Review round 1 found the fingerprint guard of the build read unproven: every
entry of the 100-key listing carried a time. The listing is now also read with
entries that say none, and no history may be read.
Builds: design 47 H2 H-A-487 (review round 1 finding)
builder_build read every key's whole history to learn when its entry was
written, one more round trip per key of the module; on nats's hundreds of keys
the read crawled. Each entry's own write time now comes with the key's read,
the history is read only for the one key a fingerprint names when its entry
says no time, and a listing cut at the 20 newest says so in its note.
Builds: design 47 H2 H-A-487
Design 51 copies an upstream image by its pinned digest, and distribution's
mirror says published with no version, so nats, hello-web, mailu and showcase
could never be released. The reservation keeps its upstream refs (Upstream,
filled at the reserve from the plan's upstream tasks) and Missing never awaits
them; the indexer releases a version once every other artifact is published,
naming each upstream artifact with the pin its manifest gives, and refuses the
release loudly when an upstream artifact names no pin.
Builds: design 47 H2 H-A-486, H-A-486 (addendum)
A version-released whose tag the seat git's holder refused was handed back at once,
spent its five deliveries in seconds and was dropped by the bus: its reservation stood
and its publish item asked again every 30 s for ever. At the start the builder now holds
every built, unreleased reservation against the catalog's verb versions and takes a
released one in as its version-released would (tag, release, bound and item ended,
findings cleared); a refused tag is handed back with the publish pause. The tag check
itself is the forge module's, already accepting a nested package path on the catalog's
trunk.
Builds: design 47 H2 H-A-484, H-A-484 addendum
With one consumer per holder an item has no class to move ahead of, so
builder_publish_first, its mode and its act go. The verb publish-queue
lists the items in stream order, the order the one consumer gives them,
and its description says so. A holder is a module name, so one ending in
-live or -again (an earlier release's class durable) is refused at the
learn, said, instead of having another holder's durable deleted.
Builds: design 47 H2 H-A-483 third addendum, §793
The two class consumers (live, again) could not order live before again
under every scheduling: the idle-dispatch row failed on the build node
after four review rounds. One consumer per learnt holder (publishes-<holder>)
gives items in order; a failed send is handed back with its pause, so an
item put meanwhile goes first by construction. The builder retires the
holder's class durables of the earlier release before making the one,
since a work queue refuses overlapping filters; their unacknowledged items
reach the one consumer. The gap read, the held again items and the live
in-flight check go.
Builds: design 47 H2 H-A-483, its addendum and second addendum, §793
A live item put after the turn's live fetch and before the again items were held
was not yet delivered at the gap read, so no gap showed and the later again item
went first (review round 3 on c37a7317, publishitem.go:792-809). With again items
held the dispatch now fetches live without waiting once more: what it returns runs
and the again items stay held; only an empty fetch leads to the gap read. The test's
idle-dispatch case is a subtest of its own, so it runs alone.
Builds: design 47 H2 §793, H-A-473 (third addendum)
On the turn after an again item ran, the dispatch fetched again and read the live gap
without fetching live first, so a live item the stream held but had not yet given to
the waiting live pull showed no gap and a later again item went first (review round 2
on b77b70a6, publishitem.go:791). The live consumer is now fetched without waiting on
every turn no live item is held. The test row adds that case and bounds its rounds by
the level's set-up share.
Builds: design 47 H2 §793, H-A-473 (third addendum)
The idle dispatch keeps a waiting pull on each consumer, the bus's only wake for an
item put or delivered again. A live item the waiting pull was given but the dispatch
had not yet received let an again item go first (review finding on 1f33a4c0,
publishitem.go:700, 761, 797). An again item taken while the live pull waits now runs
only once the live consumer's delivered sequence is no further than the live items
received; a gap holds it until the live item arrives and runs.
Builds: design 47 H2 §793, H-A-473 (third addendum)
The test ran the builder at a 1 s step bound throughout, so a loaded build
node could fail a build that must end done. Those builds now run under
testbus.SetupBound; the hung step's part starts the builder again at 1 s,
and the build again after it under the set-up bound.
Builds: design 47 H2 §793; H-A-473 (addendum)
The dispatch read the live consumer's pending count against items the
pull's callback had not yet handed it, so at one processor an again item
was sent before a live one already delivered. It now fetches the live
consumer without waiting and pulls the again consumer only when that
returns nothing; an item is received when its fetch returns it.
