On 14/11/14 14:54, Tom Lane wrote:
> Jeremy Harris <jgh@wizmail.org> writes:
>> On 14/11/14 00:46, Simon Riggs wrote:
>>> Limit (cost=.... rows=20 width=175) (actual time=.... rows=20 loops=1)
>>> -> Sort (cost=.... rows=568733 width=175) (actual time=....
>>> rows=20 loops=1)
>>> Sort Method: top-N heapsort
>
>> Going off on a tangent, when I was playing with a merge-sort
>> implementation I propagated limit information into the sort
>> node, for a significant win.
>
> I'm not entirely following. The top-N heapsort approach already
> makes use of the limit info.
Having gone back to look, you're right. It was Uniq nodes I merged
(the sort handles both bounded-output and dedup).
--
Cheers, Jeremy