Tom Lane wrote:
>
> Vadim Mikheev <vadim@krs.ru> writes:
> > In UPDATE backend inserts index tuple for new version of heap tuple
> > and adjusts all index scans affected by this insertion.
> > Something is wrong in nbtscan.c:_bt_adjscans()...
>
> Could be; maybe there's one boundary case that fails to advance the
> index scan? I hope there's someone who's looked at nbtree recently
> who can take the time to debug this.
I'll try to look there...
> Another thing that struck me while looking at the update code is that
> an update deletes the old tuple value, then inserts the new value,
> but it doesn't bother to delete any old index entries pointing at the
> old tuple. ISTM that after a while, there are going to be a lot of old
> index entries pointing at dead tuples ... or, perhaps, at *some other*
> live tuple, if the space the dead tuple occupied has been reused for
> something else. This certainly seems to present a risk of returning
> the wrong tuple. I looked through the code to find out how such an
> error is prevented, and didn't find anything. But maybe I just don't
> know where to look.
Vacuum deletes index tuples before deleting heap ones...
Vadim