Builds: design 47 H2 §793; H-A-473
The builder bound the seat mesh-builds' worker once, under a sync.Once, and stopped hearing the bus's advisories after
that, so a worker deleted and made again by the grants part was never pulled again and the asks piled up. The
advisories CREATED and DELETED are now heard for as long as the builder runs: a deletion lets the worker go and says
so, and a creation binds it again. The step test deletes the worker, makes it again and has an ask ended through it.
Builds: design 47 H2 H-A-468 (2)
The publish queue's whole test reserves m4 and waits on the seat
mesh-builds' version-reserved by its subject, said as the process says it
(sayReservedOn, now one function main and the test share); without the
event the step fails. Also gofmt of the pipeline's order_test.go.
Builds: design 47 H2 §821, §805; review round 2 finding 2
A process account is granted only what its bus names, never its seat's
work, so a builder on an account of its own could build nothing at its
first start. The process declares no bus and takes the module's broker
credential, as before B13; as the module's one process it is the one
reader of the work queue publishes. Class, caps, settings and the
publish-queue removal stay. The validator built at this head accepts the
builder in the change case (the trunk case refuses it: no class, uncapped
states).
Builds: design 47 H2 §793, §820; H-A-464 (corrects H-A-454 second addendum); review round 2
The review found §794's two bounds fixed in code: the builder's 100 ms
for saying a slow event's handling, and builder_health's five minutes
for naming a consumer behind. Both become settings read at run time,
slow-event (100ms) and events-behind-bound (5m), declared in the
manifest's config with those defaults; a value that is no duration
refuses the builder's start, and makes the health say it.
Builds: design 47 H2 §794, H-A-461 (2)
The review found §794's backlog and §796's tag at another commit proven
only by unit tests. Both become steps of the builder's whole test on the
embedded bus: a reservation skips 0.2.0 whose tag stands at another
commit, takes 0.2.1 with the finding version-tagged-elsewhere, the
release writes modules/m4/v0.2.1 at the build's commit and the
version-released heard from the bus clears the finding; a message the
handler asks again is kept on its consumer's ledger in the state
consumers on the bus, and the health held six minutes names it.
The health judgement moves from the tools program's main into
internal/versions (BuilderHealth), which the tools program calls, so the
whole test can prove it. The test's comment no longer says the old state
is dropped through the bus module (F1).
Builds: design 47 H2 §794, §796, H-A-453 (3), H-A-435, H-A-462
The bus module's keeper made no entry for the builder's old-section
publish-queue, so nats_drop refused it at every start and the move
re-put its items after the 24 h window. The builder, which made the
bucket, now deletes it itself once every item is on the work queue;
the manifest no longer names publish-queue nor invokes nats.nats_drop.
The whole test proves the move and the removal on the embedded bus,
and a second move of the same items after a stopped start puts each
once. Also corrects the publishes comment: the process consumes the
work queue on its own account (H-A-454 second addendum).
Builds: design 47 H2 §793, H-A-461 (1), H-A-454 (second addendum)
The tools bundle follows the seat-holders view of the claimant of
mesh-module-assigner, names each build seat's holder per node from it, asks
each holder's current at its node one after another and answers them
together, a holder that does not answer listed with why; it reads no presence
and chooses no node. The manifest lists the tool and the three build seats'
current among the module's invokes, outside the process's bus.
Builds: design 47 H2 §801; H-A-446 (1)
The manifest says class core (it claims a seat), caps every state by the fill
read on the bus (builds 128 MiB, findings and versions 32, bounds 16, the rest
8; H-A-437), keeps the findings history at 8 MiB for 7 days, invokes the bus
module's nats_drop for the moved bucket, and bounds publish-tries at 10 with 5
as its default. The process gains a bus of its own naming the work queue
publishes in queues with the module's emits, consumes, state and invokes, its
credential broker-builder, so the validator accepts the work-queue stream read
by exactly one process. Verdict: the judge at feat/47-H2-i ce6c39e4 accepts the
builder in the change case and refuses it in the trunk case (no class,
uncapped states).
