> I think that's OK, because it's still protected by the WAL CRC, and
> there's no expectation that the checksum is correct in shared buffers,
> and the correct checksum should be set on the next checkpoint. Just an
> observation.
We'd need to document that emphatically. Otherwise folks running on ZFS
and/or FusionIO with atomic writes (and, in the future, BTRFS) will
assume that they can turn "full_page_writes" off and checksums on, and
clearly that won't work with the current code. I think that's an
acceptable limitation, I just think we need to document it carefully,
and maybe throw a warning if people start up in that configuration.
> Perhaps we don't allow this to be turned per page, but rather per
> cluster, and per-cluster would require the entire cluster to be
> rewritten.
We dicussed this last year, and options which require a total rewrite of
the database in order to turn on the option were rejected as impractical
for users.
People did say it was desirable to have a manual option which says
"rewrite this entire table with checksums". However, that's not
required for the initial patch. For that matter, it will become
desirable to turn on checksums only for specific tables by they user
(again, future feature).
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com