mstate: add primary watchers

mstate: add primary watchers

This change adds {Machines,Services,ServiceUnits,ServiceRelations}Watcher.
Each watcher has the test ported from state, and a new, more comprehensive
test.

https://code.launchpad.net/~aramh/juju-core/70-mstate-watchers-nwo/+merge/124072

(do not edit description out of merge proposal)

[aram, 2012-09-12 23:00:02]

The check for the initial event is commented out in tests because watchers don't
deliver it yet. The reason is that mstate/watcher won't deliver an empty even at
client watch instantiation. mstate/watcher will only deliver something when
there's something to be delivered, but client watchers need to deliver an empty
initial event right away if the database is empty.

There are various fixes for this, but I'd rather change mstate/watcher behavior.

[aram, 2012-09-12 23:01:55]

Please take a look.

[fwereade, 2012-09-13 12:17:37]

I don't see any Kill()s, and I see Addeds and Removeds everywhere. I thought all
these watchers were meant to be becoming lifecycle watchers, and all the events
were meant to just have LifeChanged (or maybe just Changed) lists?

https://codereview.appspot.com/6492110/diff/3001/mstate/service_test.go
File mstate/service_test.go (right):

https://codereview.appspot.com/6492110/diff/3001/mstate/service_test.go#newco...
mstate/service_test.go:237: // BUG(aram): mstate/watcher doesn't deliver empty
initial event.
Pretty serious bug IMO... I'd prefer to fix the foundations before building on
top of them. Is that fix in progress?

[fwereade, 2012-09-14 06:35:47]

On 2012/09/13 12:17:37, fwereade wrote:
> I don't see any Kill()s, and I see Addeds and Removeds everywhere. I thought
all
> these watchers were meant to be becoming lifecycle watchers, and all the
events
> were meant to just have LifeChanged (or maybe just Changed) lists?

Sorry, I meant to follow this up yesterday.

I accept that matching APIs is the sanest way forward; I am, personally, a
little freaked out at the prospect of further delays before the agents can start
to respond to lifecycles correctly, but I don't have any better ideas.

I still consider the no-initial-event thing to be a bit of a showstopper,
though: AIUI that assumption is baked in throughout the code. Do we know why
it's happening, of if it's easy to fix, at all?

[niemeyer, 2012-09-17 09:33:35]

There are a few important issues in the implementation, discussed live.

As we talked, rejecting in favor of individual watches per branch.