Replication and on-line recovery

Hi!

I will be setting up a 3-way replicated server setup over the next
week or two, so this is probably a good time to make sure I know
exactly what I'm in for. To my knowledge, rserv is the only currently
available replication server for Postgres - is that correct?

1) Is there a way to designate 1 server as master, and do transparent
clustering by just connecting to that one server? Or does the
clustering have to be handled on the client end by making a connection
to each server and doing a "round-robin" for picking the server for
each query (multi-threaded application)?

2) If a server fails, how can it be resynchronised with the rest of
the cluster? I am asuming that inserts all have to be channeled
through the same server, in order to avoid race conditions. Is this
correct? If a server fails, and inserts are constantly being made,
then I am guessing that a dump/restore will not work properly if the
master (insert) server is not taken off-line and inserts are stopped.
Is this the case? How can this be done without taking the master
server off-line? Taking any server off line would take it out of sync,
so just doing a restore from another secondary server would then
result in two servers being out of sync. How is this worked around?

3) I know this has been asked before, and I've managed to get a few
responses about how to implement a quick and useable solution to this,
but I seem to have misplaced the emails, so please forgive me for
asking this again.

Is it possible to run Linux + Mosix + GFS to achieve the functionality
of a truly distributed database system? The bandwidth of communication
between the servers is not a huge problem, because I have the option
of connecting them either in a "star" configuration with 100 Mb
ethernet, connect them to a switch, or use a fiber link between them.
I've been told that "postmaster" won't get migrated properly due to
IPC and shared memory issues. Can anyone suggest a work-around? DIPC,
perhaps? I can't see how to work around the shared memory, though...

I know Oracle has a full distributed clustering support, but I have
made a decision to stick with open source software all the way (plus
the cost of running Oracle on a commercial setup is quite
prohibitive).

Still, even if the answer to the fully distributed database question
here is still just big, fat, flat "NO!", I'd really appreciate some
input regarding failure recovery and insert handling on a replicated
database cluster.

Regards.

Gordan