Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Mar 4, 2014 at 2:39 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
>> Your earlier claim that the dump is inconsistent just isn't accurate.
>> We now have MVCC catalogs, so any dump is going to see a perfectly
>> consistent set of data plus DDL. OK the catalogs may change AFTER the
>> snapshot was taken for the dump, but then so can the data change -
>> that's just MVCC.
> Unfortunately, this isn't correct. The MVCC snapshots taken for
> catalog scans are "instantaneous"; that is, we take a new, current
> snapshot for each catalog scan. If all of the ruleutils.c stuff were
> using the transaction snapshot rather than instantaneous snapshots,
> this would be right. But as has been previously discussed, that's not
> the case.
Yeah. And that's *necessary* for catalog lookups in a normally
functioning backend, because we have to see latest data (eg, it wouldn't
do for a backend to fail to enforce a just-added CHECK constraint because
it was committed after the backend's transaction started).
However, it seems possible that we could have a mode in which a read-only
session did all its catalog fetches according to the transaction snapshot.
That would get us to a situation where the backend-internal lookups that
ruleutils relies on would give the same answers as queries done by
pg_dump. Robert's work on getting rid of SnapshotNow has probably moved
that much closer than it was before, but it's still not exactly a trivial
patch.
Meanwhile, Andres claimed upthread that none of the currently-proposed
reduced-lock ALTER commands affect data that pg_dump is using ruleutils
to fetch. If that's the case, then maybe this is a problem that we can
punt till later. I've not gone through the list to verify it though.
regards, tom lane