Just noticed that PageIndexTupleDelete does not check that
pd_special != MAXALIGN(pd_special)
whereas PageIndexMultiDelete() does this. Both routines state that
"paranoia seems justified", so this omission looks like an error.
Looking a little deeper at this...
We only call PageIndexTupleDelete() in cases where
PageIndexMultiDelete() has only a single element. We also call it
directly (in btree case) when recovering page splits and when updating
parent pages as part of page removal.
Having a special function that exists only for use in rare occasions
seems like a great recipe for sporadic corruption, if we are taking the
"paranoia seems justified" approach.
I would be inclined to dispose of PageIndexTupleDelete altogether. We
may yet be able to optimise PageIndexMultiDelete for low values of
nitems, if that is important.
-- Simon Riggs www.2ndQuadrant.com