csv port experiments

CSV port experiments

CSV port benchmark

This harness measures the public Mojo read_csv API and pinned Polars 1.44.2 on one caller-supplied file. There is no internal-reader selection: the public API is the sole implementation in this checkout. Each run uses this fixed explicit schema:

id:Int64,value:Float64,active:Bool,label:String

The file must have that header. The full scenario reads all four columns. The projected scenario reads id,label, so Float64 and Bool parsing is excluded in every engine. Each process warms once, retains raw read samples as CSV, and checks full frame equality (Mojo) or Polars frame equality plus shape/null counts. The timed region contains only the read call.

Create the deterministic fixture outside the timed process (numeric nulls, quoted separators, escaped quotes, Unicode, and occasional embedded newlines):

python3 experiments/csv_port/prepare.py build/csv_port/mixed.csv --rows 1000000

Build the public Mojo harness without timing it:

pixi run mojo build -I . experiments/csv_port/bench_csv_port.mojo -o /tmp/bench_csv_port_current

With an existing build/csv_port/mixed.csv, collect raw samples in separate processes. DATAFRAME_THREADS must be set before launching Mojo and POLARS_MAX_THREADS before importing Polars.

mkdir -p experiments/csv_port/results
DATAFRAME_THREADS=1 /tmp/bench_csv_port_current build/csv_port/mixed.csv 8 full > experiments/csv_port/results/mojo_public_current_full_t1.csv
POLARS_MAX_THREADS=1 pixi run -e oracle python experiments/csv_port/bench_csv_port_polars.py build/csv_port/mixed.csv 8 full > experiments/csv_port/results/polars_full_t1.csv

DATAFRAME_THREADS=32 /tmp/bench_csv_port_current build/csv_port/mixed.csv 8 projected > experiments/csv_port/results/mojo_public_current_projected_t32.csv
POLARS_MAX_THREADS=32 pixi run -e oracle python experiments/csv_port/bench_csv_port_polars.py build/csv_port/mixed.csv 8 projected > experiments/csv_port/results/polars_projected_t32.csv

Repeat both commands for both scenarios at each desired thread count. Keep every output file; calculate medians only afterward:

python experiments/csv_port/summarize.py experiments/csv_port/results/*.csv

For a before/after Mojo claim, build this same public-only harness in an unchanged main checkout (or use a recorded baseline binary), then use the same fixture, environment, and invocation:

git worktree add --detach /tmp/dataframe-mojo-main-bench origin/main
cp experiments/csv_port/bench_csv_port.mojo /tmp/dataframe-mojo-main-bench/experiments/csv_port/
cd /tmp/dataframe-mojo-main-bench
pixi run mojo build -I . experiments/csv_port/bench_csv_port.mojo -o /tmp/bench_csv_port_main
DATAFRAME_THREADS=32 /tmp/bench_csv_port_main /path/to/mixed.csv 8 full > /path/to/mojo_public_main_full_t32.csv

Record both commits and the binary build commands with new samples. The retained result directories are historical evidence from earlier internal-port stages; their legacy and explicit labels are deliberately unchanged and must not be mixed with a public-API comparison without recording the build provenance.

Untimed cross-engine validation:

pixi run mojo build -I . experiments/csv_port/verify_reader.mojo -o /tmp/csv-verify-reader
pixi run -e oracle python experiments/csv_port/verify_reader_oracle.py /tmp/csv-verify-reader

The initial source-port readout and raw samples are in results/initial-source-port.

Expanded deterministic differential matrix (24 comparisons):

pixi run mojo build -I . experiments/csv_port/verify_reader_differential.mojo -o /tmp/csv-verify-reader-differential
pixi run -e oracle python experiments/csv_port/verify_reader_differential.py /tmp/csv-verify-reader-differential