tpcgen-cli

tpcgen-cli is a high-performance, parallel command line interface for generating TPC-H and TPC-DS data.

It generates data in multiple formats including Apache Parquet, CSV, and pipe-delimited (tbl for TPC-H, dat for TPC-DS).

This tool is more than 10x faster than the next fastest TPC-DS generator we know of (duckdb), 4x-27x faster than the official C dsdgen, and 12x-69x faster than the Java port used by Trino. On an 8 core AWS instance it generates data at over 1 GB/s, faster than a single EBS volume can write. See BENCHMARKS.md for more details on performance and benchmarking.

Try with uvx

uvx tpcgen-cli tpch parquet -s 1 --output-dir /tmp/tpch

Install with pip

python -m pip install tpcgen-cli

Examples

# TPC-H Scale Factor 1, all tables, in `tbl` (pipe-delimited) format in the `/tmp/tpch` directory
# Note: `tpcgen-cli tpch tbl` also works explicitly
tpcgen-cli tpch -s 1 --output-dir /tmp/tpch

# TPC-H Scale Factor 1, all tables, in CSV format
tpcgen-cli tpch csv -s 1 --output-dir /tmp/tpch

# TPC-DS Scale Factor 10, all tables, in Apache Parquet format,
# 10 part(itions) written to `/tmp/tpcds/<table_name>/<table_name>.<part>.parquet`
tpcgen-cli tpcds parquet -s 10 --parts 10 --output-dir /tmp/tpcds

# TPC-DS Scale Factor 100, store_sales table, in Apache Parquet format,
# 10 part(itions) written to `/tmp/tpcds/store_sales/store_sales.<part>.parquet`
tpcgen-cli tpcds parquet -s 100 --tables store_sales --parts 10 --output-dir /tmp/tpcds

# TPC-H Scale Factor 1, Per-column encodings 
# (overrides Parquet writer defaults for named columns)
tpcgen-cli tpch parquet -s 1 --tables lineitem --column-encoding=l_comment=DELTA_LENGTH_BYTE_ARRAY --output-dir /tmp/tpch

# TPC-DS Scale Factor 1, all tables, in CSV format in the `/tmp/tpcds` directory
# Note: `tpcgen-cli tpcds` with no format writes `dat` (pipe-delimited) files
tpcgen-cli tpcds csv -s 1 --output-dir /tmp/tpcds

Performance

Times to create TPC-DS tables in Parquet format using tpcgen-cli and duckdb for various scale factors on an 8 core AWS c8gn.2xlarge. See BENCHMARKS.md for details.

• DuckDB (proprietary) is the time required to create TPC-DS data in the proprietary DuckDB format (CALL dsdgen()). DuckDB is that time plus EXPORT DATABASE (FORMAT parquet).
• At Scale Factors 100 and 300 the DuckDB export to Parquet failed on this 16 GB machine, so only the proprietary format time is reported.