Re: [HACKERS] path toward faster partition pruning

On 12 December 2017 at 22:13, Amit Langote
<Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
> Attached updated patches.

Hi Amit,

I'm sorry to say this is another micro review per code I'm stumbling
over when looking at the run-time partition pruning stuff.

1. In get_partitions_from_clauses_recurse(), since you're assigning
the result to the first input, the following should use
bms_add_members and not bms_union. The logical end result is the same,
but using bms_union means a wasted palloc and a small memory leak
within the memory context.

/*
* Partition sets obtained from mutually-disjunctive clauses are
* combined using set union.
*/
or_partset = bms_union(or_partset, arg_partset);

2. Also in get_partitions_from_clauses_recurse(), it might also be
worth putting in a bms_free(or_partset) after:

/*
* Partition sets obtained from mutually-conjunctive clauses are
* combined using set intersection.
*/
result = bms_intersect(result, or_partset);

Also, instead of using bms_intersect here, would it be better to do:

result = bms_del_members(result, or_partset); ?

That way you don't do a bms_copy and leak member for each OR branch
since bms_intersect also does a bms_copy()

The resulting set could end up with a few more trailing 0 words than
what you have now, but it to be a better idea not allocate a new set
each time.

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