Re: ALTER TABLE lock strength reduction patch is unsafe

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Thu, Jun 16, 2011 at 6:54 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> 4. Backend #2 visits the new, about-to-be-committed version of
>> pgbench_accounts' pg_class row just before backend #3 commits.
>> It sees the row as not good and keeps scanning. By the time it
>> reaches the previous version of the row, however, backend #3
>> *has* committed. So that version isn't good according to SnapshotNow
>> either.

> <thinks some more>

> Why isn't this a danger for every pg_class update? For example, it
> would seem that if VACUUM updates relpages/reltuples, it would be
> prone to this same hazard.

VACUUM does that with an in-place, nontransactional update. But yes,
this is a risk for every transactional catalog update.

regards, tom lane