Hi,
On 2022-04-11 13:51:38 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > One way to address it in a way not requiring an API break would be to pass
> > SnapshotAny to heap_fetch and then do an explicit visibility check "ourselves"
> > in heapam_lock_tuple().
>
> I'm not really interested in fixing this without an API break (going
> forward anyway), because as it stands heap_fetch is just an invitation
> to make this same mistake again.
My suggestion was about the back branch situation... But it seems viable going
forward as well, if we we reset tuple->t_data in the !valid case. As you say:
> It should never return a tuple pointer if we don't keep the pin on the
> associated buffer.
Agreed. If tuple->t_data were reset in the !valid case, not just the
!ItemIdIsNormal() case, bug would have been noticed immediately (isolation
tests do fail, I checked).
Another approach is to extend the SatisfiesDirty approach and store the tid of
the next tuple version in addition the xmin/xmax we already store. And have
heap_fetch() always set t_data to NULL if the snapshot check fails.
Greetings,
Andres Freund