[C++][Parquet] Add configure to limit the row group size

[wecharyu]

Describe the enhancement requested

Arrow has one config max_row_group_length_ to limit the row number of a row group, but does not limit the size, which is even more important, see parquet doc.

And other implementations like Java has such config:
https://github.com/apache/parquet-java/blob/79977453b8cd65e6244f16316fac3a510aa87aa8/parquet-hadoop/src/main/java/parquet/hadoop/ParquetWriter.java#L35

Component(s)

C++

[tusharbhatt7]

Hi, I’d like to work on this issue.

To confirm the expected behavior: the idea is to add a new ParquetWriter property (similar to parquet-mr’s maxRowGroupSize) that limits the byte size of a row group, complementing the existing max_row_group_length (which limits the number of rows).

My plan:

1. Add a new optional config field in parquet::WriterProperties for max row group size in bytes.
2. Enforce this condition in parquet::RowGroupWriter so a new row group is started once the accumulated buffered size exceeds the limit.
3. Mirror semantics from parquet-mr implementation.
4. Add unit tests verifying correct splits for variable-size columns.
5. Update documentation.

Please let me know if this approach matches what you had in mind or if you’d prefer a specific naming/config structure.
Thanks — Tushar

[wecharyu]

Hi @tusharbhatt7, I have created a PR, could you help review it?

[tusharbhatt7]

Hi @tusharbhatt7, I have created a PR, could you help review it?

Thanks for tagging me! I'm still new to the Arrow codebase, but I will go through the PR and try to understand the changes. I may not be able to give a full review yet, but I’ll share any observations I have.

Thanks again for the update!

[wgtmac]

Issue resolved by pull request 49527
#49527