[C++] Switching the column ordering of "segment key" and "hash key"

[icexelloss]

Describe the enhancement requested

Recently, this PR changes the ordering output columns of aggregation to
(hash)keys + segment_keys + aggregates
(https://github.com/apache/arrow/blob/main/cpp/src/arrow/acero/aggregate_node.cc#L654)

I wondering it makes more sense to change this to
segment_keys + (hash)keys + aggregates

My reasoning is as follows:
This only really affects the result in the case of segmented aggregation so I will solely reason about this in the context of segmented aggregation (in non-segmented aggregation, segment_keys is empty so the result wouldn't change regardless).

In segmented aggregation, the hash key is used within each segment (i.e., we don't combine grouped from different segments together) so in a way, the hash key feels like "second key", i.e., data is first split up by segments, then by hash keys. And therefore, it feels more natural to order the output columns: primary key (segment keys) + secondary key (hash keys) + aggregates.

To give an example of time ordered aggregations (an application of segmented aggregations). where segment key = "time" and hash key = "id", we almost always wanted the output to be "time, id" instead of "id, time" because the full stream is time ordered so it makes sense to make that the first column.

@westonpace I wonder what is your thoughts on switching this up? If you feel strongly about keeping the current order I think we might still be able to tweak this at the substrait level with emit output to get the order we want, but I figured it's probably easier / more consistent if we just switch up the default ordering.

I am happy to make the change if that sounds ok but want to discuss first.

Component(s)

C++

[westonpace]

I'm fine with this switch. I had no idea which should come first so I left the relative ordering the same when I changed measures | keys | segment_keys to keys | segment_keys | measures. However, your reasoning seems pretty solid to me.

Another option, though trickier, would be to keep keys and segment_keys ordered as they were in the input schema. So, for example, if the user has a, b, c, d, e, f and a and c are segment_keys and b is a key then you could have the output a, b, c,.

As long as time was first in the input then time would be first in the output. I'm +1 either way.

[icexelloss]

Thanks Weston!