cudf icon indicating copy to clipboard operation
cudf copied to clipboard

[FEA] Add GZIP compression support to parquet writer

Open GregoryKimball opened this issue 2 years ago • 0 comments

Is your feature request related to a problem? Please describe. The parquet format in Apache Spark supports many compression codecs (link), including: none, uncompressed, snappy, gzip, lzo, brotli, lz4, zstd.

cuDF has both internal implementation and an nvCOMP integration to provide compression and decompression codecs. For the parquet format, GZIP compression is DEFLATE plus a header. nvCOMP does not support the deflate version with this header, so the reader still uses the internal gzip decompression implementation. We don't have internal gzip compression implementation. To support GZIP in the PQ writer we would need to use nvCOMP GDEFLATE codec + write the header on our own.

Describe the solution you'd like Add support for GZIP compressioning to the cuDF parquet writer by adding a header writing implementation and using nvCOMP deflate.

Describe alternatives you've considered n/a

Additional context Also see Spark-RAPIDS request here: https://github.com/NVIDIA/spark-rapids/issues/9718

GregoryKimball avatar Nov 28 '23 05:11 GregoryKimball