Robert Haas <robertmhaas@gmail.com> writes:
> Department of second thoughts: I think I see a problem.
Um, yeah, so that doesn't really work any better than my idea.
On further reflection, there's a problem at a higher level than this
anyway. Even if we can get a single SnapshotNow scan to produce
guaranteed-self-consistent results, that doesn't ensure consistency
between the results of scans occurring serially. An example here is
ALTER COLUMN DROP DEFAULT, which is currently imagined to impact only
writers. However, suppose that a concurrent relcache load fetches the
pg_attribute row, notes that it has atthasdef = true, and then the ALTER
commits before we start to scan pg_attrdef. The consistency checks in
AttrDefaultFetch() will complain about a missing pg_attrdef entry, and
rightly so. We could lobotomize those checks, but it doesn't feel right
to do so; and anyway there may be other cases that are harder to kluge up.
So really we need consistency across *at least* one entire relcache load
cycle. We could maybe arrange to take an MVCC snap (or some lighter
weight version of that) at the start, and use that for all the resulting
scans, but I think that would be notationally messy. It's not clear
that it'd solve everything anyhow. There are parts of a relcache entry
that we fetch only on-demand, so they are typically loaded later than
the core items, and probably couldn't use the same snapshot. Worse,
there are lots of places where we assume that use of catcache entries or
direct examination of the catalogs will yield results consistent with
the relcache.
I suspect these latter problems will impact Simon's idea as well.
regards, tom lane
