Re: [HACKERS] path toward faster partition pruning

Hi David.

On 2018/03/31 0:55, David Rowley wrote:
> On 30 March 2018 at 18:38, Amit Langote <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> Please find attached the updated patches.
>
> There's a bit of a strange case with v45 around prepared statements.
> I've not debugged this yet, but in case you get there first, here's
> the case:
>
> create table listp (a int, b int) partition by list (a);
> create table listp_1 partition of listp for values in(1) partition by list (b);
> create table listp_1_1 partition of listp_1 for values in(1);
> create table listp_2 partition of listp for values in(2) partition by list (b);
> create table listp_2_1 partition of listp_2 for values in(2);
>
> explain select * from listp where b in(1,2) and 2<>b and 0<>b; -- this
> one looks fine.
> QUERY PLAN
> ----------------------------------------------------------------------------
> Append (cost=0.00..49.66 rows=22 width=8)
> -> Seq Scan on listp_1_1 (cost=0.00..49.55 rows=22 width=8)
> Filter: ((b = ANY ('{1,2}'::integer[])) AND (2 <> b) AND (0 <> b))
> (3 rows)
>
> prepare q1 (int,int,int,int) as select * from listp where b in($1,$2)
> and $3 <> b and $4 <> b;
> execute q1 (1,2,3,4);
> execute q1 (1,2,3,4);
> execute q1 (1,2,3,4);
> execute q1 (1,2,3,4);
> execute q1 (1,2,3,4);
> explain (analyze, costs off, summary off, timing off) execute q1 (1,2,2,0);
> QUERY PLAN
> --------------------------------
> Result (actual rows=0 loops=1)
> One-Time Filter: false
> (2 rows)
>
> My best guess is that something ate the bits out of a Bitmapset of the
> matching partitions somewhere.

Hmm. It is the newly added inversion step that's causing this. When
creating a generic plan (that is when the planning happens via
BuildCachedPlan called with boundParams set to NULL), the presence of
Params will cause an inversion step's source step to produce
scan-all-partitions sort of result, which the inversion step dutifully
inverts to a scan-no-partitions result.

I have tried to attack that problem by handling the
no-values-to-prune-with case using a side-channel to propagate the
scan-all-partitions result through possibly multiple steps. That is, a
base pruning step will set datum_offsets in a PruneStepResult only if
pruning is carried out by actually comparing values with the partition
bounds. If no values were provided (like in the generic plan case), it
will set a scan_all_nonnull flag instead and return without setting
datum_offsets. Combine steps perform their combining duty only if
datum_offset contains a valid value, that is, if scan_all_nonnulls is not set.

Attached updated version of the patches.

Thanks,
Amit