On 30 June 2018 at 00:22, Alexander Kuzmenkov
<a.kuzmenkov@postgrespro.ru> wrote:
> On 06/29/2018 03:25 AM, David Rowley wrote:
>> I've attached a patch that uses SearchSysCacheAttName to speed up
>> these translations in the planner.
>
> Good idea. On my desktop, this gives 270 tps dropped vs 610 tps plain (for
> updates). If you combine it with persistent inner loop index, it's probably
> going to be even faster, because it will only require one catalog access for
> each index shift. Now it looks like it goes to catalog for every column
> after the dropped one.
Thanks for testing.
> What about convert_tuples_by_name_map, do you plan to switch it to catalog
> lookups as well?
Syscache? I didn't really see an obvious way to get the relids down to
the function. e.g the call through ExecEvalConvertRowtype() ->
convert_tuples_by_name() does not have a Relation to work with, just a
TupleDesc.
I think further work might be diminishing returns. I think your idea
to reduce the loops in test 6 from 2000 down to 1001 should be worth
it. I'll try the idea out next week.
--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services