On Mon, Apr 21, 2008 at 04:57:27PM -0400, Christopher Browne wrote:

> > Our requirement was that nodes 1-3 were not allowed to know about or be
> > influenced by any events on nodes 5 and 6.  5 and 6 should communicate
> > only with node 4, which in turn would communicate with nodes 1-3.  Is
> > this possible?  Doesn't 'store path()' create listen paths between all
> > cluster nodes?
> 
> 
> Log data cannot be purged from nodes 1-3 until they *know* that the
> events have been processed by nodes 4, 5, and 6.
> 
> If you "win the fight" and get Slony-I configured to *not* have listen
> paths between 1-3 and 5,6, you'll find that things break down because
> obsolete data won't ever get purged.

This discussion seems to be hinting at a cool feature, which is a certain
amount of node isolation.  I think it won't work today, but it'd be a nifty
enhancement.

If I read him right, the OP described a network wherein 1, 2, 3 can talk to
4, and 5, 6 can talk to 4.  Moreover, 1,2,3 can all talk to each other, and
5,6 can talk to each other.  But communication between any of 1,2,3 and any
of 5,6 is restricted.

What is interesting about this is that you can use 4 as a gateway for
"translucency", so that data in 5,6 can be replicated through 4 such that
1,2,3 can get summaries, but not everything; and conversely.  

Now, there's no reason in principle, I think, why you couldn't just decide
that 5,6 simply do not need to know all the things going on in 1,2,3.  They
get their reports of "doneness" from 4, and they accept 4's statements as
gospel.  This would solve the "no purging" problem.

Whether the niftiness of this feature (and its potential benefits in
limited-permission data exchanges) would be worth the probably significant
complication is another matter entirely.  But I can certainly see the
application.

A