The core committee has agreed that this bug
http://archives.postgresql.org/pgsql-hackers/2004-08/msg00639.php
is serious enough that we'd better push out update releases for
all the affected versions. I want to hold off a day or so and see
if a couple of currently-open bug reports can be resolved, but it's
going to happen soon. If anyone has any back-patches that they've
been thinking of getting in, now would be a good time.
BTW, I realized this morning that 7.1.* also has a version of the bug,
because it too writes and flushes an XLOG COMMIT record before it does
anything about marking the transaction complete in pg_log. So it
likewise would have the problem that a checkpoint occurring just after
the COMMIT might not include the pg_log update, leading to possible loss
of the pg_log update in event of crash and replay from that checkpoint.
Core's feeling is that we will not try to backpatch 7.1. The patch used
for the later versions will not apply to 7.1 because it doesn't have
LWLocks, so some nontrivial effort would be needed to develop a fix.
Given the age of that release, and the other serious problems it has
(the transaction-ID-wraparound problem for one), it's time to officially
state that it is broken and unsupportable.
Comments anyone? Backpatches for other bugs?
regards, tom lane