Hi Jehan-Guillaume,
Thanks for your opinion.
At first glance, i may use for automatic failover PAF, a proxy HAproxy and for fencincg, i am a bit disappointed, i don't know what to do/use
How about you, do you have any preference about tools/solutions to use ?
now, I am aware that i will have to check and document limitation/risk...
Le mer. 5 sept. 2018 à 11:38, Jehan-Guillaume (ioguix) de Rorthais <ioguix@xxxxxxx> a écrit :
Hi all,
On Tue, 4 Sep 2018 15:09:51 +0000
ROS Didier <didier.ros@xxxxxx> wrote:
> Hi
> I have made a lot of PostgreSQL High Availability tests (more
> than 20 by solution) and the two following products respond well to the need :
>
> (1) Repmgr (2ndQuadrant)
>
> (2) Pglookout (aiven)
Both solutions use a simple and naive implementation, which makes them easy to
use and admin. However, it gives the responsibilities to the admin to deal with
fencing, which is a mandatory piece in almost all kind of DB cluster if you
want to cover most of the failure cases and avoid split brain.
So yes, they are simple, because complexity is left to the admin skills. It
kind of require you rewrote and test yourself part of the fencing stack of
Pacemaker. Good luck.
And I'm not speaking about watchdog here, which I just fail to imagine how the
admin could implement it himself.
Just consider how messy it is to deal with "logical fencing" when considering
doing it with pgbouncer.
In short: if you are ready to spend many dev/admin hours to build a safe HA
cluster for your DB and set strict requirements, those are fine.
> About PAF, the product is hard to install and set up . It need a linux
> cluster and a system engineers team to use it.
Indeed, Pacemaker has a steep learning cuve and documentation still requires
some improvement. But HA is not an easy subject. Just look at RHEL or Suse
requirements before their team accept to support your DB cluster (spoiler:
fencing).
Whatever solution you pick, you must **know and document** its limitations and
risks.
