Use Row Format in SortExec

[tustvold]

Is your feature request related to a problem or challenge?

Currently SortExec::sort_batch_stream uses lexsort_to_indices to sort the produced RecordBatch. For multi-column sorts this makes use of LexicographicalComparator. The branching and dynamic dispatch involved in this comparator is relatively expensive. Converting to the row format first, and comparing these rows has been found to offer significant performance advantages in similar applications - #3386.

Describe the solution you'd like

SortExec should:

• If single sort column, use sort_to_indices to sort the input batches
• If multiple columns, convert to the row format and sort using this representation
• If performing a subsequent merge, preserve the row encoding to avoid redundant work

Describe alternatives you've considered

No response

Additional context

This is likely not a good first issue, and I do not recommend people pick it up, creating primarily for tracking purposes. I will likely pick it up at some point in the near-ish future.

[Dandandan]

AFAIK this is implemented

[Dandandan]

Hm looking at the code, this doesn't seem to be the case

[Lordworms]

take

[Lordworms]

I probably need a efficient way to efficiently spill Rows to disk. Right now I am using a dumb way to spill Rows (len + row_data)

[Lordworms]

Current design is

1. substitute SendableRecordBatchStream between SortPreservingMergeExec and SortExec to RowOrColumnStream

pub enum RowOrColumn {
    Row(Rows),
    Column(RecordBatch),
}

/// Contains a Rows or a Recordbatch
pub type RowOrColumnStream = Pin<Box<dyn Stream<Item = Result<RowOrColumn>> + Send>>;

2. In the begining of SortExec, we build the RowConverter, (if it is a single column sort, we don't build this and send a Recordbatch)
3. for every recordbatch SortExec recieved, we convert it into Rows and do spill logic using Rows format(I implemented a rudimentary reader and writer for Rows)
4. in SortPreservingMergeExec we convert the rows to [[ArrayRef]] (We have to do this since I didn't find any arrow methods to directly build Recordbatch from Rows(probably we can implement one and don't need to store extra columns?)) and didn't break any loser_tree logics.