Scott Marlowe wrote:

> In my usage, all I've seen on the subscribers is that there's no data
>
>there until the set finishes syncing up.  There's no problem accessing
>the empty tables, just nothing to see there.
>  
>
In 1.1.5, it gets more interesting...

It uses TRUNCATE, which will grab a full lock on each table as it works
on it.

The Evil Condition that we have seen in production is that we'll have
monitoring/maintenance scripts that get in edgewise, and hold some sort
of access lock that ultimately leads to the SUBSCRIBE SET event hitting
a deadlock, and restarting.

This happened last week; we were rebuilding a node, and it wasn't
sufficiently locked down, with the result that a monitoring script got
in and busted the SUBSCRIBE SET event.

I went in and added in pg_hba.conf rules similar to:

host all slony 0.0.0.0 0.0.0.0 md5
host all all 0.0.0.0 0.0.0.0 reject

I'd have to say, based on this, that there *is* a problem with accessing
possibly-empty tables.  The "lock 'em fully, lock 'em early" approach
presently in CVS HEAD would make the above policy unnecessary.