Workspace discovery benchmarks

workloads.rs measures the ignore operations used by Python development tools:

The synthetic trees contain Python modules, package metadata, non-Python files, ignored bytecode directories, a virtual environment, and build outputs. The layouts vary independently: a small project, many sibling packages, deeply nested packages, and packages with their own ignore rules and negations. The parallel visitor counts files locally and merges once per worker, avoiding an artificial shared counter on every entry. One- and four-worker cases supplement the 12-worker configuration used by Ruff and ty. The unfiltered walk models disabling ignore rules and provides a control for traversal overhead. The source-archive case honors .gitignore without a .git directory, matching uv’s require_git(false) configuration.

Fixtures are created before timing and removed afterward. Every walk checks its Python file count against the fixture’s expected count. The benchmarks retain parent ignore discovery, but disable global Git ignore configuration so that results do not depend on a user’s Git configuration. Runs use a warm filesystem cache. They measure ignore, excluding parsing, configuration resolution, and other application work; they are not end-to-end application speedups.

incremental_cold includes building matchers and loading directory rules on first use. incremental_warm checks 256 paths after their directories have been cached. incremental_warm_noncanonical checks the same paths with an internal ./ component, exercising normalization during cached queries. incremental_ignored checks a path beneath a pruned virtual environment. incremental_warm_deep checks 768 paths in a 12-level package tree, and incremental_ignored_parent directly hits an ignored directory’s cache entry. incremental_ignored_nested checks a deeper descendant of a cached ignored virtual environment inside an otherwise allowed package. The walk_dense_ignore_* cases put rules in every directory of a deep tree, providing a control for optimizations that skip directories without rules. incremental_dense_ignore_cold and incremental_dense_ignore_warm use the same tree to check incremental matching when nearly every cached directory has rules. The walk_multiple_roots_* cases pass 64 sibling package directories as roots, covering parent-matcher reuse when Ruff receives multiple directory arguments. walk_with_overrides_and_types exercises explicitly configured override and file-type matchers as a control for their more common empty configuration. The build_* and match_* microbenchmarks help distinguish compilation and matching costs from filesystem traversal. The single-pattern cases cover bytecode exclusions, the directory whitelist used in Ruff’s and uv’s root .ignore files, and ty’s vendored-typeshed whitelist. build_single_directory_whitelist builds that single-rule matcher from an already populated builder. The *gitignore_only* controls remove .ignore and .git/info/exclude, covering projects that rely solely on a .gitignore with multiple patterns. walk_selected_python_files uses two positive override patterns, exercising file-selection rules that match most visited files.

Like the globset benchmarks, these use Rust’s nightly test harness:

cargo +nightly bench -p ignore --bench workloads
cargo +nightly bench -p ignore --bench workloads incremental

For comparisons, use the same compiler, profile, fixture parameters, and CPU affinity for both revisions. Alternate baseline and candidate runs, and retain the per-run results instead of relying on a single measurement. Parallel walks also include worker startup and shutdown, which can dominate small projects.

CPU affinity also affects match-index scratch-pool sizing: available parallelism is cached when the first pool is constructed. Pinning to one CPU therefore changes pool construction costs as well as scheduling. Report affinity with results, especially when comparing construction or cached matching costs.

Ruff, uv, and ty enable fat LTO in release builds, while their profiling builds omit cross-crate LTO. Check promising changes in both configurations, especially when small matching benchmarks give inconsistent results:

CARGO_PROFILE_BENCH_LTO=fat CARGO_PROFILE_BENCH_CODEGEN_UNITS=16 cargo +nightly bench -p ignore --bench workloads

For focused comparisons, pass one benchmark name with --exact so that each case runs in a separate process:

cargo +nightly bench -p ignore --bench workloads -- --exact incremental_ignored_parent