🚀 go-pugleaf

cis115@ws001a.cis.squared.com (Tony Gast) writes:

> The vast majority of the data we access, and I believe this is a
> definite industry trend, is in raw partitions being accessed by
> Sybase (Oracle, Informix, what have you).  The OS does not cache this
> info.

Um, so what?  There's still metadata, and the issue of ordering of
writes so that the metadata and the data are consistent.  You can do
it in the app, or you can do it in the OS; but if it isn't done, (and
lying about whether a write is committed to stable storage is a
sure-fire prescription for that) you can and will end up with
inconsistent disk data structures in the event of a crash.

(I happen to think that raw DB partitions are evil, because they
require a whole 'nother suite of tools for their maintenance, but
let's not get onto side issues; performance requirements often dictate
raw partitions and I can live with them.)

BTW, all the business of just how wonderful your UPS safeguards are is
beside the point; what if the failure is in the cache RAM parts
themselves, or their controller, or the raid box's backplane, or the
scsi disk controller that ties to the disks proper?  Power outage
isn't the only failure mode on the planet.

regards,
Ross
--
Ross Alexander, ve6pdq  --  (403) 675 6311  --  rwa@cs.athabascau.ca