Re: [HACKERS] path toward faster partition pruning

Hi David.

On 2018/02/01 8:57, David Rowley wrote:
> On 31 January 2018 at 21:03, Amit Langote <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> Update patch set attached. Thanks again.
>
> (My apologies for being slow to respond here. I've been on leave this
> week and I'm off again next week. I now have a little time to reply)

No worries.

> Thanks for incorporating my changes into the patchset. A while ago I
> was rebasing the run-time pruning patch on top of this but ran into a
> few problems which are all results of my changes.
>
> 1. remove_redundant_clauses collapses the PartClause list into the
> most restrictive set of clauses. This disallows multiple evaluations
> of the PartitionClauseInfo during runtime pruning. I've written a
> proposed fix for this and attached it.

I've incorporated it in the latest patch I posted today.

> 2. PartitionClauseInfo->keyclauses is a list of PartClause which is
> not a node type. This will cause _copyPartitionClauseInfo() to fail.
>
> I'm still not quite sure the best way to fix #2 since PartClause
> contains a FmgrInfo. I do have a local fix which moves PartClause to
> primnodes.h and makes it a proper node type. I also added a copy
> function which does not copy any of the cache fields in PartClause. It
> just sets valid_cache to false. I didn't particularly think this was
> the correct fix. I just couldn't think of how exactly this should be
> done at the time.
>
> The attached patch also adds the missing nodetag from
> PartitionClauseInfo and also fixes up other code so as we don't memset
> the node memory to zero, as that would destroy the tag. I ended up
> just having extract_partition_key_clauses do the makeNode call. This
> also resulted in populate_partition_clauses being renamed to
> generate_partition_clauses

I started wondering if it's not such a good idea to make
PartitionClauseInfo a Node at all? I went back to your earlier message
[1] where you said that it's put into the Append node for run-time pruning
to use, but it doesn't sound nice that we'd be putting into the plan
something that's looks more like scratchpad for the partition.c code. I
think we should try to keep PartitionClauseInfo in partition.h and put
only the list of matched bare clauses into Append.

Thanks,
Amit

[1]
https://www.postgresql.org/message-id/CAKJS1f8T_efuAgPWtyGdfwD1kBLR-giFvoez7raYsQ4P1i2OYw%40mail.gmail.com