-
Notifications
You must be signed in to change notification settings - Fork 1.1k
Description
Is your feature request related to a problem or challenge? Please describe what you are trying to do.
MutableArrayData provides an API to copy chunks of data from source arrays into an output array, it is extremely flexible, however, its performance leaves a lot to be desired. In particular its reliance on dynamic dispatch per-run not only incurs dispatch overheads, but prevents the compiler from properly vectorising the code. There are also a number of other issues relating to capacity allocation (#347), null mask handling (#1230), dictionary handling (#506)
There has been a fairly concerted effort over the last year or so to move away from use of MutableArrayData:
- Improvements to the Array Builders
- Ability to construct Array from iterators
- Specialized take and filter kernels that only use MutableArrayData as a fallback for nested types
Describe the solution you'd like
I would like to eventually deprecate and remove MutableArrayData. I think the only major blockers to this are:
- The filter kernel uses MutableArrayData as a fallback for filtering nested types
- The concat kernel uses MutableArrayData
- No kernel for interleaving multiple arrays together (
takekernel that works across multipleRecordBatches #1523)
Describe alternatives you've considered
Additional context
This can be viewed through the lens of porting @jorgecarleitao 's ideas from arrow2 and bringing them into arrow-rs