[Spread-users] Spread for HA rpc services

[Colin Meyer]

On Wed, Aug 08, 2007 at 04:47:02PM -0400, John Robinson wrote:
> Colin Meyer wrote:
> 
> >It seems easy enough for the non-lead processes to ignore requests. What
> >happens when the current lead dies unexpectedly? The pool will receive
> >the disconnect message, but how do all of the processes come to a
> >consensus on which process is going to be the next lead?
> 
> Spread makes this pretty easy.  Each new membership list arrives at the 
> same time, and is in the same order relative to regular messages (with 
> high enough delivery class), at every process.  So simply designate the 
> new process 0 as the leader from that point forward.  In-band membership 
> change messages make this work nicely without need of a distributed 
> voting algorithm.

Right. Each listening process can easily perform the same calculations
and come to the same conclusion of who is the current leader (as well as
who else is in the pool). That makes perfect sense.

I am still evaluating whether or not to use Spread for the message queue
for my rpc. The main argument against it is the extra maintainence. Each
client of the rpc will need to be running the spread daemon. The main
argument for it is that Spread will easily fill the need that each
client will be able to contact a live server without needing to be aware
of the pool configuration.

Thanks a bunch to everyone who has answered this thread.

-Colin.