77f3a4cea7514f89b1f11873a487d0e88b0abc34
111
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 jschoubben
|
77f3a4cea7 |
Consolidate: 65 decision records to 23
Every remaining cluster merged. Each was one design that had been split across
several records because it was worked out over days rather than at once.
the node host 8 -> 1 applies not decides, depends on nothing,
per operating system, root service, the
launcher, episodic, what a declaration is,
actions from the bundle only
a node and how it joins 4 -> 1 what a node is, joining, the link as
security boundary, the enrolment token
modules and the graph 7 -> 1 everything is a module, no domain modules,
three edges, provisioning, the core library
substrate and control 6 -> 1 the test, seven contexts, one control plane,
plane the authority is not a database, the named
products, the pinned bundle
connectivity 3 -> 1 a route is a grant, reachability declared,
filter rules
delivery 5 -> 1 reconciliation not a pipeline, artifacts,
the three silos, a failed step, the verdict
the lab 5 -> 1 (earlier)
how this repository 10 -> 1 (earlier)
works
Nothing was dropped. Each consolidated record carries the reasoning of the ones
it absorbs -- the measurements, the incidents, the alternatives rejected --
because that reasoning is the only reason to keep a record at all. What is gone
is the fragmentation: eight files to read to understand tier 0, when tier 0 is
one component.
The four superseded records went too. They existed to point at their
successors, and the successors now contain what they said.
The checker made this safe. Each merge left dangling links -- 38 files after
the host merge alone -- and it named every one. Nothing was found by reading,
and a manual pass would certainly have missed some, including references inside
AGENTS.md which every session loads.
|
||
|
jschoubben
|
5e83ac2c22 |
Consolidate: 65 decision records to 52
Jochen: a normal application has 3-5 ADRs, maybe 10 for a large one, and we are at 65. Fair, and the cause is mine -- I recorded every FINDING as a decision rather than every fork in the road. Two merges, both cases where one decision had been split across many records because it was taken over several days rather than at once. 0019 absorbs ten records about how this repository works: what it is and that it is public, the folder flow, the two design layers, the issue front door, status in frontmatter, playbooks, the naming rule, the product name. Those were never ten decisions -- they were one, seen from ten angles as the repository took shape. 0016 absorbs the five about the lab: a node is a virtual machine, a router is scenery, a scenario declares the underlay, a scenario is a closed address space, and the two scenario classes. Same pattern -- one design, split by the order it was worked out in. The consolidated 0019 also raises the bar for what earns a record, since that is what produced 65: a record is warranted when there is a genuine fork -- a direction reversed, an alternative that will be proposed again, something contested. A finding is not a decision, and a bug is certainly not. Everything else belongs in the design document where the reasoning is actually read. The checker earned its place here. Deleting nine records left 13 dangling links across the repository and it named every one, including in AGENTS.md. Nothing was found by reading. Remaining clusters worth the same treatment: the host (8 records), delivery (5), modules (6), connectivity (4), substrate and control plane (4). That would be 52 down to roughly 30. |
||
|
jschoubben
|
10365f2eae |
Consolidate the design layer: one place per topic
Jochen: a jungle of specs that slightly contradict or patch each other, and what matters is a working state rather than history. Both are fair and both are mine. Measured rather than assumed. 05-the-node-host and 09-the-node-lifecycle both covered enrolment, the install commands, the unit file, the launcher and reconcile -- I wrote 09 without taking anything out of 05, so the same things were said twice and could drift apart. Split by what each document IS. 05 is the component: what the host is, its parts, the declaration vocabulary, the build order, how it is verified. 09 is what happens to it: install, enrol, run, upgrade, retire. The whole "The process" section left 05, and the unit file moved to 09 where installing is described. 05 goes from 338 lines to 245 and now points at 09 rather than restating it. 09 also carried a 105-line "Resolved" section -- six mechanisms framed as "these were open and here is the answer". The content is needed; the framing is history, and history is what makes a document read as a changelog rather than a description. Renamed to what it actually is and the was-open phrasing removed. Also added 10-delivery.md, which did not exist: four accepted decisions -- 0054, 0063, 0064, 0065 -- had no design document at all, which is the specific reason the delivery picture felt scattered. It is now one document covering modules, the three edges, the core library, and how a change becomes a running thing, with a table of what each property is designed against and what must exist before it can be built. |