Builds: design 47 H2 §783, §812, §814, §820; H-A-394 (1), H-A-437, H-A-446 (2), H-A-454 second and third addenda
The work queue publishes is declared in the manifest's old streams section
(retention work-queue, duplicates 24 h, max-deliveries 10, 64 MiB, 7 days), the
module's own stream consumed on the module's account, so the bus binding asked
of the broker seat goes; the consumers per holder and class are still made at
run time. The seat mesh-builds gains the read verb publish-queue, served by the
builder from its own stream (counts per holder and class, oldest and newest, a
list capped at 50, capped past 4,000 items read), and builder_publish_queue asks
it in place of the runtime's stream and work-queue reads. Two whole tests of a
gone artifact are folded into one, so the builder's integration count stays at
18 with the publishes whole test.
Builds: design 47 H2 §793, §819, §820; H-A-440 (1), H-A-446 (2), H-A-454 and its addendum
The bucket publish-queue listed and reread itself before each send; its items
now ride the builder's work queue publishes, pulled per holder and class by
consumers the builder makes as it learns a holder, the live class first, a
failed send handed back with the pause, an item acknowledged when its publish
is heard and terminated when its build is withdrawn, superseded or dropped.
builder_publish_drop and builder_publish_first act through reasks after the
hand-act log; builder_publish_queue reads the queue through the runtime. The
old state's items are moved once at the start, then dropped through the bus
module.
Builds: design 47 H2 §793, §820; H-A-290, H-A-291, H-A-440, H-A-446 (2), H-A-448
The artifact store's sweep read each done key's history to learn when its build ended.
A done or failed key now carries ended and outcome, written with its end (kept as it was
when the same end is written again), so a reader of the builds reads no key history.
Builds: design 47 H2 §797
A number whose tag stood at another commit was reserved again, and a Go package outside
mesh-sdk's go directory was tagged go/v<version>, which the Go toolchain never reads. The
builder now reads the Go package's tags once per catalog revision, skips a number tagged at
another commit with the finding version-tagged-elsewhere (cleared at the next release), and
tags <directory>/v<version>, the directory read from the package's module path.
Builds: design 47 H2 §796
A tool-only health is refused by the validator (ADR 0227 rule 8, ADR 0240 rule 2), as the
runner's manifest has it: the unit check stays, the tool reads the consumers beside it.
Builds: design 47 H2 §794; H-A-435 second addendum
The indexer re-pinned every importer at each publish; it re-pins once at the reservation
when the builder says version-reserved, with the package's path and the indexer's revision name.
Builds: design 47 H2 §805; H-A-420 (2)
The builder said nothing when its events backed up or one handling stood still; its health
now reads each durable consumer's oldest unacknowledged message from its state consumers,
and an event handled over 100 ms is said with the part that took it.
Builds: design 47 H2 §794, §820; H-A-435 and its addendum
A line per module drowned the journal at 6,500 plans; the plans live on the bus, where
builder_unbuildable reads them, and the builder says them once per start and per seat list change.
Builds: design 47 H2 §795, §820; H-A-425
The wait answered "no verb to read" for every step whose end is kept as its
answer object: the process never set the asker's reader and its manifest
invoked no build seat's answer.
Builds: design 47 H2 §791, H-A-412 (addendum)
The delivery SDK's build-step part keeps a step's end over the steps state's
value as the object answer.<ask>, read through the build seat's verb answer;
OutputObject is gone. The builder's wait names the answer object it could
not read.
Builds: design 47 H2 §775, §791; H-A-412
A caller whose step ended by its bound could not stop the builder working on
its ask, and nobody could see what waited on the seat's work queue. The
builder now serves the seat's verb drop: an ask it runs is cancelled, one not
pulled yet is kept dropped in its state taken (dropped.<ask>, for
ask-taken-bound); both end the step could-not-run naming the drop and who
asked it. builder_asks reads the waiting asks through the runtime and the
taken ones and the drops from the state taken; builder_ask_drop asks the
seat's drop once the hand act is recorded.
Builds: design 47 H2 §777, §778; H-A-401, H-A-404, H-A-410
The line a build step's finish wrote named no ask, so an answer could not be
matched to the ask the runner put. BuildStep now takes the ask's id and its
finish writes "answered <ask> <pipeline> <movers> <step> <outcome>".
Builds: design 47 H2 §792
An ask put while the builder is down waited for nobody: the plain subscription
heard only what came while it ran. The builder now pulls every waiting ask from
the seat's one worker, takes it once by its id in its state taken (for
ask-taken-bound), says ask-taken at the pull, sends in-progress while it builds
and acknowledges once the step's end is said. Its publish to a registry seat is
a kept call carrying the publish's ask id. The manifest declares taken and the
kept-call pace and ask-taken-bound as runtime settings.
