[ getting back to this documentation issue finally ]
Teodor Sigaev <teodor@sigaev.ru> writes:
> I disagree with last affirmation: inner pages of index contains fair union of
> keys and enough helpful to select. Mailware ( http://www.pgsql.ru/db/mw )
> sucsessfully use combined GiST index (date, tsvector) for searching.
> GiST's split algorithm is good for unique leading keys, not so bad for small
> number of non-unique values and bad for all equals leading key. But "bad" means
> that itsn't optimal as picksplit for other keys may be. If there is several keys
> which can be moved on left or right page without changing union of first key for
> each page then GiST try put its on page (left or right) with smallest penalty
> calculated by other keys. This algorithm is very similar to defining page to put
> tuple with normal processing (without page split).
> With unique leading key GiST's split is fully similar to BTree - it looks only
> at leading key, but gistchoose isn't. Gistchoose (gistutil.c:622) chooses child
> with smallest penalty and it looks to other keys if several leading keys has the
> same penalty. In a GiST tree different keys may have the same penalty value with
> new key.
OK, how about this text then?
A multicolumn GiST index can only be used when there is a query condition
on its leading column. Conditions on additional columns restrict the
entries returned by the index, but the condition on the first column is the
most important one for determining how much of the index needs to be
scanned. A GiST index will be relatively ineffective if its first column
has only a few distinct values, even if there are many distinct values in
additional columns.
regards, tom lane