Competitor benchmark evidence

This compares overlapping workloads adapted from fdir and tinyglobby’s own benchmarks against the measured pre-release Vooya public entry. It is a baseline for choosing optimizations. These reports measure the reviewed release build, including the documented glob compatibility corrections.

Findings

• On the real typescript-eslint source tree, native sync glob wins the three full-tree matching workloads against the other tested synchronous APIs.
• Native async glob beats Node’s built-in glob but loses to the fastest tested fdir/tinyglobby implementation. These two facts must not be conflated.
• Default native recursive readdir loses to fdir in both modes on that tree. A follow-up concurrency sweep below reduces native async time by about 61–62% with four workers and sharply reduces CPU cost, bringing it close to fdir.
• packages/*/tsconfig.json uses Vooya’s Node fallback in both modes. Its results demonstrate compatibility/routing overhead, not Rust acceleration.
• Tiny and wide/flat trees remain important counterexamples. A single deep-tree win is not evidence of a universally faster crawler. Close timings should be treated as comparable until repeated in the target application.

Provenance and equivalent work

The pinned upstream scripts, full MIT notices, modifications, fixture setup and reproduction commands are in test/performance/competitors/SOURCES.md .

The upstream tinyglobby script uses a typescript-eslint clone. We pin it to 338fad98047cf80b88e2090c85dab9def3e2e782 and use the source archive without installing dependencies. The benchmark verifies a manifest of names, types and sizes before timing. Original benchmark defaults sometimes return files only, all entries, or Dirents; this adaptation returns relative strings for matching files AND directories throughout, with no order requirement. Glob excludes dot entries; plain readdir includes them. Gitignore is disabled. Symlinks/special files are rejected by fixture validation. This is a deliberately shared subset, not a complete library conformance claim.

A directory named folder.js, hidden entries, duplicates and omitted entries are covered by adapter tests. An independent fixture manifest selects exact expected paths for the simple patterns. Every warmup and sample is checked. Only slash spelling/trailing directory slash are normalized outside timing. If an application needs canonicalized paths, its conversion cost must be measured separately.

Method

• Apple M4 Pro, 14 logical CPUs, macOS arm64. Node 22.22.0 and Node 24.21.0.
• Release native build; fdir 6.5.0, picomatch 4.0.7, fast-glob 3.3.3, tinyglobby 0.2.17, glob 13.0.6. The pnpm lockfile fixes transitive versions.
• Two warmups, ten samples per library/case. Tables show median milliseconds. Engine order rotates per sample and per case; raw evidence also includes p90, CPU samples and RSS deltas. No speed thresholds are asserted in CI.
• Each call constructs its crawler/matchers and fully materializes the result. Sync timings do not include an artificial await; async Node glob is fully collected. Caller-managed persistent glob caches are not exercised.
• Warm filesystem and library-internal caches; setup, GC and validation excluded. These are repeat-call measurements, not cold storage or process startup.
• Library defaults: native glob uses four workers, recursive readdir the default Rayon pool. UV_THREADPOOL_SIZE and RAYON_NUM_THREADS were unset. The table is a comparison of these adapter/default configurations, not every library’s best possible manually tuned configuration.
• RSS is immediately-after minus immediately-before, not peak memory. Zero does not mean no allocation. All libraries are loaded into the same process; process CPU time includes native worker threads.
• Vooya routing is verified by untimed instrumentation restored before timing. Other than choosing the public API and documented options, the runtime code is unchanged. Source, lockfile and native binary hashes are included in each report.
• Not measured: peak memory, event-loop delay, startup, streaming first result, cold-cache I/O, filesystem mutation/error races, metadata-heavy scan, Rust-only libraries or Windows/Linux speed. Platform CI tests adapter correctness only.

Fixtures

Full results

all = **/*; configs = packages/*/tsconfig.json; javascript = **/*.js; multi = four patterns for .ts, .tsx, .js, .md; readdir = recursive listing without glob. The multi-pattern and synthetic cases are Vooya supplements. A dash means the library was not entered in the plain-readdir comparison.

v22.22.0 — tiny

v22.22.0 — typescript-eslint

v22.22.0 — wide-20000

v24.21.0 — tiny

v24.21.0 — typescript-eslint

v24.21.0 — wide-20000

Focused readdir concurrency diagnostic

The default real-tree readdir run consumed substantially more CPU than fdir. This follow-up keeps the exact recursive relative-path result and public async entry, varies only Vooya’s documented concurrency option, and remeasures Node and fdir alongside it. Same two warmups, ten samples, rotating order and validation. These measurements are separate runs; compare baselines within each table.

Four workers reduce wall time by about 61–62% and process CPU time by about 85–86% versus the native default on this machine and fixture. Their 18.405/20.174 ms medians are close to fdir’s 19.343/20.251 ms on Node 22/24; these small differences do not establish a repeatable lead. Eight workers regress relative to four. The default’s high CPU cost supports investigating scheduling/oversubscription, but does not prove which internal operation dominates or justify changing defaults for all machines. No default was changed.

What to investigate next

1. Profile recursive readdir scheduling, default worker count and path/result allocation. The concurrency sweep identifies a concrete low-scope experiment.
2. Profile async glob against fdir/tinyglobby, separating traversal, matching and N-API result conversion. The large Node baseline gap hides the smaller, more relevant competitor gap. Sync wins do not establish async competitiveness.
3. Test targeted subtree planning. fdir’s generic root crawler in this adapter scans the full source tree for configs, while the other glob planners can constrain the search. Manual fdir root/depth tuning is not included, so that row is not a claim that fdir cannot be used efficiently for a known layout.
4. Keep the tiny and flat cases as counterweights. The full API scan comparison now covers metadata-heavy work; extend it to larger real repositories and additional platforms before expanding performance recommendations.

Reproduce and inspect raw data

Prepare the pinned fixture using SOURCES.md, then:

pnpm build
pnpm perf:competitors --fixture .perf/competitors/typescript-eslint --output .perf/competitors-node22.json
npm exec --yes --package=node@24.21.0 -- node --expose-gc scripts/benchmark-competitors.mjs --fixture .perf/competitors/typescript-eslint --output .perf/competitors-node24.json
node --expose-gc scripts/benchmark-readdir-concurrency.mjs .perf/competitors/typescript-eslint .perf/competitors-readdir-concurrency-node22.json
npm exec --yes --package=node@24.21.0 -- node --expose-gc scripts/benchmark-readdir-concurrency.mjs .perf/competitors/typescript-eslint .perf/competitors-readdir-concurrency-node24.json

Use Node 22.22.0 for the unprefixed commands. Run each command separately with no concurrent builds/tests. pnpm test:competitor-contract validates adapters without running timing benchmarks, and is wired into the platform CI test matrix.

• Node 22 full samples
• Node 24 full samples
• Node 22 concurrency samples
• Node 24 concurrency samples