PatchedArray: basics and wiring

[a10y]

Summary

This is the first PR in a series addressing the PatchedArray RFC: vortex-data/rfcs#27

This PR adds a new PatchedArray array variant, which is slated to only be used with FastLanes array types. The design is largely documented in the RFC, but briefly

• Wraps another child array
• Holds a buffer of offsets and patch values. They are both sorted by chunk/lane so there is efficient random as well as bulk access for CUDA

We are able to pushdown the following at reduce time:

• Filters pushes through to the child, and we trim the patch indices to the range of chunks covered by the filter
• Slicing that reduces the chunk range
• Compare always pushes through to the child
• Take pushes through to the inner then constructs a hashmap of the patches

There will be follow ups to add the wiring into CUDA and to update how BitPacked and ALP arrays are written.

Testing

There are unit tests for all of the reducers and kernels

[codspeed-hq]

Merging this PR will not alter performance

✅ 1009 untouched benchmarks
⏩ 1515 skipped benchmarks¹

---

_{Comparing aduffy/patched-array (35bfb5f) with develop (8efe1dc)}

Footnotes

1. 1515 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[joseph-isaacs]

I really think we should reconsider this.

This array (+ constant) should be used to replace the sparse array so I think we really want to store things other than PTypes?

[a10y]

I'm not sure if this array is well-suited for non-primitive types. How many lanes does a List<Struct{u32,utf8}> have? I was tempted to actually put this into the vortex-fastlanes crate tbh

[a10y]

i've updated to use child array for values, which should give us some flexibility to change things in the future

[joseph-isaacs]

you cannot do this

[a10y]

why? this doesn't access buffers just slices

[joseph-isaacs]

You really want to take(1) and on the slices iterator and backwards slice iterator here

[joseph-isaacs]

my feeling is that we want this to be an array too in-case we later do want to compress this array.