Repository navigation
tests: the #266 bench-freq-median black-box test flakes under heavy CPU contention #292
Description
Two independent agent sessions observed the '#266 black-box: tune bench freq median doesn't false-trip the #62 throttle guard' test (tests/run.sh, FIFO-fed scaling_cur_freq) fail once under heavy host CPU contention (several concurrent suite runs on one Mac), passing cleanly on re-run. Likely the 0.02s poll windows / FIFO feed racing the bench sampling loop when starved.
Not yet observed in CI (dedicated runners). If it appears in CI, deflake by making the feeder/sampler handshake fully event-driven instead of sleep-bounded, or fall back to asserting via the deterministic unit-level freq-writer test (same PR #282) and dropping the timing-sensitive black-box half.
🤖 Generated with Claude Code
Activity
test: Linux-gate the two timing-sensitive tune black-box blocks (#292)
VijitSingh97 commentedon Jul 17, 2026
Confirmed in CI: the macOS job on PR #289 failed both timing-sensitive tune black-box blocks in one run (the #277 zero-hashrate test with a SIGPIPE-driven
exit 141ERR-trap line, and this issue's #266 freq-median test with a false throttle flag). Could not reproduce locally in 140 iterations, including 80 under 10x CPU load.Resolution (pushed to PR #289, commit effda8b): both blocks are now Linux-gated with an explicit SKIP line. Rationale: they drive cpufreq/scaling_cur_freq plumbing — a Linux sysfs interface that never executes on a real Mac — so the macOS job was testing the simulated fixture on a slow bash-3.2 runner, not the product. Full coverage remains in the Linux CI job and the kcov container; the deterministic unit-level tests (freq-writer flooring, compute_build_jobs) still run on every platform.
Leaving this open until the Linux jobs prove stable over a few more runs; close after the v1.9.0 release CI is green.
fix: cap build jobs on sub-1GB hosts; quiet the ERR trap on a zero-ha…
VijitSingh97 commentedon Jul 17, 2026
Closing: the Linux-gated resolution shipped in #289 and the two blocks ran green in every Linux CI job since (including the v1.9.0 release PR run — verified the block headers execute, no SKIP). macOS runs the deterministic unit-level halves only, by design.