All API performance comparisons

All 26 exported operation families are measured through the public package, in synchronous and Promise forms. Each API has its own table below, including every measured loss. This is a baseline for choosing native work, not a claim that replacing every Node call with Rust is beneficial.

Reading the results

Measured on Apple M4 Pro, 14 logical CPUs, darwin/arm64, release build. Runtimes: v22.22.0, v24.21.0. 2 warmups and 10 samples per cell; tables show median milliseconds for the entire workload, lower is better. The reports include p90, raw wall/CPU samples, memory deltas and logical throughput. These are local warm-cache results; Windows/Linux performance has not been measured.

• single is one call; batch-256 is 256 distinct paths (sync sequential, async bounded to eight in flight). This is caller batching, not a Vooya bulk-call extension.
• Tree cases have 4 or 1,000 files of 4 KiB each, nested directories and one empty directory. File cases also cover an 8 MiB Buffer. UTF-8 cases contain non-ASCII text; stat/lstat separately exercise links.
• fs-extra and graceful-fs are Node wrappers for ordinary calls. They are useful migration baselines, not independent Rust engines. rimraf and mkdirp have both their default and explicit manual implementations shown.
• Vooya c4 explicitly selects four internal traversal workers. Node fallback means the public Vooya API delegates to Node; sync cp/rm have no native acceleration, including their c4 column.
• copyFile / preserve is a composed stat + copy + timestamp/mode restoration workload matching copy-file. scan peers compose crawl + lstat + complete metadata objects + sorting; Node has no identical scan method.
• Every sample starts with freshly prepared equivalent state. Setup, GC, independent result/content/metadata validation and cleanup are excluded. Mutations are fully checked after timing. No fsync/flush is requested; copies may use filesystem cloning. Logical throughput is not physical disk bandwidth.
• RSS/heap/external deltas are not peaks. Sub-millisecond single-call rankings are particularly noisy. Neither cold I/O, process startup, error races nor event-loop responsiveness is measured.

Versions: fs-extra 11.4.1; graceful-fs 4.2.11; rimraf 6.1.3; mkdirp 3.0.1; copy-file 11.1.0; fdir 6.5.0; tinyglobby 0.2.17; fast-glob 3.3.3; glob 13.0.6.

Adapter contracts, peer sources and reproduction instructions . The earlier real-repository glob/readdir comparison provides additional tree shapes and multi-pattern workloads.

Batch overview

One representative async row per API/runtime: batch-256 (Buffer or default variant), or tree-1000. The full tables below also show synchronous, tiny, large-file and other variants. The peer column is the fastest non-Vooya implementation in that row, including Node and its wrappers; close differences are not robust rankings.

Per-API tables

access

F_OK on existing regular files. Permission failures are not timed.

appendFile

Append one payload to an existing file containing an identical payload. Fresh state per sample; no flush.

chmod

Change regular-file permissions to 0640. This is a POSIX workload.

chown

Set each file to the current uid/gid. This measures the allowed syscall path, not a privileged ownership change.

copyFile

Default rows copy bytes/mode to a fresh destination. Preserve rows also restore timestamps/mode and request advisory cloning. copy-file only participates in the latter contract. Filesystem cloning can dominate large-file results; these numbers are not disk bandwidth.

cp

Recursive tree copy, new destination, timestamps not explicitly preserved. fs-extra uses copy; graceful-fs uses cp. Vooya sync is Node fallback; c4 does not change that route.

exists

Existing and missing paths are separate. Node async is access mapped to a Boolean; fs-extra uses pathExists.

glob

Pattern **/*, relative strings for files and directories. No dotfiles or symlinks in this synthetic fixture. fs-extra/graceful-fs wrap Node glob.

link

Create a hard link and verify its inode and link count.

lstat

Separate regular-file and symbolic-link metadata cases. This suite also checks file/link classification.

mkdir

Create missing a/b parents recursively. Compare directory effects; peer-specific return values are outside the shared contract.

mkdtemp

Create a unique directory for every prefix; verify each returned path exists.

readFile

Whole-file materialization, Buffer or mixed non-ASCII UTF-8. No streaming or line-range extension measured.

readdir

Recursive relative-string listing, including directories. fdir is an independent crawler; fs-extra/graceful-fs wrap Node.

readlink

Read a relative file-symlink target; no dereferencing.

realpath

Resolve existing absolute regular-file paths. No symlink chain in this workload.

rename

Move within the same temporary directory/filesystem; verify destination bytes and source disappearance.

rm

Remove a freshly populated tree. fs-extra remove and rimraf default use Node on this platform; explicit rimraf-manual is also shown. Vooya sync is Node fallback.

rmdir

Remove empty directories only; recursive deletion is in rm.

scan

Return sorted path/name/kind/size/mode/mtime/depth for files and directories. Node, fs-extra, graceful-fs, fdir and tinyglobby are explicitly composed pipelines with lstat; async metadata concurrency is eight. All construction/sorting is timed.

stat

Metadata for regular files and following symbolic links, measured separately.

symlink

Create relative file symlinks and verify their targets.

truncate

Shrink existing 4 KiB files to 1 KiB; verify retained bytes.

unlink

Remove existing regular files, one path per call.

utimes

Set fixed access and modification timestamps and verify both.

writeFile

Write to fresh distinct paths, Buffer or mixed non-ASCII UTF-8. No flush or overwrite-race semantics measured.

Raw evidence and reproduction

• Node 22 raw samples
• Node 24 raw samples

pnpm install --frozen-lockfile
pnpm build
pnpm test:competitor-contract
pnpm perf:all-apis --output .perf/all-apis.json
# Focus a follow-up:
pnpm perf:all-apis --apis cp,rm,scan --output .perf/batches.json

No runtime default or implementation was changed for these measurements. Before promoting a fast path, repeat the relevant distributions in the target application and platform, with the required options and durability.