Robert Kruszewski
Robert Kruszewski
Once you split compute from the main crate you can split arrow from compute
Seems this would only get the effort it needs if we forced ourselves to go through that api layer. Right now it's simpler to use binding generators so we might...
Seems like #5198 was majority of it
There's still 10% regression in some queries
When I was looking at this before the locking cost was non negligible and ideally we’d take immutable snapshot and have a read only version
CI runs in AWS, you can get whatever machine you want, right now by default we use m7i/m7a which are sapphire rapids/zen4 which have AVX512. On benchmarks we run c6i...
I think if you do this you might still want a threaded runtime (my point of reference is that in spark parquet still spins up a forkjoinpool per task), just...
We have coverage for search_sorted and binary_numeric fwiw.
This is a fun one - the fix here is to widen offsets
Tihs is blocked on lance and tpchgen-rs because of our benchmark setup