|From:||Kouhei Kaigai <kaigai(at)ak(dot)jp(dot)nec(dot)com>|
|Cc:||Etsuro Fujita <fujita(dot)etsuro(at)lab(dot)ntt(dot)co(dot)jp>|
|Subject:||PassDownLimitBound for ForeignScan/CustomScan|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
The attached patch adds an optional callback to support special optimization
if ForeignScan/CustomScan are located under the Limit node in plan-tree.
Our sort node wisely switches the behavior when we can preliminary know
exact number of rows to be produced, because all the Sort node has to
return is the top-k rows when it is located under the Limit node.
It is much lightweight workloads than sorting of entire input rows when
nrows is not small.
In my case, this information is very useful because GPU can complete its
sorting operations mostly on L1-grade memory if we can preliminary know
the top-k value is enough small and fits to size of the fast memory.
Probably, it is also valuable for Fujita-san's case because this information
allows to attach "LIMIT k" clause on the remote query of postgres_fdw.
It will reduce amount of the network traffic and remote CPU consumption
once we got support of sort pushdown.
One thing we need to pay attention is cost estimation on the planner stage.
In the existing code, only create_ordered_paths() and create_merge_append_path()
considers the limit clause for cost estimation of sorting. They use the
'limit_tuples' of PlannerInfo; we can reference the structure when extension
adds ForeignPath/CustomPath, so I think we don't need a special enhancement
on the planner stage.
NEC Business Creation Division / PG-Strom Project
KaiGai Kohei <kaigai(at)ak(dot)jp(dot)nec(dot)com>
|Next Message||Andres Freund||2016-08-29 02:00:26||Re: VACUUM's ancillary tasks|
|Previous Message||Vik Fearing||2016-08-29 01:34:36||Sample configuration files|