Re: [HACKERS] path toward faster partition pruning

David

On 2017/12/21 21:04, David Rowley wrote:
> On 21 December 2017 at 23:38, Amit Langote
> <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> Attached updated set of patches.
>
> Looks like the new not equals code does not properly take into account
> a missing NULL partition.
>
> create table ta (a int not null) partition by list (a);
> create table ta1 partition of ta for values in(1,2);
> create table ta2 partition of ta for values in(3,4);
> explain select * from ta where a <> 1 and a <> 2;
> ERROR: negative bitmapset member not allowed
>
> -- Add null partition
> create table ta_null partition of ta for values in(null);
> explain select * from ta where a <> 1 and a <> 2; -- works now.
> QUERY PLAN
> -------------------------------------------------------------
> Append (cost=0.00..48.25 rows=2525 width=4)
> -> Seq Scan on ta2 (cost=0.00..48.25 rows=2525 width=4)
> Filter: ((a <> 1) AND (a <> 2))
> (3 rows)
>
> This code appears to be at fault:
>
> /*
> * Also, exclude the "null-only" partition, because strict clauses in
> * ne_clauses will not select any rows from it.
> */
> if (count_partition_datums(relation, boundinfo->null_index) == 0)
> excluded_parts = bms_add_member(excluded_parts,
> boundinfo->null_index);

Oops, must check before going to count datums that a null-partition exists
at all. Will post the fixed version shortly, thanks.

Regards,
Amit