Skip to content

monorepo-reconciliation-v1-2026-07-20.md

Latest commit

 

History

History
79 lines (60 loc) · 4.04 KB

monorepo-reconciliation-v1-2026-07-20.md

File metadata and controls

79 lines (60 loc) · 4.04 KB
Raw

Monorepo reconciliation decision

Date: 2026-07-20

Experiment: monorepo-reconciliation-v1

Workflow: GitHub Actions run 29760207560

LeanToken ran from Git commit b5879d2fe826102d680036cef1613ae0e8c9f4dd. GitHub checked out PR merge commit 829b641566cfe00ff50da18e500678c86f594dbb for the measurement.

The measured merge revision and LeanToken head have the same Git tree, 2ac6379e0c54ec6b8eacbe5a4123c7883c6c7c65. The merge revision is recorded because GitHub Actions checked out the pull-request merge ref. All six raw jobs and the strict aggregate job completed successfully.

Frozen decision rule

The manifest was committed before the matrix ran. A changed-path journal or directory invalidation prototype was eligible only if:

  1. full fallback reached at least 250 ms p95;
  2. discovery plus hash/planning consumed at least 50% of mean full-noop time;
  3. both conditions held in at least two corpus/platform pairs.

The profile used 10 iterations, one read sample, a one-file hot set, and a 50 ms watcher debounce. It ran release binaries from clean worktrees against Tokio 9cae638de6dc8dd9779c450201df8c102247a242 and Vue core 31d0f23757afb410c638a9c29d44d76d0944e18f on GitHub-hosted Linux, macOS, and Windows runners.

Results

Platform Corpus Files Initial index (s) Full no-op p50 / p95 (ms) Scan share Targeted modify p95 (ms) 30-file directory rename p95 (ms) Watcher p95 (ms) Peak RSS (MiB)
Linux Tokio 865 6.23 14.38 / 15.62 99.69% 2.84 331.40 51.86 67.16
macOS Tokio 865 5.95 24.22 / 29.01 99.67% 1.81 298.60 67.41 68.45
Windows Tokio 865 10.41 43.22 / 51.89 99.72% 60.61 1,790.14 57.10 57.35
Linux Vue core 703 16.05 12.48 / 13.74 98.29% 14.05 2,304.62 51.58 83.86
macOS Vue core 703 12.45 24.80 / 32.45 98.46% 11.49 2,728.42 72.17 85.48
Windows Vue core 703 19.11 41.12 / 46.89 99.02% 59.20 4,886.09 60.72 73.85

All native watcher samples produced one changed-path delivery after the quiet period: 10 changed messages and zero full-reconciliation messages per pair. The final SQLite/WAL/SHM logical footprint ranged from 66.79 to 68.14 MiB for Tokio and from 77.80 to 81.60 MiB for Vue core. Windows checked out the same commits with platform line endings, so its visible byte totals differ while file counts and revisions remain fixed.

Hash/planning was the dominant full-noop phase in every pair. The high scan share is not material by itself: every full-noop p95 was between 13.74 and 51.89 ms, at most 21% of the frozen absolute threshold. Therefore zero of six pairs qualified. The strict aggregate decision was:

no-go: 0 corpus/platform pairs met the frozen materiality threshold; retain targeted reconciliation with bounded full fallback

The directory-rename timings do not overturn that result. They include parsing and publishing the moved files plus affected importers, not just rediscovering metadata. They motivate keeping targeted directory handling bounded and correct, but do not isolate a journal-addressable full-scan bottleneck.

Correctness evidence

Portable overflow and interruption performance events were not forced in the matrix. Their safety properties remain deterministic gates:

  • watcher input overflow becomes one sticky full reconciliation;
  • a full output queue degrades changed paths to full reconciliation;
  • startup bursts collapse to one quiet reconciliation and repeated fallback is rate-limited;
  • streamed cancellation, storage failure, and panic roll back uncommitted work;
  • a follower rebuilds after the elected leader is killed during reconciliation.

Absolute timing and RSS values are runner-specific. This two-corpus matrix does not represent cold disks, network filesystems, antivirus configurations, or all repository shapes. Reconsider the no-go only with a newly frozen experiment showing a material full-fallback cost on target workloads.