Builds: design 47 H2 §775, §777, §820; H-A-381
The build modules' shared steprun parts of §799: again-delays (15s,60s) read
into a Puller's Delay; a fetch's 401/5xx or connection fault and a publish's
no responders or exited bundle are transport faults whose why opens fetch: or
publish:; the assigner call that named a silent holder goes. The builder reads
an exited bundle as no responders and keeps the part in a key's fault words.
Builds: design 47 H2 §799 (H-A-390)
tags answers items {tag, commit}; trees' description says a missing directory
answers not-found; the error kind away joins refused, not-found and failed,
every answer carrying part and words with it; tag's description names the
<directory>/v<version> form. The seat's form rises to 2. The builder's,
bridge's and indexer's readers take the tag's name from the item.
Builds: design 47 H2 §776, §796; H-A-350 DMC-2, DMC-3, H-A-354
The node enroller's and the watchdogs' generated decoders answer the form
they read (mesh-contracts K1); their readers here take it.
Builds: design 47 H2 §780
The seat generator gives every value, event and answer the wire field form
(left out while 1) and a tolerant decoder answering the form it read; each
contract carries its forms and a digest of its fields, and the compare fails
a definition whose fields changed while its form did not. Every delivery seat
is regenerated, and the readers take the form.
Builds: design 47 H2 §780; H-A-352
The runner's ended line was written only for a heard step-ended and its run's
readers; the unanswered-bound end, the end at placing, a supersede, a re-pin,
the approval's end, the start pass and the verbs said their ends silently. It
is now written in Hearer.say, the one place that says step-ended, and at the
starter's two sayers that call stepEvent directly (repin, sayNotAsked). The
builder's answered line sat on four of buildEntry's returns; the ask's answer
is now one line in BuildStep's finish, naming the pipeline, the movers, the
step and the outcome, the per-entry existing/built-now line kept beside it.
Builds: design 47 H H-A-320 (1), H-A-333
A working 36-mover pipeline was read as stalled twice: the runner wrote no
line when a step ended or was asked, and the builder none when it answered a
step ask. The runner now writes "<pipeline> <mover> <step> ended <outcome>
(<node>, <duration>)" at each step end it writes (the asking step, every
reader of its test run, and each reader EndRun ends), and "<pipeline> <mover>
<step> asked on <seat>" for each ask put (the re-ask paths keep their own
line). The builder writes "<module>@<fingerprint> build step answered
<outcome> for <pipeline> (existing build / built now)".
Builds: design 47 H H-A-320 (1)
The verb published answered no time, and after a release no digest,
while the seat form says both come when published. The builder now
keeps when it heard or confirmed each publish on the mark, and each
published digest on a kept release; marks and releases written before
answer without them. A number a take-over withdrew is not published.
The runner drops a pipeline's kept answers at its end.
Builds: design 47 H H-A-331, H-A-328
A package published before a pipeline existed is in no record of the
runner, so the builder, which owns publishes, answers whether an
artifact built at a fingerprint is published, from its reservations
(marked from the registries' record) and its releases.
Builds: design 47 H H-A-328, H-A-326 (1), H-A-274
The review found two faults. The tool paged on through a state of any size, one replay of the bucket per page; it now
reads one page of each of its own states and, when the state holds more, answers that page capped, with the keys read
and the state's size from its info, never asking a second page and never refusing. And the retiring of superseded done
builds keyed the newest reservation and release by module alone, so a standing 0.2 would retire a done 0.1.x; both are
now found per module and series (MAJOR.MINOR).
Builds: design 47 H H-A-292, H-A-293, H-A-291 (1), H-A-290 (5)
Two tests failed under -race on main too: the re-ask state fake's maps were written by a try on the watch's goroutine while
the test read and deleted them, and the operator-ask test read the build store unlocked and waited on the builder while
the build its answer starts could still be counting itself in. The fake's maps are locked and read through it, the store
is read under its lock, and the builder is waited on only once that build has claimed its key.
