Thread: [ADMIN] Postgres HA using Shared Disk Solution

On Fri, 4 Aug 2017 13:30:24 +0530
Purav Chovatia <puravc@gmail.com> wrote:

> Hello All,
>
> I have been researching for possible HA solutions for postgres. I find
> shared disk solution (either using Linux-HA or RHCS) to be a very good
> solution. Especially for us, because our databases are big and they do a
> lot of IO so we deploy a lot of disks.
>
> However, research on the internet shows that a shared disk solution is not
> very commonly deployed when it comes to postgres. People mostly go for
> Streaming Replication/Warm/Hot-Standby which to me are more like DR
> solutions.
>
> Is my research correct or that is not so? for e.g. this was the only post
> that I came across, and that too as old as 2012:
> https://www.postgresql.org/message-id/4F68E989.7080101%40d2.com
>
> Given our IO requirement, if we go for Streaming Replication instead, then
> our storage needs double and that increases the cost.
>
> Can someone confirm, if shared disk clustering is fairly common or no in
> the postgres world? If no, why is it so - can somebody enlighten me pls.

Shared disk is totally fine. It just requires a warm node useless for
day-to-day usage compare to HA on a streaming replication architecture where
the standby can be involved as a RO node.

Just make sure that:
* your SAN does not allow more than one server at a time to access the data
* you are able to fence the other node or its access to the disk
* extensive failover scenarios has been tested
* all required documentation has been written on your side to manage the cluster

As a bonus, make sure to use Pacemaker, with a recent distro (RHEL 7, CentOS 7,
Debian 9).

Regards,
--
Jehan-Guillaume de Rorthais
Dalibo

Purav Chovatia wrote:
> I have been researching for possible HA solutions for postgres. I find shared disk
> solution (either using Linux-HA or RHCS) to be a very good solution. Especially for us,
> because our databases are big and they do a lot of IO so we deploy a lot of disks.
> 
> However, research on the internet shows that a shared disk solution is not very commonly
> deployed when it comes to postgres. People mostly go for Streaming Replication/Warm/Hot-
> Standby which to me are more like DR solutions.
> 
> Is my research correct or that is not so? for e.g. this was the only post that I came
> across, and that too as old as 2012:
> https://www.postgresql.org/message-id/4F68E989.7080101%40d2.com
> 
> 
> Given our IO requirement, if we go for Streaming Replication instead, then our storage
> needs double and that increases the cost.
> 
> 
> Can someone confirm, if shared disk clustering is fairly common or no in the postgres
> world? If no, why is it so - can somebody enlighten me pls.

This is just my opinion, but I'd say that shared storage cluster solutions
have fallen from favor with the advent of streaming replication.

The less you share between cluster nodes, the fewer single points of failure you have.

I personally have seen a shared storage cluster fail because a hardware
problem caused the file system to get corrupted, and the file system
is a single point of failure with all shared storage solutions.

But yes, you'd have to double the storage.

Yours,
Laurenz Albe