Re: real multi-master replication?

Stefan Kaltenbrunner <stefan(at)kaltenbrunner(dot)cc> wrote:
>
> Bill Moran wrote:
> > Stefan Kaltenbrunner <stefan(at)kaltenbrunner(dot)cc> wrote:
> >> Bill Moran wrote:
> >>> "hubert depesz lubaczewski" <depesz(at)gmail(dot)com> wrote:
> >>>> hi,
> >>>> i read about some replication system for postgresql, but - as far as i
> >>>> know there is none real multi-master replication system for
> >>>> postgresql.
> >>>> all i have seen are based on "query replication" with various "hacks"
> >>>> for specific constructions (like now()).
> >>>> my question is - is there any (even fully commercial) multi-master
> >>>> replication system for postgresql that will work with all possible
> >>>> constructs, triggers, random data and so on?
> >>>> i mean - i dont want to bother with choosing to 'note' somehow that
> >>>> 'this particular query' has to be replicated somehow.
> >>>> i'm thinking about working solution that will allow multi-master connections.
> >>>>
> >>>> anything? anywhere?
> >>> Have you looked at pgpool?
> >> afaik pgpool is statement based and not really multimaster either ...
> >
> > Well, it's multi-master to the degree that all servers are read/write,
> > and therefore any server can take over.
>
> not sure I follow - pgpool will simply replay the queries to each
> backend-server that are going through it.

I guess I'm comparing it to Slony, which has a clear delineation between
master and slave. With pgpool, you don't have the failover procedure in
the same way, in that each server can be read/write at all times.

> You cannot directly write to the servers (well you can - but that will
> likely cause inconsistent data) - and you have all the problems with
> non-determinstic queries as well as problems of getting a node back in
> sync after a downtime or connection loss.

Right.

> > How would you define multi-master?
>
> for true multimaster one would expect to be able to write to all the
> nodes and keep the data consistent/sync or have some sort of conflict
> resolution for an async solution.

Something more like pgcluster II.

--
Bill Moran
http://www.potentialtech.com