Tony Nelson wrote:
> This morning my server (ubuntu 10.04 LTS, pg 8.4 from apt)
> experienced a disk corruption which caused errors like this:
>
> Feb 18 08:50:07 ihdb1 postgres[13317]: [2-1]
> user=postgres,db=instihire ERROR: invalid page header in block 10
> of relation base/16384/8082305
Do you know why? If not, that machine is not to be trusted.
> I am ok, because I have a dump from last night, and wal since then.
That dump was taken according to PITR instructions, not from pg_dump
or pg_dumpall, I hope.
> As a test, I tried doing a dump like this from the broken server:
>
> pg_dump -Fc mydb > busted.dump
>
> I restored this dump on a test server, and all of the data looks
> ok, and my application is able to run the operation that was
> failing in production just fine.
>
> My question is simple, is my dump good?
Maybe; maybe not. I would try to restore from that dump you
mentioned, run pg_dump on that, and compare the two files which came
out of pg_dump. Or just use the dump if you can restore it and
you're sure it's up to date.
> Is it possible that pg_dump was able to correctly save the data
> that "select .. " and "update.. " couldn't.
pg_dump essentially just does a SELECT from each table. If the
corrupted relation was an index, you might have dodged the problem
mentioned above, but if a page in that index got corrupted, why
assume damage was localized to that one place?
You really have two problems to address here: how best to recover
your data, and how to prevent a recurrence.
What version of PostgreSQL is this exactly? (It's best to SELECT
version(); and copy/paste the results.)
-Kevin