Fuse grouped aggregate and filter operators for improved performance

[andygrove]

Is your feature request related to a problem or challenge?

When we perform a grouped aggregate on a filtered input (such as with TPC-H q1), the filter operator performs two main tasks:

• Evaluate the filter predicate (usually very fast)
• Create new batches and copy over the filtered data (very slow if the filter is not very selective, as in q1)

I wonder if we would see a significant performance improvement if we could avoid creating the filtered batches in this case.

One idea would be to create the filtered batches by copying the arrays and mutating the validity bitmap to hide the rows that are filtered out. This would potentially change the semantics in some cases though so we can probably only do this under certain conditions.

Another idea is to update the aggregate logic to perform the predicate evaluation and then use the resulting bitmap to determine which rows to accumulate.

Describe the solution you'd like

I am working on a small prototype of this, outside of DataFusion, that I will share once the code is less embarrassing.

Describe alternatives you've considered

It would be worth seeing how other engines handle this.

Additional context

No response

[andygrove]

@Dandandan @alamb wdyt?

[Dandandan]

One other source of inefficiency (and should be relatively easy to change) is that currently we output the entire RecordBatch, including the columns that are needed to evaluate the filter, while throwing those columns away in the following Projection.

See:
See #5436

[Dandandan]

In q1 this would remove l_shipdate from the list of 7 columns to be copied, so I would expect a smaller improvement here.

[Dandandan]

I remind seeing some issue/papers about a similar approach before to this, maybe those were shared by @alamb ?

[alamb]

One idea would be to create the filtered batches by copying the arrays and mutating the validity bitmap to hide the rows that are filtered out. This would potentially change the semantics in some cases though so we can probably only do this under certain conditions.

I think this basic idea is called a "selection vector" in the literature -- and as you hint at, it is not quite the same as the null mask as it has different semantics.

One approach might be to add another enum type to ColumnarValue that had an additional selection mask

https://github.com/apache/arrow-datafusion/blob/bbc71692fcd8dd9f3a9686162e59d092b37031f2/datafusion/expr/src/columnar_value.rs#L33

After @tustvold 's recent work in Arrow, I think this would just be a https://docs.rs/arrow/latest/arrow/buffer/struct.BooleanBuffer.html and should be straightforward to use.

To really take advantage of a selection vector, however, the underlying compute kernels need to be updated to know how to ignore the selection vectors (and likely only do so when they are sparse)

Another idea is to update the aggregate logic to perform the predicate evaluation and then use the resulting bitmap to determine which rows to accumulate.

While not exactly the same, @yjshen 's has been workking to add filtering to the aggregate input here, which is similar: #5868

[tustvold]

apache/arrow-rs#3620 may be related

[tustvold]

This would potentially change the semantics in some cases though so we can probably only do this under certain conditions

I'm curious about this, in what situation would the nullability or not of a non-selected value matter? It is just going to be discarded regardless? See apache/arrow-rs#3620