Hi,
On 2022-11-15 13:54:24 -0800, Peter Geoghegan wrote:
> On Tue, Nov 15, 2022 at 12:29 PM Andres Freund <andres@anarazel.de> wrote:
> > ... I strongly dislike latestCommittedXid. That seems at least as misleading
> > as latestRemovedXid and has the danger of confusion with latestCompletedXid
> > as you mention.
>
> > How about latestAffectedXid?
>
> I get why you don't care for latestCommittedXid, of course, but the
> name does have some advantages. Namely:
>
> 1. Most conflicts come from PRUNE records (less often index deletion
> records) where the XID is some heap tuple's xmax, a
> committed-to-everybody XID on the primary (at the point of the
> original execution of the prune). It makes sense to emphasize the idea
> that snapshots running on a replica need to agree that this XID is
> definitely committed -- we need to kill any snapshots that don't
> definitely agree that this one particular XID is committed by now.
I don't agree that it makes sense there - to me it sounds like the record is
just carrying the globally latest committed xid rather than something just
describing the record.
I also just don't think "agreeing that a particular XID is committed" is a
good description of latestRemovedXID, there's just too many ways that
"agreeing xid has committed" can be understood. To me it's not obvious that
it's about mvcc snapshots.
If we want to focus on the mvcc affects we could just go for something like
snapshotConflictHorizon or such.
> Perhaps something like "mustBeCommittedCutoff" would work better? What
> do you think of that? The emphasis on how things need to work on the
> REDO side seems useful.
I don't think "committed" should be part of the name.
Greetings,
Andres Freund
