Обсуждение: Adivice on master - master replication.

Hi
I'm new to Posgresql.
I have been requested to setup application level HA. The picture is, I
will have 2 web servers behind 2 new work load balancers, so should
one fail the NLB's will redirect traffic to the remaining server. I
need a similar scenario for the Pgsql servers.

From my reading I need a synchronise master - master. so both servers
need to be hot. Such that the web servers are unaware of which server
the are communicating with, so should one  pgsql server go down there
would be no interruption to the web servers.

The processes need to be handled automagicly , well at least the fail
over where the one server takes over the failed servers work load.

Everything I have looked at has a single point of failure, generally
the management application that looks after the replication.

How do I achieve this .. I'm looking for a comprehensive howto .

Thanks
G

On 10/28/10 3:42 PM, Gregory Machin wrote:
> Hi
> I'm new to Posgresql.
> I have been requested to setup application level HA. The picture is, I
> will have 2 web servers behind 2 new work load balancers, so should
> one fail the NLB's will redirect traffic to the remaining server. I
> need a similar scenario for the Pgsql servers.
>
> > From my reading I need a synchronise master - master. so both servers
> need to be hot. Such that the web servers are unaware of which server
> the are communicating with, so should one  pgsql server go down there
> would be no interruption to the web servers.
>
> The processes need to be handled automagicly , well at least the fail
> over where the one server takes over the failed servers work load.
>
> Everything I have looked at has a single point of failure, generally
> the management application that looks after the replication.
>
> How do I achieve this .. I'm looking for a comprehensive howto .

the traditional method is a HA cluster with master/standy, sharing
storage via a fully redundant and dual pathed SAN.  all ethernet
connections are dual pathed, too, with redundant network switches.   the
master and slave run a cluster management package, such as linux
heartbeat, or veritas cluster service, which has distributed management
capability.   for robustness, you use a fencing systemm on the shared
storage so only one server is physically capable of accessing the shared
storage (this prevents a class off failures where the other server also
thinks its active and tries to mount and access the database storage).
The cluster manager also shares a single IP such that it reaches the
current master server, and your webfarm uses that IP for all database
operations.