Re: Detecting DB corruption

Am 01.11.2012 06:47, schrieb Craig Ringer:

On 11/01/2012 01:10 PM, Scott Ribe wrote:

On Oct 31, 2012, at 8:50 PM, Craig Ringer wrote:

Seriously, if you're facing DB corruption then something is already
horribly wrong with your setup.

"Horribly" is not strong enough a word IMHO when we're discussing double primary key values... except if Raj is not using sequences to generate them. Although on the other hand, in that case, it's in turn an even more horrible setup; questionable DB design on unreliable hardware.

Raj, would you mind pasting your schema somewhere, at least of the tables you experienced the corruption?

True, but. In a past life, complaints from the db (it was a db that stored a checksum with every block) were the very first symptom when something went horribly wrong with the hardware. (Partial short between wires of an internal SCSI cable; eventually we determined that about every 1MB, 1 bit would get flipped between the controller & disk.)

So, if there were an official db verifier tool for PG, I for one would have it run periodically.

If there were a way to reliably detect corruption, so would I. As things
stand there are no block checksums, so if a bit gets flipped in some
random `text` field you're never going to know, corruption-checker or
no. Some forms of random corruption - like bad blocks on disks causing

I think checksums are currently being worked on and are to be expected for 9.3. Might be interesting to scan -hackers for that once more...

I/O errors, zeroed blocks, truncated files, etc - will become apparent
with general checking, but others won't be detectable unless you know
what the expected vs actual data is.

If page checksumming or any other reliable method of detecting possible
incipient corruption were available I'd quite likely want to use it for
much the same reason you outlined. For that matter, if there were a
general "sanity check my tables and indexes" tool I'd probably use that
too. However, no such tool exists - and in a good setup, none should be
needed. I'd want to use one anyway purely out of paranoia.

--
Craig Ringer

On a side note, Raj, you might want to read the descriptions of MVCC and WAL once more, then re-think about your idea of updating all rows and rolling back the transaction. That would potentially produce the effect you're looking for with InnoDB or Oracle, but not with PG.

Cheers,

--
Gunnar "Nick" Bluth
RHCE/SCLA

Mobil   +49 172 8853339
Email: gunnar.bluth@pro-open.de
__________________________________________________________________________
In 1984 mainstream users were choosing VMS over UNIX.  Ten years later
they are choosing Windows over UNIX.  What part of that message aren't you
getting? - Tom Payne