> There's a great deal about this in the list archives (probably more in
> pgsql-hackers than in -performance). Most of the current interest has
> to do with building in-memory bitmaps on the fly, as a way of decoupling
> index and heap scan processing. Which is not quite what you're talking
> about but should be pretty effective for low-cardinality cases. In
> particular it'd allow AND and OR combination of multiple indexes, which
> we do poorly or not at all at the moment.
Is this called a star join ?
It would also allow to access the data pages in a more sequential order
if the rows are not required to be retrieved in index order, which would
potentially be a large speedup for index scans concerning more than the
usual very small percentage of rows in a table : if several rows to be
retrieved are on the same page, it would visit this page only once.
