Boolean <-> Integer duality

[gatesn]

We should support converting between strict sorted integers and boolean masks. We may need an array type to go in both directions?

This could allow us to remove the RoaringUInt array

[a10y]

Capturing from slack:

Currently our tableprovider's pushdown is bottlenecked by take(varbin)

DataFusion defers to arrow's filter_bytes function to turn the predicate mask into new ArrayRef:

https://github.com/apache/arrow-rs/blob/920a94470db04722c74b599a227f930946d0da80/arrow-select/src/filter.rs#L660-L689

We want to have our own boolean builder to construct these masks, calculating run lengths, and using that to alternate between slicing/indexing in our implementation of take()

[danking]

I think we resolved this in #1327 and #1351. #1327 switched the predicate/mask argument of filter to FilterMask. I'll close but obviously re-open if there's more work to do!

I profiled the entire TPC-H suite and it seems to be mostly DataFusion stuff. _target_release_tpch_benchmark 2024-12-16 11.51 profile.json.gz