Repository navigation
Flaky: markdown.test.ts measures the parser against 1000 ms of wall clock #1901
Description
src/lib/markdown.test.ts ("text no one wrote by hand") guards the parser against quadratic input with a wall-clock check: within(1000, run) asserts the run took under 1000 ms of performance.now(). On a busy CI runner that can fail with no slowdown in the code.
Seen on 2026-10-24 in the Fast checks of #1895, a PR that does not touch the markdown code (run 40105758528, job 120907036514):
(fail) text no one wrote by hand > looks once for what closes a bracket, an inline formula or an underscore [1422.41ms]
error: expect(received).toBeLessThan(expected)
Expected: < 1000
Received: 1108.9343639999997
at within (src/lib/markdown.test.ts:248:39)
at <anonymous> (src/lib/markdown.test.ts:319:5)
The case that failed parses "__a b_ ".repeat(150_000) (about 1 MB). Linear parsing of that input takes a few hundred ms on an idle machine, so the 1000 ms margin is small for a shared runner; the quadratic behaviour these tests guard against took one to four minutes, so a much wider bound (or a bound measured against a linear-size baseline, or counting steps instead of time) would still catch it. There are 13 within(1000, ...) calls in the file (lines 279 to 388).
Not caused by #1895 or the PRs it was opened beside; found while working on #1866.
Activity
devswha commentedon Oct 25, 2026
Seen again on the first CI run of two unrelated PRs (1515 ms and 1517 ms against the 1000 ms budget), passing on rerun: #1930 job 121219542941, #1858 job 121225515154.
devswha commentedon Oct 26, 2026
This test has failed twice today on GitHub-hosted runners (ubuntu-24.04), which are in use while the self-hosted P620 runners are off:
The test is "text no one wrote by hand > looks once for what closes a bracket, an inline formula or an underscore" (
src/lib/markdown.test.ts:248,within(1000, …)). The #1956 failure passed on a re-run. The 1000 ms wall-clock budget is too tight for these runners. An operation count, or a budget relative to a calibration run, would hold on any runner.fix(tests): bound the markdown parser's growth, not its milliseconds (d…
fix(files): keep a large scratch file's length when many wait to be c…