David Parker wrote:

>Sigh. OK. Can you point me to where exactly that code is? I see a delete
>from sl_log_1 in cleanupThread_main, but I'm not sure when that gets
>invoked? 
>  
>
That gets run roughly every 10 minutes.

>It just seems like you should be able to say that these 2 nodes
>subscribe to the same set, but don't care about each other. It's
>replication without failover semantics. I realize the log shipping is
>designed to address this, but I'm not quite seeing why the listening
>relationships could not be made lighter....
>  
>
We'd have to make the sl_confirm table aware of activity on a set-by-set 
basis, which probably implies having each sync set be broken up on a 
by-replication-set basis.  That would be rather a large change.

In any case, the listening relationships between nodes that _aren't_ 
interested in one another (e.g. - where there's no direct 
provider/subscriber relationship) shouldn't be overly expensive.

There's probably room for some tuning of slon parameters; thinking out 
loud, there's the check interval and interval timeouts; it may be that 
increasing these times on a distant "slave node" would diminish the 
effort that would be generated on notifications.

I expect there's some tuning that can be done, and that this would be 
worthwhile, as, irrespective of whether the cost is too high or not, 
there sure are a lot of pg_listener entries generated in a day.  That's 
probably worth an IRC discussion at some point.

>Of course, there's plenty that I don't understand, so I probably just
>need to blunder around in the code for a while hurting myself before I
>realize why things have to be the way they are!
>  
>
Unfortunately, there's little other way of figuring out carefully 
designed systems :-(.