ARROW-9054: [C++] Add ScalarAggregateOptions

[rok]

ARROW-9054.

This adds ScalarAggregateOptions to control null behavior of mean and sum kernels.

[github-actions]

https://issues.apache.org/jira/browse/ARROW-9054

[rok]

As per comments in Jiras this should introduce the following:

• sum - don't promote type (e.g. int32 to int64) if possible
• sum / mean - should have min_count and return NA if the count of non-nulls is lower than min_count.
• sum / mean - should short circuit when it becomes impossible for non-null count to be below min_count.
• mean - should have skipna and keepna options.

[nealrichardson]

Should MinMax be updated to use ScalarAggregateOptions?

You should also be able to remove https://github.com/apache/arrow/blob/master/r/R/compute.R#L113-L117 and search for any other references to that jira number in the R code/tests.

[rok]

Should MinMax be updated to use ScalarAggregateOptions?

I'll look into it. Count would be a candidate as well.

You should also be able to remove https://github.com/apache/arrow/blob/master/r/R/compute.R#L113-L117 and search for any other references to that jira number in the R code/tests.

Will do.

[rok]

This is slowly coming together. Remaining todo:

• min/max must cover R edge cases
• sum - don't promote type (e.g. int32 to int64) if possible
• sum / mean - should short circuit when it becomes impossible for non-null count to be below min_count.

[jorisvandenbossche]

sum - don't promote type (e.g. int32 to int64) if possible

Is this desired behaviour? I think in general the resulting data type should be known based on the input types and operation (and not depend on the values)

[jorisvandenbossche]

Just to make sure I am reading this one correctly, is it basically testing mean([], min_count=0) == 0 ?

I am not sure that should give 0 as result? I would have expected either NaN or Null.
Basically mean of empty is dividing 0/0, which is NaN.

[rok]

Agreed. I think I've put it there to have it documented while I was changing other things.

[jorisvandenbossche]

It seems you changed it to return null now? Should it rather return NaN? (I assume it could be consistent with a manual mean from sum/count with min_count=0)

[jorisvandenbossche]

Looking forward to this! Added a few comments on naming and exact expected semantics.

[rok]

This requires some more work on C Glib but is ready for review otherwise.

[rok]

@jorisvandenbossche & @nealrichardson could you take a look at this proposal?
I'm not very confident about the c_glib and Ruby changes, who could review that?

[rok]

Should we introduce ScalarAggregateOptions in All and Any as well?

[rok]

Some checks timed out. Rebased to restart the tasks.

[kou]

The "R / AMD64 Windows RTools 35" failure was occurred on master too: https://github.com/apache/arrow/runs/2634288057

It's unrelated to this. I'll merge this.

[rok]

Thanks for the help everyone!