Re: path toward faster partition pruning

On 2017/09/30 1:28, Robert Haas wrote:
> On Thu, Sep 28, 2017 at 5:16 AM, David Rowley
> <david(dot)rowley(at)2ndquadrant(dot)com> wrote:
>> I'd imagine, for
>> each partition key, you'd want to store a Datum with the minimum and
>> maximum possible value based on the quals processed. If either the
>> minimum or maximum is still set to NULL, then it's unbounded at that
>> end. If you encounter partkey = Const, then minimum and maximum can be
>> set to the value of that Const. From there you can likely ignore any
>> other quals for that partition key, as if the query did contain
>> another qual with partkey = SomeOtherConst, then that would have
>> become a gating qual during the constant folding process. This way if
>> the user had written WHERE partkey >= 1 AND partkey <= 1 the
>> evaluation would end up the same as if they'd written WHERE partkey =
>> 1 as the minimum and maximum would be the same value in both cases,
>> and when those two values are the same then it would mean just one
>> theoretical binary search on a partition range to find the correct
>> partition instead of two.
>
> I have not looked at the code submitted here in detail yet but I do
> think we should try to avoid wasting cycles in the
> presumably-quite-common case where equality is being tested. The
> whole idea of thinking of this as minimum/maximum seems like it might
> be off precisely for that reason.

I agree. Equality checks are going to be common enough to warrant them to
be handled specially, instead of implementing equality-pruning on top of
min/max framework.

Thanks,
Amit