Re: Ordered Partitioned Table Scans

Hi,

On 2019/03/28 7:29, David Rowley wrote:
> On Wed, 27 Mar 2019 at 19:48, Amit Langote
> <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> Sorry if this was discussed before, but why does this patch add any new
>> code to partprune.c? AFAICT, there's no functionality changes to the
>> pruning code.
>
> You're right. It probably shouldn't be there. There's a bit of a lack
> of a good home for partition code relating to the planner it seems.

partprune.c is outside the optimizer sub-directory, but exports
planning-related functions like prune_append_rel_partitions(),
make_partition_pruneinfo(), etc.

Similarly, partbound.c can grow bit of planning-related functionality.

>> Regarding
>>
>> +bool
>> +partitions_are_ordered(PlannerInfo *root, RelOptInfo *partrel)
>>
>> I think this could simply be:
>>
>> bool
>> partitions_are_ordered(PartitionBoundInfo *boundinfo)
>>
>> and be defined in partitioning/partbounds.c. If you think any future
>> modifications to this will require access to the partition key info in
>> PartitionScheme, maybe the following is fine:
>>
>> bool
>> partitions_are_ordered(RelOptInfo *partrel)
>
> It does need to know how many partitions the partitioned table has,
> which it gets from partrel->nparts, so yeah, RelOptInfo is probably
> needed. I don't think passing in int nparts is a good solution to
> that. The problem with moving it to partbounds.c is that nothing
> there knows about RelOptInfo currently.

Maybe, this could be a start. Also, there is a patch in nearby thread
which adds additional functionality to partbounds.c to be used by
partitionwise join code in the optimizer [1].

Thanks,
Amit

[1] https://commitfest.postgresql.org/22/1553/