On Tue, 2003-02-25 at 23:53, Ed L. wrote:
> On Tuesday February 25 2003 11:52, Tom Lane wrote:
> >
> > Also, there are nontrivial licensing issues involved. The PG-R
> > design depends on an underlying "group communication" system, which
> > is a nontrivial bit of software that none of the core team wants to
> > rewrite. But none of the available GC systems are BSD-license open
> > source. We had had some hopes of getting Spread to offer BSD
> > terms, but that seems to have fallen through. So right now, PG-R
> > is on the outside looking in, as far as inclusion in the core
> > distribution goes :-(
>
> Is anyone aware of particular reasons why the group is pushing on a
> syncronous solution? I'm sure they have good reasons, but I would've
> assumed an asyncronous solution would be far more applicable for
> simple redundancy as opposed to syncronicity for high-performance
> clusters, not too mention being far simpler implementations.
Some business needs absolutely must have synchronus replication. You're
right that most users want async so they can have clusters serving data
out, but there are some very good reasons for sync:
1. Financial transactions MUST have off-site sync replication.
2. If you have a cluster of servers, sync guarantees that all the data
are in a consistent state. Ease of administration and reliability may
be worth the slight performance penalty.
3. For servers on a fast LAN, sync may not make much of a difference on
performance.
I personally think that sync should be used in more cases than is
generally thought. We can get very reliable LANs and even WAN links,
and if I know that the data are always consistent, that makes my life
much much easier than having to worry about cases where they are not.
