hq issue 335: a build that passed its gate was failed and put back when the
first machine's later reports went quiet. Proved: fails on 7f65e62, passes
on becd40b.
hq issue 331: every declared tool check read no link to the bus is open while the
node-engine's link was up. Proved: fails on 824911d^1, passes on 824911d.
hq issue 325: assign said no module of that name a minute after the merge
that asked for the module's build. Proved: fails on a759ac6, passes on b93dea8.
It dialled the old broker's guest account, which the builder no longer reads and the mesh no longer
runs. And the walk stops stocking a packet-filter runtime no container of the module names.
The tool runtime is a Go binary that launches each bundle through the launcher the builder writes; the
repository root has no package.json any more, so the old script failed at its first step.
The tests set MESH_LAB_INCUS in their body, which runs after the client module has read it, so they
asked the real incus: green where none is installed, red on the workstation that runs the lab.
A catalogue module is a Go module of its own, so the prover runs a replay
in the directory its register entry names. R305 fails on the commit before
the fix and passes on it.
Two machines: the anchor raises the foundation, joins over its own
loopback and becomes the hub; the joiner makes its tunnel key, is issued
a token for it and enrols with the bus's port closed to its own address,
so the enrolment can arrive only over the tunnel (novox/hq ADR 0169). The
check that had been added to the two-node walk moves here, closing the
port the bundle publishes the bus on rather than the bus's own.
And the uplink's range may be named: behind a VPN client that routes
every private range, incus had none left to pick and no scenario could
be raised.
The filter module now serves its verbs from a runtime the mesh builds
(novox/hq ADR 0170), and a bed registered its raw manifest, which the
mesh refuses as unbuilt. The bed stocks mesh-runtime-nftables and the
filter helper registers the module through the stocked image.
A token carries the bus's address, and at genesis that is the anchor's
loopback, which no other machine reaches. The anchor joins locally and
becomes the hub; the laptop makes its tunnel key, is issued a token for
it and joins over the tunnel (novox/hq ADR 0169, issue 146).
This bed raises genesis by hand, so it places the composed user list
after each token and each enrolment, as the trust bed does (novox/hq
issue 146); without it the first join was refused.
A third machine makes its tunnel key, is issued a token for it, and
enrols while the anchor drops its packets to the bus at the first hook;
it can only have arrived over the tunnel (novox/hq ADR 0169).
The lab's check ran in the TypeScript toolchain, which holds no Go compiler,
so the replays register was never compiled before a merge. merge-check.sh now
declares replays/merge-check.sh as its Go part, which the build seat runs in
the Go toolchain: gofmt, go vet and the register's own tests. Both new
replays are proved: each fails on the commit before its fix and passes on it.
Two of nineteen image checks read unhealthy while working in the mesh's hands;
adopted without proof they would have put back two good builds. The replays
every catalogue merge check runs now start each long-running container whose
declared check or image the change touches, alone, with what its module
declares and nothing of the mesh's, and require the check to see the program
within its grace: a program that stays up while its check does not see it
fails the change; one that does not stay up alone is said and left to the
first machine's gate; a check needing a provider must only reach the program.
R-studio replays the studio's false unhealthy — a server bound to HOSTNAME's
address and an image check asking localhost — failed on the bed, and proved
with HOSTNAME=0.0.0.0.
For eleven hours a web application's port was open and its program ran while
every request hung; liveness and a TCP check both say it fine. R145 raises a
web server whose application never answers, has the node-engine at its commit
look at it with the module's declared HTTP check, and the controller at its
commit raise the module's condition on the second look. Proved: it fails on
the trunk before Phase B and passes on it.
The agent server crash-looped about a hundred times behind every passing check.
R-crashloop raises a container whose program exits at start, has the node-engine
at its commit judge it through the runtime, and the controller at its commit
judge the gate from what the engine said — two repositories, each at its own
commit before the fix and on it.
A merge-check.sh, the repository's layer of the mesh's merge check (mesh/repo-check), in
the TypeScript toolchain: install from the lock file, type-check, the unit suite. The
integration suite and the replays need a lab and the container runtime, which unapproved
code is not given: said as not run, never passed silently.
Phase 5 is done when the replays of 236, 262, 263 and 266 fail on the commit before their fix and
pass after. The register names every replay with its issue and fix; the bus replay (266) runs a
consumer filtered like the controller's against a bus of a given release, the resolver replay (262)
renders the catalogue's machine list and asks every machine's name by getaddrinfo under musl and
glibc, and the prover runs each at both commits: all five (236, 262, 263, 266, 273) proved.
jochen
mesh-admin
jschoubben