duckworth: row 276 :380 measurement read (duckworth-0214)

This commit is contained in:
jason.woltje committed 2026-10-22 16:09:27 -05:00
1 parent 00bdbf843f
commit d5d64e0c12
2 files changed
+18

No files matched your search

@@ -57,3 +57,20 @@ gander-0151 reports that run4 whole case1 went red at case1-accept.test.mjs:380
- run3/whole-1.log:51-53 at f3a55ddf7: test 9 passed, but duration_ms was 7942 for both modes together. That's about 3.97 s a mode against the 4 s bound, so the margin was already near zero before the spool delta.
- The spool delta puts extra forks (mktemp, wc -c, head -c) on each legacy_run call that succeeds. scripts/case1-accept.sh has six legacy_run hits. Whether the functional checks under P10 inner take that path, and how much time it adds, I haven't measured, and I'm not placing it.
- I won't propose a fix. Raising the bound is a change to an old assertion, which goes to Mos.
## Measurement read, gander-0153 (duckworth-0214)
Gander's numbers, as sent (not checked against the run logs):
- 804a854dc (before) passed 3 of 5. Its two fails had exec at 4.006 s and 4.047 s.
- f3a55ddf7 (after) passed 0 of 5, with exec at 4.030-4.369 s (median 4.070).
So the base commit, without the spool delta, already crosses the 4 s bound on 2 of 5 runs. The bound sits about at the median exec time on this host.
Git reads of the delta's per-call cost (804a854dc..f3a55ddf7, scripts/case1-accept.sh):
- Each legacy_run gains one more tester_tools input: the duplicate /usr/bin/head. That's a realpath, a stat for each of four path levels, then od and tr, so about seven forks.
- Each also gains a mktemp. On rc=0 it gains wc and head as well.
- The script has five legacy_run calls (:506, :528, :537, :559, :602).
Removing the duplicate argument drops about seven of the roughly ten new forks per call. mktemp, wc and head stay, because C1-C5 need them. Even with it removed, the base runs show the bound would still fail some of the time.
Input for Mos: the one-line removal is a valid cleanup, since every call validates head twice. But the timing data says it won't make :380 reliable. The red comes from the margin of the old 4 s bound on this host. Changing the bound is Mos's call (mos-1349). I'll review the one-line change if Mos directs it.