[POC] Add scripts to Partition and Sort tpch datasets

[GregoryKimball]

Summary

Add GPU tooling to create a TPC-H dataset with:

• orders range-partitioned by o_orderdate and sorted by (o_orderdate, o_orderkey)
• lineitem range-partitioned by l_shipdate and sorted by (l_shipdate, l_orderkey)
• File rewriting is done by pylibcudf in a RAPIDS docker container. This is to avoid adding a cudf python dependency to velox-testing
• Source Parquet schema, codec, encoding, row-group target, row count, and file count are preserved
• partsupp, supplier, part, etc are symlinked or copied from the source dataset
• The generic sort_partition_table.py operates on one Parquet table directory. The TPC-H-specific shell wrapper rewrites the tables into an output directory

Testing

• Generated and benchmarked an SF1000 sorted dataset, with conditions (1) pure cuDF rewrites (no repartition), (2) repartition+sort on the date key, and (3) repartition on date key, sort on date plus primary key

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.