Re: PANIC: corrupted item lengths

Simon Riggs wrote:

> I would much prefer:
>
> * VACUUMs seeing these errors would perform as with zero_damaged_pages.
> * other backends seeing those errors should just ERROR out.
>
> We can do this by having a new function: boolean PageIsValid() which
> performs the sanity checks. This can then be executed by
> heap_page_prune() prior to entering the critical section. That will then
> be called correctly by both VACUUM and other code. VACUUM can then
> optionally zero out the block, as is done with PageHeaderIsValid().

I tend to hate automatic zeroing of pages because there's no way to get
the contents later for forensics. I would support your proposal if we
had a way of saving the block elsewhere before zeroing it (say create a
directory corrupted+zeroed similar to lost+found in the database dir and
save it annotated with the OID of the table and the block number).

The main problem I see with this is that if you don't immediately act to
examine the data, some of the pg_clog files that you need to be able to
read these files may be gone.

--
Alvaro Herrera http://www.CommandPrompt.com/
The PostgreSQL Company - Command Prompt, Inc.