Repository navigation
fix: test: our tooling peaks at ~9 cores in one land; cap it at 4 and run components and screenshots in one Vitest start #632
Description
Follow-up to #628/#630 (codeatlasdev/atlas#124 resource governor seam). A land runs Vitest (3 workers), Playwright (3 workers), the site build (~1.5 cores) and bun test at once: load average reached 29-63 on 12 cores during lands. Target: our tooling uses <= 4 cores combined, machine-wide. Also: components and screenshots start Vitest twice (~3 s cold start each).
Activity
mvilacad commentedon Oct 25, 2026
Working on this in #633 (draft). Branch
feat/632-fix-test-our-tooling-peaks-at. Regression test first, then the fix.mvilacad commentedon Oct 25, 2026
Measured, not built. Our tooling peaks at 8 cores in a shared land (p50 4, p90 6). The route sweep does not scale with workers: 111.7 s at 3, 96.1 s at 6 (one flaky timeout), 85.4 s at 2. So capping the sweep at 2 workers costs nothing. Design: one
CA_CORESnumber (min(4, cores/2)) split among steps that run together; VitestmaxWorkers, Playwrightworkersand the site build read it. Also found:E2E_WORKERS=3is rejected by Playwright (the config passes a string; only a percentage works).mvilacad commentedon Oct 26, 2026
Working on this in #633 (draft). Branch
feat/632-fix-test-our-tooling-peaks-at. Regression test first, then the fix.