Re: BUG #17245: Index corruption involving deduplicated entries

On Sun, Oct 24, 2021 at 6:33 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> Anything is possible, but this doesn't look like logical index
> corruption. If we assume that it was something like a faulty
> deduplication pass, then we have to explain how the spurious TID ended
> up in that particular leaf page to begin with. We also have to explain
> why it is a valid TID though totally distinct TID, which is unlikely
> to occur by chance.

I discussed the robustness of some of the posting list split code
privately with Heikki earlier today. Heikki suggested that I more
consistently invalidate the insert state that is used across
nbtinsert.c. Although I don't have any reason to believe that that
makes a difference, I still agree with his conclusions.

Attached patch does this. It also makes the special posting split
related call to _bt_binsrch_insert() at the end of _bt_findinsertloc()
more robust. There is a similar defensive elog() within
_bt_swap_posting(), but that only kicks in when the posting offset is
totally out of bounds for the tuple. _bt_binsrch_insert() could still
give answers for some other reason, at least in theory, and so it
seems like a good idea to be extra careful, and throw an error at the
earliest opportunity.

--
Peter Geoghegan