The object store reads an object's metadata and purges its old chunks when it replaces one, so the hand path is
granted the direct and message gets and the purge of each node's object stream; a test on a bus running the composed
user list walks the hand path end to end.
The new controller says its first start's deletions as an event, a node-engine hears only its backup holder, and the
way back right after the switch, a want and a composition written at the terminal, runs on the user list today's
controller composes (ADR 0539 clauses 2, 10 and 13). The operator is the administrator person, the one second
writer ADR 0539 names.
The switch to release 6 is today's controller's last plan, and the bus runs the user list it composes until the new
controller writes one: so the new controller's writes of each node's compositions and their digests, each new
node-engine's reads of its own compositions and wants and its backup holder's verbs and word, and the assigner's read
of the digests must be granted by this controller already (ADR 0539 clauses 2, 6, 9 and 16; the review's A1).
mesh/delivery-group group fix/1937-done-key-blobs rejected: composed together: the checker's verdict at the members' heads (dbd9c3771cd8): integration: github.…
chainBuildByID answered a trimmed build without SourceFingerprint since builds moved to the builds bucket (issue 1090), so
currentFingerprints read every fingerprint empty and grantsFrom skipped every module: mesh-controller_grants has never held
a key, and TestTheGrantOfAnIssuedCredentialIsOnTheBus failed on main.
mesh/delivery-group group fix/1937-done-key-blobs rejected: composed together: the checker's verdict at the members' heads (c20fdcb8c623): integration: github.…
A done key the controller never took in holds no manifest, and the store's nightly collector deleted its bare blob:
node-tools', mesh-host's and the forge bridge's keys name blobs that answer 404. Since issue 1933 the take-in refuses such a
key, and nothing built it again. held (in the serving process, HEADs only) lets the builder ask before it reuses a key; a
refused take-in now names every gone archive and says artifacts-gone with the module, key, fingerprint and references.
A sweep read what it keeps before what was ever recorded, so a take-in's
build recorded between the two reads, by a concurrent take-in off the
plans' hold, was eligible and let go of seconds after the take-in read
it; a new module's build is unkept until it is registered. Now the sweep
reads the records first and keeps what a build recorded within the hour
or an open plan names, and names every reference it lets go of; a take-in
holds every artifact in the store before recording, failing on one the
store lacks; and RecordBuild clears artifact_collected as RecordMirrored
does.
mesh/check failed: g14: novox/mesh-catalog#464: checks/health/go.mod names mesh-sdk a9560890ee26, which is neither on mesh-sdk's trunk nor the group's…
The catalogue's manifests moved their content into template files and fall back to predefined keys, so a copy of
module.json alone no longer passes the catalogue's own check; the bus's Dockerfile is gone from the catalogue.
The controller's readable states, its bucket lists and the composed golden file keep lane C's removal of the
checks-running and plans-building buckets and the build agent's check subjects, and add lane B's grants bucket.
The controller asks no build of anybody (ADR 0501 §5, ADR 0525 clause 6, ADR 0503 §3). Every module of a walk is taken from the
builder's done key; a key not done yet waits, named in the plan, and never fails it.
- C2b/C2c: internal/builder and cmd/mesh-controller's asks go (askABuild, onlyTheBuildersOwn, buildOneAsked, takeIn from an
outcome, buildAgentsOwn, the builder's say and its fallback, the daemon hearing an outcome). The server follows the retired
seats' built subjects only to let go of them, because the installer's carried user list names them.
- C2d: replay, queue, cancel, clear, kill, pause, resume, delivery-check, check-here and merge-gate are refused by name.
- C3: the check queues (CHECKS_READY, CHECKS_OTHER) and their workers, the checks-running and plans-building buckets and
their grants, the build queue and its hold (plans_building.go, queue.go, link/holds.go, builds_nats.go, checks_nats.go).
- C4: merge_gate*.go, the pull request's repo-check path, the Checked event and its layers; a pull request asks the
controller nothing but to tell mesh-config what its head declares.
- C5: the builds bucket is written only by key_takein.
- C10 (help and refusals): the terminal's help lists only what is served.
Integration tests of the plan's tiers, sends, stops and retries now run on a stand-in builder with a done key for every
module (walk_keys_scaffold_integration_test.go); the tests of the asks, the queue, the merge gate and the replays are deleted.
mesh/merge-gate head 185d0294ebd4 merged into trunk 89abce5df955 — pass: builds build-agent, mesh-controller → ace, g14, novox, shanks; no bus step; the ch…
mesh/delivery covered after its stop: a later merge that contains it was delivered: novox/mesh-controller@0b312218f90f (merged as 791786d5 into main, wal…
mesh/check pass 185d0294ebd4 on 89abce5df955 | unit 1756 run, 0 kept, 4 skipped | integration 3034 run, 0 kept, 16 skipped | over bound: integration (…
The checker's integration store answers on a unix socket named in the query of MESH_TEST_POSTGRES
(?host=<socket dir>&sslmode=disable). The identity, inventory and licences fixtures cut the address at
its last slash and appended ?sslmode=disable, dropping host=, so pgx dialled its default socket where no
server is: every store test waited out its 20 s readiness bound, each package ran into go test's
10-minute timeout, and the verdict named only the first failure,
identity.TestACertificateChainsToTheMeshsOwnAuthority.
store.DatabaseForTest now builds that address for every fixture, keeping the query as the template raise
already did; identity's fresh is ForTest rather than a second copy. The readiness bound and its error,
which wraps the driver's cause, are unchanged.
R4e, resumed at tier 1 after a stop of five hours, named ace and shanks silent within seconds of tier 1's send: the
walk judged each machine on every send it had made it, tier 0's of 00:23 too, held against tier 1's templates, and
measured the bound from it. A send carries the whole declaration, so the last one is what a machine is judged on.
The resume kept the tier's entry of before the stop, so the plan read waiting for 5h48m. And plans retry refused a
walk failed only by the gate's silence, though its machines had since applied.
R4e's own delivery plan waited for ever in tier 1: the controller took every
module but build-agent's own from a referenced builder key, and only R4e's
builder and build modules, delivered by that same plan, write references.
The builder now says so in its bucket of instances; until it does, a walk asks
build-agent as before R4e, and an open walk waiting on a key the running
builder cannot reference is judged again at the next look.
ADR 0520 narrowed after its review: mesh-tools-go builds from its own
directory on a pinned base and has its key from mesh-build-image, and
mesh-tools-python is no module, so the set is named, not derived from
the toolchain bases.
The builder writes no key of build-agent or of the toolchain bases its
builds stand on, so a walk moving one would hold its tier for ever.
Until R4g they are build-agent's own, as the builder's module is; the
bases are read from the toolchains, not copied.
A build or check failed now and then on 'Could not resolve host:
novox.internal' between clones that resolved the same name, and its
whole delivery failed with it. A clone whose lookup failed reached
nothing, so it is cloned again with a growing pause for up to 60 s,
each retry said in the build's log as the resolver not answering;
past the bound it fails saying so. Any other git failure fails at
once, as before. Covers a module's own clone, an artifact's context
and a check's repositories.
jschoubben
mesh-tools