On Mon, Mar 2, 2009 at 3:02 PM, Jeff Davis <pgsql@j-davis.com> wrote:
> On Mon, 2009-03-02 at 21:14 +0200, Heikki Linnakangas wrote:
>> If I'm reading the code correctly, item pointers of all matching heap
>> tuples are first collected into a TIDBitmap, and then amgetnext returns
>> tuples from that one by one. If the bitmap becomes lossy, an error is
>> thrown. gingetbitmap is a dummy implementation: it creates a new
>> TIDBitmap and inserts all the tuples from the other TIDBitmap into it
>> one by one, and then returns the new TIDBitmap.
>
> Do you think that might be the cause of the extra startup overhead that
> Robert Haas observed for bitmap scans?
I don't think this is the same thing. My point was that an index scan
wins big time over a bitmap index scan when the index scan doesn't
need to be run to completion - that is, when the query is a semi-join
or an anti-join, or when using LIMIT without ORDER BY.
This is true with or without Teodor's patch, and is the reason why I'm
not sure that removing index scan support is such a great idea.
...Robert
