Re: Problems with plan estimates in postgres_fdw

(2018/12/17 22:09), Etsuro Fujita wrote:
> Here is a set of WIP patches for pushing down ORDER BY LIMIT to the remote:

> For some regression test cases with ORDER BY and/or LIMIT, I noticed
> that these patches still cannot push down those clause to the remote. I
> guess it would be needed to tweak the cost/size estimation added by
> these patches, but I didn't look at that in detail yet. Maybe I'm
> missing something, though.

I looked into that. In the previous patch, I estimated costs for
performing grouping/aggregation with ORDER BY remotely, using the same
heuristic as scan/join cases if appropriate (i.e., I assumed that a
remote sort for grouping/aggregation also costs 20% extra), but that
seems too large for grouping/aggregation. So I reduced that to 5%
extra. The result showed that some test cases can be pushed down, but
some still cannot. I think we could push down the ORDER BY in all cases
by reducing the extra cost to a much smaller value, but I'm not sure
what value is reasonable.

Attached is an updated version of the patch. Other changes:

* Modified estimate_path_cost_size() further so that it accounts for
tlist eval cost as well
* Added more comments
* Simplified code a little bit

I'll add this to the upcoming CF.

Best regards,
Etsuro Fujita