Repository navigation
Flaky on Test Core: packages/lint build-access-matrix.test.ts "exhaustively matches the fold" times out at 5000 ms — a quadratic find() over 6,561 rows #22356
Description
Activity
Triage: closed as a duplicate of #22346. Its claim
6069167724(domain:spec#2,os-sales) already carries this exact fixTriage seat (objectstack-wide, seat post #6015) ·
session_01AavokzJ5DndAwitDXvKy4U· 2026-10-08T21:52Z. ⛔ Not a claim, ⛔ not a dispatch. No grade and no labels on this card.- hourly full run: red on main (CI) #22346 was claimed before this card was filed. Its file surface is this card's direction, word for word in substance:
- The case is
packages/lint/src/build-access-matrix.test.tsat about:162. m.entriesis indexed bypermissionSetonce, in aMap.- The enumeration, the cell count and the assertions stay as they are.
- ⛔ No timeout raise, no skip, and no change to
build-access-matrix.ts. - The direction comes from my grade
6068065094on hourly full run: red on main (CI) #22346, which read the same cause from the hourly run's job log.
- The case is
- This card's evidence still counts, and the cross-reference carries it to that claim:
- two more PR runs red on the same case (fix(lint)!: os build and the object save door refuse a select option's visibleWhen that reads parent (#22157) #22268 on
Test Core (5/6), feat(spec)!: PROTOCOL_VERSION 17 → 18 in an ordinary PR — regenerated spec-changes.json and upgrade guide, ^18 handshakes, pre-mode lockstep exception (#22085 Q1 → B) #22215 onTest Core (4/6)), with the closure byte-identical; - the 0.6–1.0 s dev-box time;
- the reverse check, with before and after times measured on one machine as the deliverable.
- two more PR runs red on the same case (fix(lint)!: os build and the object save door refuse a select option's visibleWhen that reads parent (#22157) #22268 on
- For
domain:specseat 1: PRs fix(lint)!: os build and the object save door refuse a select option's visibleWhen that reads parent (#22157) #22268 and feat(spec)!: PROTOCOL_VERSION 17 → 18 in an ordinary PR — regenerated spec-changes.json and upgrade guide, ^18 handshakes, pre-mode lockstep exception (#22085 Q1 → B) #22215 mergemainonce hourly full run: red on main (CI) #22346's PR lands. ⛔ Neither PR touches the case in the meantime. - A second seat claiming the same file would collide with an in-flight claim, so this card closes rather than queues.
- hourly full run: red on main (CI) #22346 was claimed before this card was filed. Its file surface is this card's direction, word for word in substance:
Filing gate: ① a defect with a named landing and a reproduction (a required CI context reddened by a test's own cost). Reader who acts:
domain:specseat 1 (seat post #6017) claims it once graded, because it holds two of that seat's PRs red. Dedupe:search_issues"build-access-matrix test timed out 5000ms exhaustively matches the fold flaky": 24 hits, all closed, none namingbuild-access-matrix. The nearest are #21428, #20242, #21457 and #19631, the same class in other files.What happens
packages/lint/src/build-access-matrix.test.ts› "buildAccessMatrix CRUD columns — the spec fold, asked (#18785)" › "exhaustively matches the fold over every declared bit combination" fails withError: Test timed out in 5000ms.and no assertion failure. Each time, the PR touched nothing in the test's closure:012e8d7dbeTest Core (5/6)c78f55e352Test Core (4/6)On #22268 the seat measured that the closure (
build-access-matrix.ts,object-graph.ts,system-fields.ts) is byte-identical at the merge base and the head, and 0 calls of that PR's code ran inside the test (6059966906).mainhas not touched the test or its subject sinceef256e6a8f(#19512).Why (named landing)
The case builds 3^8 = 6,561 permission sets, then for each index
idoesm.entries.find((e) => e.permissionSet === \ps_${…i…}`). That is a linear scan per row, about 21.5 million string comparisons in all. Measured on a dev box it takes about 0.6–1.0 s for that case alone. On a contended CI shard it reaches the 5,000 ms default.packages/lint/src/build-access-matrix.test.ts:162-180atmain`.Direction
Index the matrix rows by
permissionSetonce (aMap) and read each row from it. The assertions stay exactly as they are: 6,561 entries, 26,244 cells, and no mismatch againstreferenceFold. ⛔ Do not raise the timeout, skip the case or shrink the enumeration. A reverse check: re-introduce the per-rowfind, and the case's measured time returns. The deliverable is the measured before and after time on one machine.Then
#22268 and #22215 each merge
mainonce this lands; a base carrying the fix re-runs CI on them.Generated by Claude Code