It seems useful to use the index type to set the starting bit width of
the builder. I guess we can preserve the behavior of expanding to the
next bit width when overflowing the smaller integer types.
On Sun, Aug 2, 2020 at 9:32 PM Kenta Murata <mura...@gmail.com> wrote:
>
> Hi folks,
>
> arrow::MakeBuilder function with a dictionary type creates a
> dictionary builder with AdaptiveIntBuilder by ignoring the bit-width
> of DictionaryType's index type.
> I want to know whether this behavior is intentional or not.
>
> I think this feature is useful when I want to use a dictionary builder
> with AdaptiveIntBuilder.
> But the result by following code is a little bit surprising.
>
> ```cpp
> #include <arrow/api.h>
> #include <arrow/util/logging.h>
> #include <iostream>
>
> int
> main(int argc, char **argv)
> {
> auto dict_type = arrow::dictionary(arrow::int32(), arrow::utf8());
> std::unique_ptr<arrow::ArrayBuilder> out;
> ARROW_CHECK_OK(arrow::MakeBuilder(arrow::default_memory_pool(),
> dict_type, &out));
> std::cout << "type: " << out->type()->ToString() << std::endl;
> return 0;
> }
> ```
>
> You can see the message below when executing this code.
>
> type: dictionary<values=string, indices=int8, ordered=0>
>
> I got `indices=int8` from a dictionary type with int32 index type.
> I guess most people expect they get `indices=int32` here.
>
> --
> Kenta Murata
