Re-chunk 2.0.0 parquet row groups for Hugging Face /blob/ viewer

#11
by heigold - opened
Google org

Re-write the 104 2.0.0/utts_*_*.parquet files and 6 single-file
2.0.0/*_index.parquet files with smaller, 100-row-aligned row groups
(~32 MiB uncompressed target per row group, write_page_index=True) while
preserving identical schemas, metadata, row ordering, and compression.

Previously, row groups targeted ~128 MiB uncompressed (for the 300 MB Dataset
Viewer scan limit), which caused 75/104 utts_*_*.parquet files to exceed the
stricter 100 MB (100,000,000 bytes) read limit of the /blob/ Parquet
preview (BlobParquetPreview disables offset-index slicing when nested columns
such as waveform or candidates are present), and 102/104 files to exceed
100 MB on 100-row pages straddling row-group boundaries.

After re-chunking, the largest compressed row group and worst-case 100-row
/blob/ page read across all files in 2.0.0/ is 31.38 MB (well below the
100 MB limit).

heigold changed pull request status to open
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment