On Wed, 31 Oct 2001, Tatsuo Ishii wrote:
> > It may be unthinkable hubris to say this, but ... I am starting to
> > notice that a larger and larger fraction of serious trouble reports
> > ultimately trace to hardware failures, not software bugs. Seems we've
> > done a good job getting data-corruption bugs out of Postgres.
> >
> > Perhaps we should reconsider the notion of keeping CRC checksums on
> > data pages. Not sure what we could do to defend against bad RAM,
> > however.
Maybe not defend against it, but at least you can detect and warn the
user that something is likely to go wrong.
> I have been troubled by a really strange problem. Populating with huge
> data (~7GB) cause random failures, for example a misterious unique
> constaraint violation, count(*) shows incorrect number, pg_temp*
> suddenly disappear (the table in question is a temporary table).
Remember the guy who had to change relnatts by hand to get a table back
on line? It was bad RAM. One may wonder just how big the coincidence was
to get exactly that bit changed... Well, a bad CRC checksum would've
warned him right away.
--
Alvaro Herrera (<alvherre[@]atentus.com>)
"Si quieres ser creativo, aprende el arte de perder el tiempo"
