Re: WIP: Failover Slots

On 23 January 2016 at 00:51, Andres Freund <andres(at)anarazel(dot)de> wrote:

> Not propagating them through the WAL also has the rather large advantage
> of not barring the way to using such slots on standbys.
>

Yeah. So you could have a read-replica that has a slot and it has child
nodes you can fail over to, but you don't have to have the slot on the
master.

I don't personally find that to be a particularly compelling thing that
says "we must have this" ... but maybe I'm not seeing the full
significance/advantages.

> I think it's technically quite possible to maintain the required
> resources on multiple nodes. The question is how would you configure on
> which nodes the resources need to be maintained? I can't come up with a
> satisfying scheme...
>

That's part of it. Also the mechanism by which we actually replicate them -
protocol additions for the walsender protocol, how to reliably send
something that doesn't have an LSN, etc. It might be fairly simple, I
haven't thought about it deeply, but I'd rather not go there until the
basics are in place.

BTW, I'm keeping a working tree at
https://github.com/2ndQuadrant/postgres/tree/dev/failover-slots . Subject
to rebasing, history not clean. It has a test script in it that'll go away
before patch posting.

Current state needs work to ensure that on-disk and in-memory
representations are kept in sync, but is getting there.

--
Craig Ringer http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services