Skip to content

dataframe-quickstart.md

Latest commit

 

History

History
124 lines (85 loc) · 3.78 KB

dataframe-quickstart.md

File metadata and controls

124 lines (85 loc) · 3.78 KB
Raw

DataFrame Benchmarking Quickstart

Run TPC-H using DataFrame APIs (Polars, Pandas) instead of SQL.

Why DataFrame Benchmarking?

  • Compare paradigms: SQL vs DataFrame execution on same queries
  • Data science workflows: Benchmark tools you actually use
  • No database required: Pure Python execution

Quick Start with Polars

benchbox run --platform polars-df --benchmark tpch --scale 0.1

This runs all 22 TPC-H queries using Polars' DataFrame API.

Available DataFrame Platforms

Platform CLI Name Best For
Polars polars-df Speed, production workloads
Pandas pandas-df Compatibility, reference
PySpark pyspark-df Distributed computing
DataFusion datafusion-df Arrow-native processing

SQL vs DataFrame Comparison

benchbox run --platform duckdb --benchmark tpch --scale 0.1 --output ./benchmark_runs

benchbox run --platform polars-df --benchmark tpch --scale 0.1 --output ./benchmark_runs

The first command runs SQL execution on DuckDB. The second runs DataFrame execution on Polars. Find the two result files with benchbox results --limit 2 and pass them to benchbox compare.

How It Works

BenchBox translates TPC-H queries into DataFrame operations:

SQL (TPC-H Q1):

SELECT l_returnflag, l_linestatus,
       SUM(l_quantity), SUM(l_extendedprice)
FROM lineitem
WHERE l_shipdate <= DATE '1998-12-01' - INTERVAL 90 DAY
GROUP BY l_returnflag, l_linestatus
ORDER BY l_returnflag, l_linestatus

DataFrame (Polars):

lineitem.filter(
    pl.col("l_shipdate") <= date(1998, 9, 2)
).group_by(
    "l_returnflag", "l_linestatus"
).agg(
    pl.col("l_quantity").sum(),
    pl.col("l_extendedprice").sum()
).sort("l_returnflag", "l_linestatus")

Both produce identical results.

Performance Tuning

Enable auto-tuning for optimal performance:

benchbox run --platform polars-df --benchmark tpch --tuning auto

Or use a custom configuration:

benchbox tuning defaults --platform polars

benchbox run --platform polars-df --benchmark tpch --tuning tuning.yaml

The first command shows the available settings. The second runs with a custom tuning file.

Pandas Example

benchbox run --platform pandas-df --benchmark tpch --scale 0.01

Pandas is slower than Polars for large datasets. Use scale factors of 0.01 to 0.1 for Pandas benchmarks.

Cross-Platform DataFrame Comparison

benchbox run --platform polars-df --benchmark tpch --output ./benchmark_runs
benchbox run --platform pandas-df --benchmark tpch --output ./benchmark_runs
benchbox run --platform datafusion-df --benchmark tpch --output ./benchmark_runs

Find the three result files with benchbox results --limit 3 and pass them to benchbox compare. The three runs and the final comparison cover the DataFrame platforms.

TPC-DS Support

DataFrame mode also supports TPC-DS (99 queries):

benchbox run --platform polars-df --benchmark tpcds --scale 1

Note: TPC-DS also supports sub-SF1 development runs with the patched generator, but those runs remain unofficial.

Programmatic Usage

DataFrame execution is driven through the CLI (benchbox run --platform polars-df ...). For embedding benchmarks in Python, use the same benchmark API as SQL mode (TPCH(...).generate_data(), get_query(n)) and execute through a DataFrame adapter. See the DataFrame Platform Guide for the adapter interface.

Next Steps