Re: 9.0.4 Data corruption issue

2011/7/17 Ken Caruso <ken@ipl31.net>:

>
>
> On Sat, Jul 16, 2011 at 2:30 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>
>> Ken Caruso <ken@ipl31.net> writes:
>> > Sorry, the actual error reported by CLUSTER is:
>>
>> > gpup=> cluster verbose tablename;
>> > INFO:  clustering "dbname.tablename"
>> > WARNING:  could not write block 12125253 of base/2651908/652397108
>> > DETAIL:  Multiple failures --- write error might be permanent.
>> > ERROR:  could not open file "base/2651908/652397108.1" (target block
>> > 12125253): No such file or directory
>> > CONTEXT:  writing block 12125253 of relation base/2651908/652397108
>>
>> Hmm ... it looks like you've got a dirty buffer in shared memory that
>> corresponds to a block that no longer exists on disk; in fact, the whole
>> table segment it belonged to is gone.  Or maybe the block or file number
>> in the shared buffer header is corrupted somehow.
>>
>> I imagine you're seeing errors like this during each checkpoint attempt?
>
> Hi Tom,
> Thanks for the reply.
> Yes, I tried a pg_start_backup() to force a checkpoint and it failed due to
> similar error.
>
>>
>> I can't think of any very good way to clean that up.  What I'd try here
>> is a forced database shutdown (immediate-mode stop) and see if it starts
>> up cleanly.  It might be that whatever caused this has also corrupted
>> the back WAL and so WAL replay will result in the same or similar error.
>> In that case you'll be forced to do a pg_resetxlog to get the DB to come
>> up again.  If so, a dump and reload and some manual consistency checking
>> would be indicated :-(
>
> Before seeing this message, I restarted Postgres and it was able to get to a
> consistent state at which point I reclustered the db without error and
> everything appears to be fine. Any idea what caused this? Was it something
> to do with the Vacuum Full?

Block number 12125253 is bigger that any block we can find in
base/2651908/652397108.1
Should the table size be in the 100GB range or 2-3 GB range ?
This should help decide: if in the former case, then probably at least
a segment disappear or, in the later, the shared_buffer turn
corrupted.

Ken, you didn't change RELSEG_SIZE, right ? (it needs to be change in
source code before compile it yourself)
In both case a hardware check is welcome I believe.
--
Cédric Villemain               2ndQuadrant
http://2ndQuadrant.fr/     PostgreSQL : Expertise, Formation et Support