"Tom Lane" <tgl@sss.pgh.pa.us> writes:
> Remember that the xmin/xmax fields are basically the first thing we can
> check with any degree of strictness when examining a tuple. This means that
> if a page is partially clobbered, but not in a way that sets off the
> invalid-page-header checks, then the odds are very high that the first
> detectable sign of trouble will be references to transaction numbers that
> are far away from what the system is really using.
I'm increasingly thinking that one of the first things I'll suggest putting
into 8.4 is a per-page checksum after all. It was talked about a while back
and people thought it was pointless but I think the number of reports of
hardware and kernel bugs resulting in zeroed and corrupted pages has been
steadily going up. If not in total than as a percentage of the total problems.
--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com