On 2015-07-16 12:40:07 -0400, Robert Haas wrote:
> On Wed, Jul 15, 2015 at 12:51 PM, Florent Guiliani <florent@guiliani.fr> wrote:
> > During slot creation, the snapshot building and exporting code seems
> > highly coupled with the logical decoding stuff. It doesn't seems much
> > reusable to retrieve the snapshot's LSN outside of logical decoding.
>
> I don't think "the snapshot's LSN" has a well-defined meaning in
> general. The obvious meaning would be "the LSN such that all commits
> prior to that LSN are visible and all later commits are invisible",
> but such an LSN need not exist. Suppose A writes a commit record at
> LSN 0/10000, and then B writes a commit record at 0/10100, and then B
> calls ProcArrayEndTransaction(). At this point, B is visible and A is
> not visible, even though A's commit record precedes that of B.
Well, in combination with logical decoding it kinda has one: It should
allow you to take a dump of the database with a certain snapshot and
replay all transactions with a commit lsn bigger than the "snapshot's
lsn" and end up with a continually consistent database.
Visibility for HS actually works precisely in commit LSN order, even if
that is possibly different than on the primary...