Jeff Davis <pgsql(at)j-davis(dot)com> wrote:
> Ok, great. When I read that before I thought that WAL might need
> to be sent for implicit RO transactions. I will read it more
> carefully again.
In looking back over recent posts to see what I might have missed or
misinterpreted, I now see your point. Either of these alternatives
would involve potentially sending something through the WAL on
commit or rollback of some serializable transactions which *did not*
write anything, if they were not *declared* to be READ ONLY. If
that is not currently happening (again, I confess to not having yet
delved into the mysteries of writing WAL records), then we would
need a new WAL record type for writing these.
That said, the logic would not make it at all useful to send
something for *every* such transaction, and I've rather assumed
that we would want some heuristic for setting a minimum interval
between notifications, whether we sent the snapshots themselves or
just flags to indicate it was time to build or validate a candidate
snapshot.
Sorry for misunderstanding the concerns.
-Kevin