Re: recovering from "found xmin ... from before relfrozenxid ..."

Hi,

On 2020-07-14 13:20:25 -0400, Alvaro Herrera wrote:
> On 2020-Jul-13, Andres Freund wrote:
> 
> > Hi,
> > 
> > On 2020-07-13 17:12:18 -0400, Robert Haas wrote:
> > > 1. There's nothing to identify the tuple that has the problem, and no
> > > way to know how many more of them there might be. Back-patching
> > > b61d161c146328ae6ba9ed937862d66e5c8b035a would help with the first
> > > part of this.
> > 
> > Not fully, I'm afraid. Afaict it doesn't currently tell you the item
> > pointer offset, just the block numer, right? We probably should extend
> > it to also include the offset...
> 
> Just having the block number is already a tremendous step forward; with
> that you can ask the customer to set a pageinspect dump of tuple
> headers, and then the problem is obvious.  Now if you want to add block
> number to that, by all means do so.

offset number I assume?


> One useful thing to do is to mark a tuple frozen unconditionally if it's
> marked hinted XMIN_COMMITTED; no need to consult pg_clog in that case.
> The attached (for 9.6) does that; IIRC it would have helped in a couple
> of cases.

I think it might also have hidden corruption in at least one case where
we subsequently fixed a bug (and helped detect at least one unfixed
bug). That should only be possible if either required clog has been
removed, or if relfrozenxid/datfrozenxid are corrupt, right?

Greetings,

Andres Freund