registeredAt compared commits, so a catalogue merge that moved only the commit
(2df05030 at 18:02, a85d9037 before it) took every catalogue module in again
and composed it: about 200 take-ins, some 25 minutes, on every mesh-catalog
merge, which held nextcloud for an hour. A module whose trunk fingerprint is
the one it is registered at now has the new commit recorded (built_from and
source_head, as a registration moves them) and composes nothing unless its
composition or grant is missing. Readers of the registered commit: the
fingerprints read the commit's build or, with none, the indexer's value at it,
which names the same fingerprint; status, modules, the board and the facts
read it as the source's commit, which it is; the upgrade path counts it as
built from that commit, which it is.
The stall watch counted only settles the bus accepted, so each refused
acknowledgement at a first start left work held for good: it now counts the
loop's every attempt. The acknowledgement and refusals at the stop are sent
past the stop, and the connection closes only after the readers and the loop
have ended.
Research 072: the take-ins on a bounded pull consumer.
A refused key stays unacknowledged until its delay ends, so with 8 a consumer
a handful refused for long (a manifest word this controller does not read
yet, a node that keeps failing) stopped that consumer for good, the build that
reads the word among what it held back. Each consumer now leaves up to 1000
unacknowledged; the store's bound is the one loop's batch of 16. A refusal
waits 30 s, doubling each time up to 10 minutes.
The stall watch judged movement by deliveries, which rise with every
redelivery: it now counts the keys this process acknowledged. A consumer that
cannot be made raises take-in-stalled instead of a log line. On stop the
gathered batch and a key in hand are refused with no delay, so the next
controller has them at once. A trunk value with no fingerprint of the new form
is said once, as every wait for a write.
Research 072: the take-ins on a bounded pull consumer.
The take-in watched the builder's and the indexer's buckets, started a follower
per module that retried on a 30 s timer, and at start took every catalogued
module in alphabetically: on 2026-10-25 a restart took in about 2 modules a
minute, about 190 followers queued behind 16 slots, and every take-in composed
two or three times even with nothing new, so a new build waited 30 minutes and
more for its credential.
Now two durable consumers named mesh-controller-take-in read the indexer's
modules bucket (every module's last trunk value when the consumer is made, then
each write) and the builder's builds bucket (each write), at most 16 keys
unacknowledged each. The bus keeps what is not taken in: the start's look is the
consumers' own catch-up, and a restart costs only what was left unacknowledged.
A batch reads the catalogue once, takes each module in once, and asks the
composer twice: once for its modules' nodes, once for the bus's node with the
grants. A module registered at its trunk whose compositions and grant exist, or
that this process took in at the same trunk, composes nothing; a key delivered
again is taken in whole. A wait for the builder's or indexer's next write is
acknowledged, a wait on a read or a failure refused with a 30 s delay. The
backlog is in status, and a backlog standing still for 5 minutes raises
take-in-stalled. The controller's account is granted the two consumers' acks.
The follower tests go with the followers: twelve tests become three.
Research 072: the take-ins on a bounded pull consumer (milestone 4, done keys
alone) and the start's look composing only what changed (milestone 2).
mesh/check pass 8e02c4726eb4 on c3db00ceac25 | unit 981 run, 171 kept, 2 skipped | integration 1641 run, 232 kept, 7 skipped | over bound: integration…
A success cleared the condition only when the same follower had raised it, so
one raised before a restart or by another follower stayed open after the take-in
worked, and the operator read that a build needed a look. Clearing a condition
that is not open costs one read. On an error return, serve now ends its
goroutines before the keeper closes and before handActConn and the bus are put
back (the review of fix 12).
inBackground ran the take-ins of done keys (now in #463's slots), the start's
look, the given values' replacement and the announced reads beside serve without
serve waiting for them: a take-in could outlive its controller as the loops did.
It now takes the context and is counted in its serving group as well, and a
serve that is stopping starts none.
Thirty urgent fingerprint-not-composed conditions were live on 2026-10-25, each a
done key's take-in a stop or a restart cut with "context canceled". A
cancellation is no fault of the build, the chain or the bucket: a take-in or a
layout whose controller stops returns and the next one does it; one whose own
step was cancelled is tried again without a condition; a send cut by a
cancellation counts as no failure to put its files.
The put test now starts and leaves the node's failure count at nothing, so a
second run of it (-count=2) finds no count left by the first.
The try of the on-change mirror failed with every test passed: goroutines a
served test started outlived its serve and met the next test's globals and its
closed stores. Under -race the served tests show it as a race on conditionsFrom
(watchTheMesh writing it while settlingPending, already running, read it), and
a take-in cut by the stop raised fingerprint-not-composed, a kind with no plain
words, which fails the package after its PASS line.
- serve counts every goroutine it and its parts start (serving_group.go) and
ends and waits for them before anything else is undone: the lease's watch,
the pending settler, the status summary, the facts, the mirrors and the new
declared-packages keeper with its two bucket watches, the declaration sweep,
the read-call counter, the watches' closers, the have watches, the vault move
on a report, and the watchdogs, doctor, healers and asker under their own
group.
- The condition store opens before anything that reads conditionsFrom starts;
askerFrom is an atomic pointer, read from the keeper's goroutines.
- A done key's take-in cut by the stop raises nothing, and fingerprint-not-composed
has plain words.
At 16:07:36 the builder rewrote about 120 done keys at once. Each woke a
take-in of its own with no limit, and each take-in read the rest of the
mesh through restOfTheMeshOf, which opened a licences and an identity
pool per call. novox's store ran out of connections at 16:08 and refused
everyone, the controller's retries keeping it full.
- The serving controller lends its stores by inventory (shareStores);
restOfTheMeshOf reads on them instead of opening pools per call, and a
verb run as its own process still opens its own. stores opens each
store once under a lock, since readers now share it. serve uses the
shared identity store instead of a second pool.
- The watch's take-ins and the start's look share one cap of 16 take-ins
at once (takeInSlots); a take-in waiting for its next try holds none.
- store.Ready backs off on a full store (SQLSTATE 53300), doubling its
wait up to 5 s, and says once that the store is full.
Issues 1945 and 1277.
With no send left nothing else writes a missed grant, so the take-in of a done key returns it and tries again every
half minute under an urgent condition until it is written. plan-moved and sent go with the plans. The controller's
first-start deletions are said on a granted subject, a node hears only its backup holder, and the operator's account
writes the hand path, an object replaced included (ADR 0539 clauses 2, 7, 10 and 13).
With no plan left to take a done key in, a write of the builder's keys registers the new fingerprint, composes the
module's nodes and the bus's user list, and writes its grant, so the assigner's rules can name it (ADR 0539 clauses
5 and 7). The new controller composes every node at its start, and the controller it replaced at its own
fingerprint, the operator's way back (clause 10); the merge window's, silent sends' and building plans' buckets are
no longer asserted and are deleted once, each said as the event bucket-deleted (clause 13). The assigner may read
the compositions' digests, and a node-engine hears its backup holder's backed-up.
jschoubben