Replace `BufferBuilder<u8>` with `Vec<u8>`

[jayzhan211]

Is your feature request related to a problem or challenge?

The functionality we need from BufferBuilder<u8> can be handled by Vec<u8>. Switching to Vec will simplify things, and maybe faster (?

• ArrowBytesMap
• ByteGroupValueBuilder
• regex replace

Describe the solution you'd like

1. Replace with Vec<u8>
2. Run the benchmark to see the performance impact

No response

Describe alternatives you've considered

No response

Additional context

Performance should not downgrade.

Suggested by apache/arrow-rs#6909 (comment)

[poweihuang0817]

take

[alamb]

Thanks @jayzhan211 and @poweihuang0817 -- sounds like a good change

Maybe we can do it as a few PRs (for example, start with ArrowBytesMap) -- that will be faster to review and will ensure once we have the pattern we can easily apply it to the other structures

[mzabaluev]

See also this epic in arrow-rs: apache/arrow-rs#10245

[Punisheroot]

Hi, is the ArrowBytesMap slice of this issue still available? I noticed that #13867 was previously assigned and later unassigned, while the related Arrow Rust epic #10245 has a recent take comment. I confirmed that ArrowBytesMap still uses BufferBuilder<u8> on current DataFusion main. I would start with a focused baseline benchmark and limit the first change to this structure. Is that still the preferred scope?

[alamb]

Sound like a good plan to me

[Punisheroot]

Thanks @alamb! I opened #24071 for the focused ArrowBytesMap slice.

It replaces BufferBuilder<u8> with Vec<u8> and adds Criterion coverage for short unique values, long unique values, and long values with low cardinality.

In the Linux comparison, the long-unique case showed 61.24% lower execution time, while the short-value and low-cardinality cases showed no stable regression.

I intentionally kept the PR limited to ArrowBytesMap, leaving the other structures in this issue for separate follow-ups.