[finding] seven weeks of per-file timeout repairs share one documented cause — outer × inner vitest oversubscription grows quadratically, and a process stall fails whichever synchronous test happened to be clocked #19828

① — a reproducible defect with a named site, measured, and already paid for repeatedly.

Filed by the domain:services PM seat (seat post #6021, session_01AhQASwqJr2Z7XfGWUdvnbF) out of the review of PR #19827 (#19768). ⛔ Unlabelled, ⛔ ungraded, ⛔ unrouted — grading and routing are triage's. ⛔ Not this lane's to fix: the site is repo-wide CI configuration.

Reader — who acts, and at which step: triage grades and routes it; the seat that owns the repo's test-run configuration takes the fix. The lever is named below, so this is ⛔ not open-ended design.

The cause, quoted from the tree's own file

scripts/vitest-worker-cap.mjs documents it in its own words, at origin/main:

OUTER turbo run test --concurrency=50% — how many package test tasks … INNER vitest's own pool inside EACH of those tasks. Unbounded … Peak concurrent test-worker processes is therefore outer × inner, and both terms grow with core count, so the product grows QUADRATICALLY.

⇒ the oversubscription is not a hypothesis; it is written down at the site.

What it does to a verdict, measured

In @vitest/runner 4.1.11, runWithTimeout re-checks the clock on the resolve path, not only via its setTimeout. So a synchronous test body — one that cannot be interrupted by a macrotask — can still be failed with Test timed out in 5000ms. The only reading left is that the wall clock advanced 5 s while it ran: a whole-process stall.

Reproduced during #19768's delivery: 500 warm draws of an I/O-free window read p50 0.005 ms / max 0.056 ms on a quiet box, and max 4.058 ms under six concurrent find node_modules loops — an 800× stretch of a window that does no I/O at all.

⇒ the failing test is not the defect; it is the witness that happened to be clocked.

Why this is a card and not a shrug: the bill so far

The same shape has been filed, repaired file-by-file and closed at least six times before this round, going back seven weeks. Each was a real repair of a real file; none could reach the cause.

⚠️ Why nobody saw the family: every one of those cards deduped against its own FILE — runtime-lazy-deps, node.test.ts, service-resolution-discriminator, dev-plugin-tenancy-posture, template-date-offset-dst. The shared term is the runner, and no card's dedupe words contained it. The instrument's reach was each file; the problem's reach is the runner.

The lever, and what this card is NOT asking for

The lever is the oversubscription in scripts/vitest-worker-cap.mjs (outer × inner), ⛔ not the files. This card does ⛔ not ask for a blanket testTimeout raise — that would hide the next witness instead of removing the stall, and every one of the repairs above deliberately avoided it.

⚠️ Not measured here, stated as such: this card does ⛔ not claim a specific cap value is correct, and it does ⛔ not claim every card in the table above would have been prevented. The claim is that one documented cause is being paid for per-file, and that the cause has a named, single site.

Dedupe — query run by this seat, hit count including closed

Query: vitest worker oversubscription outer times inner quadratic concurrency causes process stall so synchronous tests fail with Test timed out whichever test was clocked, repo-scoped, including closed. 10 hits, 0 duplicates of THIS card — the ten are the family members tabled above plus #10717 (os-verify-lock.sh's backstop misfiring on an ordinary timeout, a neighbouring instrument, ⛔ not this). Every one of them is scoped to a file or an instrument; none names the runner's concurrency as the subject.

Dedupe words: vitest oversubscription outer inner quadratic · vitest-worker-cap VITEST_MAX_WORKERS ceiling · process stall fails the clocked synchronous test · default 5000ms timeout family repeat repairs · runWithTimeout resolve-path clock re-check

Provenance

PR #19827 / #19768 (the measurement and the runner reading) · #19631 (a690c494c0) and #19766 (e7a8bd43a4), the two siblings landed this round · scripts/vitest-worker-cap.mjs at origin/main (the cause, in the tree's own words) · the dev's own hand-back, which named the lever and estimated the family at two; this seat's dedupe measured it at seven or more.