Builds: design 47 H H-A-290 (the module's tests pass with -race, as its checks ask)
A start asked a publish again for every unreleased done build, older ones whose digests a newer reservation or release
of the same module made dead included, and a live publish queued during a long start pass waited for its end. A done
build a newer reservation or release (the builder's own record) supersedes is now retired at the start and at the pass:
its key dropped, its publish item ended, said in the journal; the start asks only each module's newest. A live enqueue
marks its holder, and a pass about to send a start-class item stops there, lets the sends in flight end and lists again,
so the live one goes next; a start-class enqueue never does.
Builds: design 47 H H-A-291 (1), (2), H-A-290, H-A-275
The tool runner's release waited forty minutes behind old versions: every start asked a publish again for every done
build, and each send relisted and reread the whole queue. A publish whose build ended in this run now goes before any
asked again at a start (the class kept on the key's value, a live key of a build before the start made a start key at
the start); a pass lists a holder's keys once and sends them four at a time; a withdrawn or taken-over build is dropped
with its publish item ended at enqueue, at the pass and at the start; a send that returns unsent says why, and every call
it makes honours its one bound; builder_publish_queue reads its states in pages and answers counts per holder and class,
the oldest and newest queue time, a list capped at fifty and one module's items with their class.
Builds: design 47 H H-A-290 (1)-(5), H-A-269, H-A-271, H-A-275
Live, a published at a version fourteen reservations share cost three seconds:
the fallback by version read every reservation there, and the journal line
read them all again, so the backlog drained at fifty a minute.
The published now hands the reservations it read (each once, a second pass
taking what the first read) to its journal line. The fallback reads the new
index key of the reservations at a version whose build has not recorded every
digest (kept at reserve and where a digest is recorded, written again at the
start), the only ones a published can be kept on; the line counts the rest from
the index by version without reading them. Before that pass the handler says
in progress to the bus.
Builds: design 47 H H-A-280, H-A-271
Every artifact carries its version from its build (a Go module zip's paths, a
program's stamped version, an image's and a bundle's build environment), so
no digest survives a change of number: the move that kept digests would have
published a binary saying 0.1.37 under another number.
The try, on a number a registry holds at another digest, withdraws it on the
reservation (a gap) and raises version-held-elsewhere naming both digests,
then asks the build of the same fingerprint again through the build-again
path of a done key whose version is not released. Its reserve never takes a
withdrawn number again, skips every number a registry holds, and the build
records fresh digests and runs its own publish; the takeover ends the old
number's publish item and bound.
Builds: design 47 H H-A-275, H-A-279, H-A-226
A publish item skipped while the registry held the version would stay skipped
for ever once its published event was lost; and node-tools 0.1.37, held by the
artifact registry at another digest than the build's, was asked again for ever
and never released.
When every missing artifact is held at its version with the built digest the
builder marks it published from the registry's record and says
package-published, as a heard published would; the release follows from the
catalog. Before reserving, the registry seats' versions of the module's
artifacts are read and every number one holds is skipped (version-held-elsewhere,
scope module); a read that fails lets the reserve proceed. A done build whose
number a registry holds at another digest moves to the next free number, its
digests kept, the old number withdrawn as a gap; its old item and bound end and
its publish is asked under the new number. The finding clears at the module's
version-released.
Builds: design 47 H H-A-274, H-A-275, H-A-245, H-A-226
The publish item's try asked the registry to publish node-tools 0.1.37 again
every few minutes while the registry had published it and the builder's own
consumer stood thousands of messages behind, adding events to the backlog.
Before each try the builder reads its durable on the events stream (pending
plus unacknowledged): at a hundred or more the try asks nothing. Below it the
registry seats' own record (their verb versions) is read: every missing
artifact held at the version with its built digest asks nothing either. A
skipped try is said in the journal, written again with its pause and spends
no try; a count or record that cannot be read holds no ask.
Builds: design 47 H H-A-269 (2), H-A-272
The builder's consumer of the registry seats' published stood thousands behind:
each published, and each one that marked nothing, read every reservation.
An index beside the reservations (three-token keys in the same state, which the
reservations' two-token filter never lists) names the reservation keys holding
a built digest and those standing at a version; it is kept at reserve, at a
digest recorded, at release and at drop, and written again once at the start.
A published reads its digest's entry and that reservation; a digest no
reservation holds is judged from its version's entry, never a scan.
The test stores and the tools' read store now read a key filter token by token
as the bus does, which showed the publish queue's resume listing three-token
keys with the two-token filter: it lists them with its own filter now.
Builds: design 47 H H-A-269 (1), H-A-271