Re: COUNT(*) and index-only scans

Thom Brown <thom(at)linux(dot)com> writes:
> So is there a chance of getting bitmap index-only scans?

Don't hold your breath. It seems like a huge increment of complexity
for probably very marginal gains. The point of a bitmap scan (as
opposed to regular indexscan) is to reduce heap accesses by combining
visits to the same page; but it's not clear how useful that is if
you're not making heap accesses at all.

Robert's sketch of how this could work, full of don't-know-how-to-do-this
as it was, still managed to omit a whole lot of reasons why it wouldn't
work. Notably the fact that the index AM API for bitmap scans is to
return a bitmap, not index-tuple data; and trying to do the latter would
break a lot of the performance advantages that exist now for bitmap
scans.

regards, tom lane