On Wed, Dec 05, 2012 at 07:43:08PM -0500, Tom Lane wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
> > On Tue, Dec 4, 2012 at 3:38 PM, Jeff Davis <pgsql@j-davis.com> wrote:
> >> After reading that thread, I still don't understand why it's unsafe to
> >> set HEAP_XMIN_COMMITTED in those conditions. Even if it is, I would
> >> think that a sufficiently narrow case -- such as CTAS outside of a
> >> transaction block -- would be safe, along with some slightly broader
> >> cases (like BEGIN; CREATE TABLE; INSERT/COPY).
>
> > I haven't looked at the committed patch - which seemed a bit
> > precipitous to me given the stage the discussion was at - but I
> > believe the general issue with HEAP_XMIN_COMMITTED is that there might
> > be other snapshots in the same transaction, for example from open
> > cursors.
>
> From memory, the tqual.c code assumes that any tuple with XMIN_COMMITTED
> couldn't possibly be from its own transaction, and thus it doesn't make
> the tests that would be appropriate for a tuple that is from the current
> transaction. Maybe it's all right anyway (i.e. if we should always treat
> such a tuple as good) but I don't recall exactly what's tested in those
> paths.
I don't see semantics preservable by freezing, yet omitting
HEAP_XMIN_COMMITTED. The "HeapTupleHeaderGetCmin(tuple) >= snapshot->curcid"
test is the one at risk. HeapTupleSatisfiesMVCC() does skip that test for
HEAP_XMIN_COMMITTED tuples, but seeing xmin==FrozenTransactionId hampers it
all the more.
What if one of the preconditions for the optimization were the equivalent of
CheckTableNotInUse()? I cannot immediately think of a older-cmin-scan source
not caught thereby. Unmodified HeapTupleSatisfiesMVCC() will then suffice.
Happily, it's not a restriction users will regularly encounter.
Thanks,
nm