On 2016-01-20 15:11:06 +0800, Craig Ringer wrote:
> Unfortunately it's not particularly simple and nobody seems to have time to
> implement it.
FWIW, I don't think it's *that* hard.
> As Álvaro pointed out, sometimes you have to do the work if
> you want the change to happen. Or find someone with the existing skills and
> convince them to want to do it, but most of those people are already very,
> very busy.
>
> As part of the failover slots work Simon noted that:
>
> "To prevent needed rows from being removed we need we would need to enhance
> hot_standby_feedback so it sends both xmin and catalog_xmin to the master."
>
> ... which means a protocol change in the walsender protocol. So you're
> looking at that plus the other comment given above, that
Not a huge problem though.
> "We need to be able to correctly and quickly identify the timeline a LSN
> belongs to"
>
> .... which is new internal infrastructure and new state in the replica that
> must be maintained. Probably persistently.
I think it just needs a more efficient lookup structure over the
existing tliOfPointInHistory(), the data is all there.
> (On a side note, failover slots probably won't be usable from a standby
> either. They have to write to WAL and you can't write to WAL from a
> standby. So if slot support is ever added on a standby it'll probably be
> ephemeral slots only.)
ephemeral slots are a different thing. Anyway, at this point failover
slots aren't really proposed / have an agreed upon design yet, so it's a
bit hard to take them into account.
Greetings,
Andres Freund
