appspot.com/5909056/diff/1/juju/state/hook.py
File juju/state/hook.py (right):
https://codereview.appspot.com/5909056/diff/1/juju/state/hook.py#newcode127
juju/state/hook.py:127: 
we discussed this in our call, but fwiw.
its not clear that hook context should be responsible for this, as it manages a
single execution context. definitely passing in a topology to a state manager
raises a red flag.
there's inherently a race here, thats currently being avoided by passing around
the topology. afaics a cleaner way to do and avoid the topology passing is 1)
calculate the full rel set upfront, 2) this lets the individual hook contexts
continue to manage their own consistency on a per relation basis.
https://codereview.appspot.com/5909056/diff/1/juju/state/hook.py#newcode162
juju/state/hook.py:162: getattr(self, "_change", EMPTY_CHANGE),
empty change should go into the spec. charm authors would need to be aware of
this. several charms dispatch on change type. for non implicit hook contexts, i
think the change should always be none, else its propagating change information
out of context.
https://codereview.appspot.com/5909056/diff/1/juju/state/hook.py#newcode197
juju/state/hook.py:197: self._relation_id = relation_id if relation_id else
change.relation_id
why would it be passed in sometimes and on the change at others.
it feels like the construction should be uniform here, the caller is making the
distinction between implicit/implied contexts.
ie. relation_id is required.