Expose set_data_page_v2_compression_ratio_threshold for Parquet
- Dominant language
- Rust
- Stars
- 9.3k
- Forks
- 2.4k
- Avg merge
- 3d 7h
- Merged PRs (30d)
- 344
Description
### Is your feature request related to a problem or challenge?
Writing Parquet V2 can benefit from [set_data_page_v2_compression_ratio_threshold](https://docs.rs/parquet/58.3.0/parquet/file/properties/struct.WriterPropertiesBuilder.html#method.set_data_page_v2_compression_ratio_threshold), which is not exposed in Datafusion.
### Describe the solution you'd like
Expose `data_page_compression_ratio_threshold` on `datafusion_common::config::ParquetOptions` and `ParquetColumnOptions` so it can be passed along to `WriterPropertiesBuilder`.
This primarily enables `datafusion_datasource_parquet::file_format::ParquetFormat` to use this setting.
### Describe alternatives you've considered
We are evaluating if we should always disable compression for sorted delta-encoded columns. This can easily be done with `ParquetColumnOptions.compression` today.
We can also create a wrapper `ParquetSink` to set more options on the `WriterPropertiesBuilder` but that's quite a painful workaround for little gain.
### Additional context
We experience poor compression ratios (<5%) with Zstd for delta-encoded columns of timestamps and other types. Disabling compression with a threshold would be ideal to avoid the decompression overhead at query time.
Contributor guide
Research direction
Start with datafusion_common::config::ParquetOptions and ParquetColumnOptions, then trace how datafusion_datasource_parquet::file_format::ParquetFormat builds WriterPropertiesBuilder. Check the linked parquet API for set_data_page_v2_compression_ratio_threshold; done means the option is exposed and passed through so Parquet V2 writing can apply the configured threshold.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- rust
- Domain
- data-engineering
- Issue type
- Feature
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 68/100