CCing to the list...
On Fri, Aug 12, 2016 at 4:10 PM, Ioana Danes <ioanadanes@gmail.com> wrote:
>> given 318220 and 318216 are just a bit away ( 4db08/4db0c ), and it
>> repeats sporadically, have you ruled out ( by having page checksums or
>> other mechanism ) a potential disk read/write error ?
>>
>>
>> > Also the index is correct on db3 as the record in case (with drawid =
>> > 318216) is retrieved if I filter by drawid = 318220
>>
>> Specially if this happens, you may have some slightly bad disks/ram/
>> leading to this kind of problems.
>>
>
> Could be. I also had some issues with an rsync between db3 and drdb a week
> ago that did not complete for bigger files (> 200MB) and gave me some
> corruption messages. Then the system was revbooted and everything seemed
> fine but apparently it is not.
> I am planning to drop & create the table from a good backup and if that does
> not fix the issue then I will rebuild the server.
I would check whatever logs you can ( syslog or eventlog, smart log,
etc.. ) hunting for disk errors ( sometimes they are reported ). This
kind of problems, with programs as tested as postgres and rsync, tend
to indicate controller/RAM/disk going bad ( in your case it could be
caused by a single bit getting flipped in a sector for the data
portion of the table, and not being propagated either because it
happened after your sync of drdb or because it was synced from the WAL
and not the table, or because it was read from the disk cache ).
Francisco Olarte.
