Re: index-only scans

On Tue, Oct 11, 2011 at 2:36 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:

> Have you given any thought to what would be required to support
> index-only scans on non-btree indexes - particularly, GIST? ISTM we
> might have had a thought that some GIST opclasses but not others would
> be able to regurgitate tuples, or maybe even that it might vary from
> index tuple to index tuple. But that discussion was a long time ago,
> and my memory is fuzzy.

In some GiST opclasses original tuple can't be restored from index tuple.
For example, polygon can't be restored from it's MBR. In some opclasses, for
example box_ops and point_ops, original tuple can be easily restore from
index tuple.
I can't remeber any implementation where this possibility can vary from
index tuple to index tuple. GiST opclass for ts_vector have different
representation of leaf index tuple depending on it's length. But, at most,
it store array of hashes of words, so it's lossy anyway.
I think GiST interface could be extended with optional function which
restores original tuple. But there is some additional difficulty, when some
opclasses of composite index provide this function, but others - not.

------
With best regards,
Alexander Korotkov.