Leita CLI filesystem scan benchmark, 2026-10-19

This is a measured scan of generated files on a real btrfs mount. The fixture is
synthetic developer-shaped data, not a sample of a person's files or a claim
about cold-disk or general user workloads. Leita is an optimized development
build, g8e42581, not a published release. The full source revision, pinned jm
submodule, Odin version, executable SHA-256 hashes, kernel, CPU and exact mount
options are in results.json.

Reproduction (Linux; run outside tmpfs on a filesystem with at least 2 GB free):

  export LEITA_REPO=/path/to/leita/checkout
  export BENCH_SOURCE=/path/to/isolated/leita-source
  export BENCH_WORK=/path/to/persistent/benchmark-data
  git -C "$LEITA_REPO" worktree add --detach "$BENCH_SOURCE" 8e42581d204c11c22724c0fdbc3757dffd34c949
  git -C "$BENCH_SOURCE" submodule update --init jm
  cd "$BENCH_SOURCE"
  LEITA_LICENCE_KEY=e83dc875975880f373348c2eccbf6b5c719736abf15727e77f3968c0c6644553 just cli release
  python3 /path/to/benchmark.py --workdir "$BENCH_WORK" \
    --leita "$BENCH_SOURCE/build/release/leita" \
    --fd /path/to/fd --ncdu /path/to/ncdu --time /path/to/GNU-time \
    --source-revision 8e42581d204c11c22724c0fdbc3757dffd34c949 \
    --jm-revision 6d85e4750a4c58ca3c1076b78b9cc78def2a4c9e \
    --build-kind 'optimized development build with repository test public key' \
    --build-command 'LEITA_LICENCE_KEY=e83dc875975880f373348c2eccbf6b5c719736abf15727e77f3968c0c6644553 just cli release' \
    --output results.json

The key above is the repository's public TEST key. The actual run used fd 10.5.0,
GNU findutils 4.11.0-modified, ncdu 2.9.2 and GNU Time 1.10. For exact binary
identity, compare their SHA-256 hashes in results.json. ncdu came from the Arch
package named ncdu-2.9.2-1-x86_64.pkg.tar.zst; its package SHA-256 is recorded.

The generator writes 2 fixture roots under BENCH_WORK: 200,000 and 400,000
regular data files. Files are arranged 60% under src, 30% under vendor and 10%
under docs, with 100 files per leaf directory. A seeded PRNG chooses language
extensions and a content length uniformly from 0, 32, 256 and 4,096 bytes;
contents are repeated x bytes. The actual 200,000-file fixture had 49,918 empty,
50,372 32-byte, 49,886 256-byte and 49,824 4,096-byte files (218,461,824 content
bytes). The 400,000-file fixture had 99,587 empty, 100,361 32-byte, 99,931
256-byte and 100,121 4,096-byte files (438,889,504 content bytes). Each root
also has 400 symlinks to directories, a .gitignore and a fixture marker. The
measured operation denominator is all entries beneath the root, excluding the
root: 203,020 and 405,634 respectively. Leita's used count and fd names-only,
fd metadata, and find counts matched those numbers; ncdu JSON had one additional
record for the root.

For each fixture the harness verifies every generated file's content, every
symlink target and the entry-type counts. It then runs each of the 5 modes twice to warm
caches, followed by 7 measured fresh-process runs per mode. Each round shuffles
mode order with the recorded seed. Caches stay warm; no global cache drop or sudo
is used. The host mount was btrfs with zstd:3 compression; full options are in
results.json. The scan commands (with $TREE as fixture root) are:

  leita --no-history --whole --color never $TREE
  fd -HI --one-file-system -0 . $TREE
  fd -HI --one-file-system --changed-before 2100-01-01 -0 . $TREE
  find $TREE -xdev -mindepth 1 -print0
  ncdu -0 --ignore-config -x -o /dev/null $TREE

Standard output goes to /dev/null. Leita writes only its first-use trial state
under an isolated XDG data directory; history/baseline reuse and report output
are disabled. All LEITA_* inherited environment variables are removed. Symlink
traversal, ignored/hidden entries and mount boundaries are aligned. Tool thread
counts are defaults: Leita reported 24 directory readers on this host; fd and
ncdu used their defaults; find uses its normal single process. The fd metadata
mode uses a far-future modification-time bound to include every generated
entry while making fd read its metadata. It runs no external stat command.
The sorted path set matched fd names-only for both fixtures. Untimed strace
checks found 112 statx calls versus 10 for names-only on a 100-file leaf, and
412 versus 10 on a 400-symlink directory.

Total time is Python perf_counter_ns around launch and wait of a fresh GNU Time
wrapper and its target; it includes wrapper startup and output costs. The
median of 7 times is reported. ns/op is that median divided by entries beneath
the root. Peak memory is GNU Time %M maximum resident set size for the target
process, with the maximum of the 7 runs shown in decimal MB (KiB * 1,024 /
1,000,000). Raw elapsed nanoseconds and peak RSS KiB for every measured run are
in results.json. Counts and units are independently checkable there.

The tools do different work: fd names-only and find enumerate names; fd
metadata reads modification times and filters against 2100-01-01; Leita builds
a disk usage tree and summary; ncdu accounts for disk usage and serializes a
JSON scan to /dev/null. The serialization cost is part of ncdu's reported
time. These per-entry figures normalize by the same traversed fixture, not
equivalent work.
Small, compressible files favor metadata-cache scans; this test does not measure
large-file reads, cold storage, network mounts or real project mixtures. Fixture
creation takes roughly 1.8 GB of allocated space on the measured host. The
script never removes an existing fixture unless --regenerate is given, and then
only when its marker exists.