Re: Research/Implementation of Nested Loop Join optimization

Gregory Stark <stark(at)enterprisedb(dot)com> writes:
> "Manoel Henrique" <mhenriquesgbd(at)gmail(dot)com> writes:
>> Yes, I'm relying on the assumption that backwards scan has the same cost as
>> forward scan, why shouldn't it?

> Because hard drives only spin one direction

Good joke, but to be serious: we expect that forward scans will result
in the kernel doing read-ahead, which will allow overlapping of
CPU work to process one page with the I/O to bring in the next page.
A backwards scan will get no such overlapping and thus be up to 2X
slower, unless the kernel is smart enough to do read-ahead for
descending-order read requests. Which seems not too probable. A fairly
typical kernel behavior is that read-ahead is triggered by successive
read() requests without any intervening seek(), and this is impossible
for a backward scan.

(Yes, we do optimize out the seek calls in a forward scan. IIRC it's
done in fd.c.)

regards, tom lane