On Thu, Jul 22, 2021 at 8:07 AM
houzj.fnst@fujitsu.com <
houzj.fnst@fujitsu.com> wrote:
>
> Hi,
>
> When I was writing an extension which need to get the median of an array, I
> tried to find if postgres provide some api that can do that. I found all the
> places in postgres invoke qsort() and then get the median. I was thinking can
> we do better by using "quick select" and is it worth it.
> Attach a POC patch about this idea. I did some simple performance tests, I can
> see about 10% performance gain in this test[2].
>
> Thoughts ?
> 1.
> entry_dealloc
> ...
> /* Record the (approximate) median usage */
> if (i > 0)
> pgss->cur_median_usage = entries[i / 2]->counters.usage;
It might be useful to be more precise here, but it seems it would be slower, too?
> -----test create index
> drop index quad_box_tbl_idx;
> CREATE INDEX quad_box_tbl_idx ON quad_box_tbl USING spgist(b);
>
> -------test results
> PATCH:
> Time: 2609.664 ms (00:02.610)
>
> HEAD:
> Time: 2903.765 ms (00:02.944)
That index type is pretty rare, as far as I know. That doesn't seem to be quite enough motivation to change the qsort template. If the goal was to improve the speed of "create spgist index", would this still be the best approach? Also, there are other things under consideration that would add complexity to the qsort template [1], and this would add even more.
