Proposal: Change `Accumulator` trait to accept `RecordBatch` / `num_rows` to allow faster `Count`

[Dandandan]

Is your feature request related to a problem or challenge?

Currently the CountAccumulator implementation requires values: &[ArrayRef] to be passed.

In order to eliminate scanning a (first) column, we need to be able to accept a RecordBatch or num_rows instead of values: &[ArrayRef].

Describe the solution you'd like

Rather than changing every method to accept a RecordBatch (and needing to update the code), I propose adding two new methods:

update_record_batch(&mut self, recordbatch: &RecordBatch) retract_record_batch(&mut self, recordbatch: &RecordBatch)

The default implementation of the methods can use update_batch and update_record_batch (i.e. assume having at least one column).

In the aggregation code, we call update_record_batch/retract_record_batch instead.

Describe alternatives you've considered

No response

Additional context

No response

[2010YOUY01]

Was that because this counting operation is possible to be done during scanning?

Looks like it's a case of aggregate pushdown. For min()/max()/count() aggregate functions on Parquet, it's possible to get the result on whole column only use metadata, without full scan.

To do that i think update_record_batch() is needed, possibly also allow RecordBatch to carry more flexible payloads

[alamb]

I think using a RecordBatch rather than &[ArrayRef] makes sense to me

If we are going to change the API anyways, I recommend considering changing the signature to ColumnarValue so it can handle either a RecordBatch or a ScalarValue

[alamb]

The other thing maybe we can think about while messing with the Accumulator trait is how we might expose GroupsAccumulator as well 🤔

[Dandandan]

I looked a bit more into this, it looks currently we're getting away mostly by converting 1 scalars as "count expression" (count(Int64(1)) to an array with to_array_of_size. This is a bit wasteful, but also not extremely bad (as long as the size is not enormous).