Thank you for this! I have definitely run across the one-byte-per-item bool in numpy, DuckDB, and cudf. I haven't heard any discussion about DuckDB here but I am fairly sure that they represent their boolean type as an int8 as well [1].
> Before the vote, I would like to see verification that this truly enables > zero-copy to/from NumPy bool arrays in Python. I think this is an implementation issue more than a specification issue...I am not personally worried about any provisions on the specification that might make this impossible. -dewey [1] https://github.com/duckdb/duckdb/blob/85a82d86aa11a2695fc045deaf4f88fc63dd4fec/src/common/arrow/appender/bool_data.cpp#L28-L37 On Tue, Jul 16, 2024 at 11:25 AM Antoine Pitrou <anto...@python.org> wrote: > > > Hi Joel, > > This looks good to me on the principle. Can you split the spec and the > implementation(s) into separate PRs? > > Regards > > Antoine. > > > Le 16/07/2024 à 13:18, Joel Lubinitsky a écrit : > > Hi Arrow devs, > > > > I'm working on adding an extension type for 8-bit booleans, and wanted to > > start a discussion about it here because it could be valuable to others if > > adopted as a canonical extension type. > > > > The native implementation of the Boolean type uses 1 bit to encode each > > value, enabling a very compact representation. This is favorable for many > > workloads, but lots of systems that want to produce/consume Boolean arrays > > use an 8-bit representation internally and are forced to copy/convert at > > their periphery. For these scenarios where zero-copy compatibility is > > important, the 8-bit representation of boolean values may be preferred. > > This can benefit interactions with existing libraries that avoid packing > > column data like 1-bit booleans for parallelization purposes, including GPU > > libraries such as libcudf. The original issue [1] identifies numpy > > conversion as a specific use-case as well. > > > > The details of the extension type can be found in the draft PR [2] which > > contains a Go implementation (WIP) and an update to the documentation for > > canonical extension types. I plan to add a C++ implementation as well but > > wanted to open this discussion first. > > > > A quick overview of the layout / semantics proposed in the PR: > > Storage Type: Int8 > > Value Semantics: 0 == false, any non-zero value is true > > > > I'd appreciate any feedback here or on the PR. If this all seems reasonable > > then I'll move forward with the next implementation and open up another > > proposal for a formal vote. Thanks! > > > > [1]: https://github.com/apache/arrow/issues/17682 > > [2]: https://github.com/apache/arrow/pull/43234 > >
