Benchmarks

This page lists commands and measurement rules. Results stay beside the methods they measure:

• Arrow schema projection and cached Field access;
• plain-text records and record write dispatch;
• placeholder overhead;
• native Scalar and Arrow boundaries;
• natural codec boundaries;
• streamed byte reads;
• structured handle values;
• gzip, zlib, and zstd;
• IPC, Parquet, Avro, and their opened-session metadata sections;
• expressions, Arrow filesystem, and page buffering;
• digests.

Running the benchmarks

RustPythonJavaScript

cargo bench --bench types
cargo bench --bench fix
cargo bench --bench holder --features "parquet"
cargo bench --bench coding
cargo bench --bench media --features "parquet iceberg"
cargo bench --bench text
cargo bench --bench uri
cargo bench --bench expression
cargo bench --bench xxhash

cd python
.venv/Scripts/python benchmarks/types.py --iterations 10000
.venv/Scripts/python benchmarks/types/arrow.py --iterations 10000
.venv/Scripts/python benchmarks/holder.py --min-time 0.2 --repeat 7
.venv/Scripts/python benchmarks/coding.py --min-time 0.2 --repeat 5
.venv/Scripts/python benchmarks/media.py --min-time 0.2 --repeat 7
.venv/Scripts/python benchmarks/text.py --min-time 0.2 --repeat 7
.venv/Scripts/python benchmarks/uri.py --iterations 2000
.venv/Scripts/python benchmarks/xxhash.py --min-time 0.2 --repeat 5
.venv/Scripts/python benchmarks/fix.py --iterations 2000

Build a release wheel with maturin build --release before timing.

npm run --prefix node bench:types
npm run --prefix node bench:holder
npm run --prefix node bench:coding
npm run --prefix node bench:media
npm run --prefix node bench:text
npm run --prefix node bench:xxhash
npm run --prefix node bench:fix

Run on a quiet machine and compare identical release toolchains. Published runs use one containerized x86_64 Linux host (Intel Xeon @ 2.10 GHz, 4 cores, 16 GiB; rustc 1.94.1 with thin LTO; CPython 3.11.15, PyArrow 25.0.1; Node 22.22.2). Regenerate numbers on the deployment host before drawing conclusions.

What each target isolates

target	isolates
types	parsing, construction, validation, mutation, cached access, and Arrow schemas
holder	byte streams, listings, buffering, and foreign-filesystem boundaries
coding	content codings beside their standard-library baselines on the same wire
media	record round trips, text projection, Avro, Parquet, Iceberg, and pushdown
text	natural whole-value and streaming codecs, field-directed parsing, and placeholders
uri	URI parsing and component access
expression	binding, row and Arrow evaluation, and statistics pushdown
fix	registry lookup, mutation, storage, and binding crossings
xxhash	digest throughput per algorithm and size, wrapper overhead over the protocol crate, constant-memory handle reads, the canonical value feed, and Arrow row digests

Rules

• Keep fixtures outside timed loops.
• Keep group identifiers stable so history remains comparable.
• State exactly what each result measures; separate setup from the operation.
• Put a benchmark in its owning domain, beside existing cases.
• Compare with a trusted external implementation on the same payload and wire whenever one exists: stdlib, PyArrow, CPython re, fastavro, or raw Arrow.
• Regenerate published output; never edit measured numbers by hand.
• Binding benchmarks measure the boundary. Recursive conversion and validation belong to the Rust target that implements them.