You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Quiesce vlt's detached children between runs, pin --allow-scripts for vlt ci, rank by median #197
Three things the round-4 install performance review found in the harness that change what the public numbers measure.
vlt's detached children run into the next timed run. After a cold or no-lockfile install vlt leaves up to four detached processes running: the cache-unzip explode child (1.5–11 s of CPU per cold install), the manifest revalidate child (3.6–10 s after a no-lockfile install with moving selectors), rollback-remove and the security-archive update. hyperfine returns when the parent exits, so the next run of the same cell, and the other package managers' runs in the same job, share the box with them. No prepare / conclude step waits for them. Add a quiesce step for every package manager (wait for the PM's child processes to exit, or kill them, before the next run), and document it.
Pin scripts off for vlt ci like for the others.suite/scripts/variations/ci.sh:26 runs vlt ci --view=silent while npm, pnpm and bun get --ignore-scripts. vlt ci no longer runs scripts by default since vlt 1.3.2, so this is hygiene, but --allow-scripts=':not(*)' keeps the cell honest if the default ever changes.
Rank by median, publish both.suite/scripts/clean-benchmarks.js computes a median but apps/www ranks and charts means (src/lib/measurements.ts). The cache cell's day-to-day CV went from 1–3 % to 10–60 % when vlt's resolve endpoint arrived (first-run spikes of 1.3–9 s); medians would have shown the floor and the spikes separately.
Also worth a line in the README: the workflow exports VLT_REGISTRY=https://registry.vlt.io/vlt-benchmarks/npm/ (.github/workflows/benchmark.yaml:208), which is what makes vlt's server-side resolve live on the runner. Users configured through the npm alias do not get it, so the public cache / clean cells measure a path most users do not run today.
From the vlt install performance review, round 4 (2026-10-08), re-checked against vltpkg/benchmarks main 654814ed on 2026-11-12.
--allow-scripts on vlt ci: not done, on purpose.vlt ci runs no dependency scripts by default, and the audit confirmed it by observation: esbuild's postinstall never ran. Passing a selector would only restate the default inside vlt's timed command.
The README note about VLT_REGISTRY is now a METHODOLOGY paragraph in #208: each tool is measured in its default configuration, which for vlt is its own registry.
Activity
lukekarrys commentedon Nov 18, 2026
Status of the three items:
vlt-cache-unzip. Quick fixes from the benchmark audit: zpm/berry caches, vlt children, ci+setup, Yarn 1 counts #208 extends them to all five children in vlt 1.8.0: cache-unzip, revalidate, rollback-remove, the security-archive update and the update check. A logging-proxy run in a CI-like sandbox showed no vlt child alive when any timed command started (0 of 3 inclean, 0 of 15 inci+setup). The 2026-11-17 audit saw no other package manager leave processes behind, so the wait only targets vlt.--allow-scriptsonvlt ci: not done, on purpose.vlt ciruns no dependency scripts by default, and the audit confirmed it by observation: esbuild's postinstall never ran. Passing a selector would only restate the default inside vlt's timed command.The README note about
VLT_REGISTRYis now a METHODOLOGY paragraph in #208: each tool is measured in its default configuration, which for vlt is its own registry.This closes when #208 merges.
🤖 Generated with Claude Code