Re: Performance regression with PostgreSQL 11 and partitioning

On 5 June 2018 at 16:44, Ashutosh Bapat <ashutosh(dot)bapat(at)enterprisedb(dot)com> wrote:
> I think the idea is brilliant. I do not have objections for trying
> something in that direction. I am suggesting that we take this a bit
> forward and try to eliminate append_rel_list altogether.

I was trying to be realistic for something we can do to fix v11. It's
probably better to minimise the risky surgery on this code while in
beta. What I proposed was intended to fix a performance regression new
in v11. I'm not sure what you're proposing has the same intentions.

Probably, if you do want an efficient way to store the children
belonging to a parent we could just have another array of Bitmapsets
which would just serve as a set of indexes into the array I've added.

I'd prefer to get a committers thoughts on this before doing any further work.

I've attached a cleaned up patch from the last one. This just adds
some sanity checks to make sure we get an ERROR if we do ever see two
AppendRelInfos with the same child relation id. I've also made it so
the append_rel_array is only allocated when there are append rels.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services