Re: [HACKERS] path toward faster partition pruning

Hi.

On 2018/04/05 0:45, Jesper Pedersen wrote:
> Hi,
> 
> On 04/04/2018 09:29 AM, David Rowley wrote:
>> Thanks for updating. I've made a pass over v49 and I didn't find very
>> much wrong with it.
>>
>> The only real bug I found was a missing IsA(rinfo->clause, Const) in
>> the pseudoconstant check inside
>> generate_partition_pruning_steps_internal.

Fixed.

>> Most of the changes are comment fixes with a few stylistic changes
>> thrown which are pretty much all there just to try to shrink the code
>> a line or two or reduce indentation.
>>
>> I feel pretty familiar with this code now and assuming the attached is
>> included I'm happy for someone else, hopefully, a committer to take a
>> look at it.

Thank you, your changes look good to me.

>> I'll leave the following notes:
>>
>> 1. Still not sure about RelOptInfo->has_default_part. This flag is
>> only looked at in generate_partition_pruning_steps. The RelOptInfo and
>> the boundinfo is available to look at, it's just that the
>> partition_bound_has_default macro is defined in partition.c rather
>> than partition.h.

Hmm, it might not be such a bad idea to bring out the
PartitionBoundInfoData into partition.h.  If we do that, we won't need the
has_default_part that the patch adds to RelOptInfo.

In the Attached v50 set, 0002 does that.

>> 2. Don't really like the new isopne variable name. It's not very
>> simple to decode, perhaps something like is_not_eq is better?

isopne does sound a bit unintelligible.  I propose op_is_ne so that it
sounds consistent with the preceding member of the struct that's called
opno.  I want to keep "ne" and not start calling it not_eq, as a few other
places use the string "ne" to refer to a similar thing, like:

/* inequality */
Datum
range_ne(PG_FUNCTION_ARGS)

Datum
timestamptz_ne_date(PG_FUNCTION_ARGS)

Since the field is local to partprune.c, I guess that it's fine as the
comment where it's defined tells what it is.

>> 3. The part of the code I'm least familiar with is
>> get_steps_using_prefix_recurse(). I admit to not having had time to
>> fully understand that and consider ways to break it.

The purpose of that code is to generate *all* needed steps to be combined
using COMBINE_INTERSECT such that the pruning will occur using the most
restrictive set of clauses in cases where the same key is referenced in
multiple restriction clauses containing non-equality operators.  So, for a
range partitioned table on (a, b):

For a query like

explain select * from foo a <= 1 and a <= 3 and b < 5 and b <= 10

Pruning steps generated to be combined with an enclosing INTERSECT step
will be as follows:

<= (1, 10)
<  (1, 5)
<= (3, 10)
<  (3, 5)

>> Marking as ready for committer.

Thank you!

> Passes check-world, and CommitFest app has been updated to reflect the
> current patch set. Trivial changes attached.

Merged these changes.  Thanks again Jesper.

Attached v50.

Thanks,
Amit