Repository navigation
0.70.2 T42: determinism holds an ARM64 runner waiting for an X64 artifact that cannot come #4747
Description
Measured 2026-10-04, read-only: all 6 determinism reds (#4623 #4677 #4705 #4707 #4716 #4724) are X64 artifact-wait timeouts, with 0 hash mismatches. ARM64 was built every time, and each wait held an ARM64 clean-room runner for about 3665 s. Cost: about 6 runner-hours of the two-runner ARM64 pool, and 6 false reds that each need a full re-run.
Fix (signed by the operator): determinism starts only after x86-main concludes, so its wait for the X64 raster drops from 3600 s to 0. A late or missing X64 raster is RED, never a skip, where it used to wait.
Status: built; review passed. In the CI lane after T43.
K^ 3. Owner aprender-a7k. Refs #4678.
Activity
fix(ci): determinism-compare downloads exactly the two host receipts …
chore(pr-review): receipt for #4755 at c2da23b (T42)
Merge origin/main into build-kaizen/t42-determinism-wait
fix(ci): determinism needs both mg-reuse and x86-main after the T43 m…
chore(pr-review): receipt for #4755 at 3ca7da1 (T42)
chore(pr-review): #4755 receipt findings use the S4.2 driver names an…
noahgift commentedon Oct 6, 2026
Closing: the bar is met on main since #4755 (acd5031, merged 2026-10-05 04:44Z).
Read-only sample: the last 40
ci.ymlruns after the merge.determinismjobs ran, and all 23 are green. On the ARM64 runner each took 67–76 s (median 70 s). Before the fix, each X64 artifact wait held that runner for about 3665 s.determinismjobs were cancelled before a runner picked them up (0 s), so they held no ARM64 runner.x86-maincompleted at 08:12:01Z anddeterminismstarted at 08:12:04Z.x86-mainfailed anddeterminismwas still green, which is correct. In the two checked (37414554056, 37406965027), the X64 raster had been uploaded early (determinism-X64at 05:06:37Z and 03:13:28Z), and only the later "wait for every one to finish" step failed.check_determinism_wait.sh --selftest23/23, mutants 10/10) and byif: ${{ !cancelled() }}.