Benchmarks

Wall-clock formatting speed of arity against other R formatters, measured with hyperfine. The default comparison is against air; styler can be added opt-in (see below). Every tool formats stdin to stdout (exit 0 regardless of changes), so the comparison is free of file-mutation and exit-code noise.

This is not a CI gate and not a parity target. Timings are machine- and run-dependent, and these numbers measure speed only, never output equivalence (see AIR_COMPAT.md or task air-compat for that). The tools also pay very different startup floors: styler runs inside an R process, so a large part of its time on small inputs is interpreter startup rather than formatting work. Treat the ratios, not the absolute milliseconds, as the takeaway.

The figures below are regenerated manually with task bench and committed as a machine-readable artifact (benches/benchmark_results.json); they are never re-measured when this site is built or in CI.

How it is measured

Each tool is invoked exactly as a user would pipe a file through it, stdin to stdout:

arity is the baseline; every other tool’s time is reported relative to it. Comparison tools absent from the machine are skipped, so a run with only air installed compares just arity and air. The timing backend prefers hyperfine (warmup plus stddev/min/max); without hyperfine and jq it falls back to a mean-only shell loop and the min/max columns become blank.

styler is an R package: it pays an interpreter startup floor plus a steep per-line cost (seconds even on the small tier), so it is not measured by default. Opt in with ARITY_BENCH_STYLER=1 task bench; even then it is skipped on tiers too large to format in reasonable time, so it may appear for only some tiers. Because the tools do such different work, this is a rough scale comparison, not a like-for-like one.

Corpus

The corpus is synthetic: every tests/fixtures/formatter/*/expected.R is concatenated (sorted, blank-line separated) into a base block, which is repeated to two size tiers. The content repeats, so it is cache-friendly and not fully representative of real code; it exists to amortize process startup and show rough scaling, not to model a real workload. (Pass ARITY_BENCH_INPUT to benchmark a single real file instead.)

Setup

• arity: 0.11.0
• air: 0.10.0
• backend: hyperfine (min runs: 3)
• host: linux/x86_64, Intel(R) Core(TM) Ultra 7 155U
• generated: 2026-07-10T00:58:49Z

Results