The canary answers "did this runner get the work done" for the CPU and
for memory, and every disk hypothesis on the node-join lane has been
closed on byte counts. A byte count says a transfer happened; it says
nothing about what the transfer cost, and nothing anywhere in this tree
has ever measured how long the storage takes to answer.
The one field already in the report that would say is
delayacct_blkio_ticks, carried verbatim in every per-thread stat line the
QEMU thread capture writes. It reads zero on all 225 threads of both
samples of a green run and of a red one, which is what it reads on any
kernel booted without `delayacct`, so it is silence rather than a
finding.
So a third arm: 4096 single 4 KiB writes under O_DIRECT, issued one at a
time, so the rate divides to operations a second at queue depth one and
its reciprocal is how long one write took to reach the volume. O_DIRECT
is the load-bearing part -- without it the write lands in the page cache
and returns, and the arm would report a healthy volume at any storage
speed, which is the one failure a collector meant to stand alone on a red
run cannot have. The probe file is written outside the report directory,
because an arm the ceiling stops leaves its file behind and a file inside
the report ships in the artifact.
Why the canary rather than a collector of its own: it is the only
instrument that runs on the passing path as well as the failing one and
brackets the wait on both. Twelve of the seventeen captures a red report
carries are failure-path only, so a green report has nothing for them to
be compared against, and a disk reading with no green baseline would
answer in one direction only.
The floor is 300 operations a second, crossing at 13.7s inside the 20s
ceiling, for the reason the memory floor sits where it does: a floor the
ceiling reaches first never fires. 300 a second is 3.3ms for one 4 KiB
write, pathological for any block device rather than merely slow.
This is the third arm and there is no room for a fourth. The phase budget
guard prices the canary from the arm count it reads out of the source, so
this moved the collectors gated ahead of the serial console from 435s to
460s against a 480s budget; a fourth arm would be 485s and would fail
that guard rather than silently overspend.
Verified in the sandbox image itself, on the overlay the QEMU disk images
live on: 4096 direct writes in 60ms, no probe file left behind.
Assisted-By: Claude <noreply@anthropic.com>
Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
Myasnikov Daniil
67600f47ef