eod status 27-jun-2012
William Reade
william.reade at canonical.com
Thu Jun 28 08:21:18 UTC 2012
On Thu, 2012-06-28 at 00:33 -0700, Gustavo Niemeyer wrote:
> > evidence for this, I would like to observe that it was only after making
> > the change that it became clear that most of the watchers just don't
> > bother checking whether a given content change actually represents a
> > real change in state.
>
> I don't understand this. I see most of the watchers necessarily doing
> diffing since they have to deliver a delta only. How can they possibly
> not be considering whether the change is real?
In several cases, we don't handle the situation in which a node change
and revert happen close together and are processed as a single event by
ZK; affected watchers are ConfigNodeWatcher, NeedsUpgradeWatcher,
ResolvedWatcher and PortsWatcher. The impact and likelihood of each of
these is arguable, but I'd consider them more valuable abstractions if
they did actually guarantee that only real changes would be emitted. The
other 3 are fine, I think (but FlagWatcher was itself only fixed
recently).
Incidentally, the (maybe) problematic watchers are ofc the ones that
send complete data, rather than diffs; and anything that's handling
their output should probably be able to deal with non-change events just
fine; this is why I'm not sure if fixing this will pay off (beyond
giving us a fuzzy sense of smugness about the quality of our
abstractions). Hence request for opinions.
> (...)
> > Ofc, if other considerations intrude, I have plenty of interesting bugs
> > related to deploy that I can work on tomorrow, so don't feel too
> > pressured :).
>
> Thanks again. I have missed just one branch from you and one from
> Frank, but otherwise all reviewed.
Many thanks!
Cheers
William
More information about the Juju-dev
mailing list