||
|
jschoubben
|
9d091c81e0 |
A build edge, a core library that is a domain, and 0063 corrected
Three things from walking a real dev cycle through 0063, all of which Jochen caught by pushing on where I had glossed. 0064 -- a build edge is a third kind. Research 011 established presence and instantiation, and both are RUNTIME edges: they answer what a module needs in order to run. Delivery needs a different question -- what has to be rebuilt when this changes -- and that relationship is fixed inside an artifact rather than negotiated when it runs. So the graph as designed could not drive delivery, which is the real reason 0063 was not approvable. It is derived rather than declared, read from what a module actually imports, because a declared list and the imports it describes drift and the imports are the true ones. The runtime edges stay declared, and that asymmetry is not an inconsistency: a runtime edge is an intention somebody has, a build edge is a fact about code that exists. It also makes design quality measurable. A module with many inbound build edges is one whose every change is expensive, and the current shared library is exactly that -- nobody could see it because nothing drew the edges. 0065 -- the core library is the mesh's domain. Jochen disagreed with 0030's "types, not behaviour" and was right: that guard is aimed at the wrong thing. A library everything depends on is a hub whether it holds types or code, and the fan-in is what makes a change expensive. So types ship with the module that owns them -- trading one wide edge for several narrow ones -- and the core library holds what is true of the mesh regardless of context, which research 011 already found: a module, a node, an assignment. The test is "would this still mean the same thing in a context that had never heard of the one it came from". A node does; a pipeline stage does not. Domain-driven is the point rather than the label: "who else might want this" always answers yes, which is how the current one grew. And it changes the check for the better. "The build output contains no runtime code" would have enforced a rule now withdrawn. Inbound build edges is a measurement rather than a prohibition, and it is visible while a hub is forming rather than after. 0063 revised on both counts, plus a third: I had written "the lab judges it" as though that were a step. A lab run takes tens of seconds, occupies a VM, and fails for environmental reasons -- and a shared-library change produces dozens. One expensive non-deterministic gate fails both ways, and neither failure looks like itself. Verdicts are now tiered, and a run that failed environmentally is explicitly not a verdict. 0063 also now carries what must exist before it can be implemented, rather than leaving that to be discovered. |
||
|
jschoubben
|
4ab8a0507f |
Delivery is reconciliation, not a pipeline; research 008 closes
Jochen: don't rebuild the current coordinator, use it as a pitfall list. That reframed the last open question rather than answering it. 0058 stopped deploy being a stage that pushes to nodes, and said plainly what it did not fix: detection. A merge that created no pipeline, and nothing said so. That is not a defect in the detector -- it is what happens when correctness depends on an event ARRIVING. 0063 applies 0058's move one level up. The control plane holds what source exists and what has been built from it, and builds the difference. A change becomes a build because source is ahead of artifacts, which is a comparison answerable at any moment. An event makes it fast; nothing makes it necessary, so a missed webhook costs latency and cannot cost correctness. The mesh becomes one idea at two layers: the control plane reconciles artifacts against source, the host reconciles machine state against declarations. The pipeline as a state machine disappears, and with it the stage list that a verify step was once omitted from. That reframing answered the three questions still open in 008, so it graduates with all six closed. A deployed state is two comparisons rather than an event. A verdict is about an ARTIFACT and gates whether it may be declared -- sharper than the question expected. And "before self-hosting" mostly dissolves, because a reconciler needs source and artifacts as bindings where a pipeline's stages name their targets. Four costs recorded, and one is a real risk rather than a trade: a reconciler that cannot reach its target retries forever, and without something noticing, the failure is silence -- the exact fault this removes, reintroduced elsewhere. Also named: the run identity people actually use is lost, and "did my change go out?" needs a replacement or this will be worse to live with than what it replaces, whatever its properties. |
||
|
jschoubben
|
f728c3fd98 |
File 009: a digest-pinned image cannot be placed in the lab
Two accepted decisions collide, and testing found it rather than review. 0046 pins images by digest and has the host refuse anything unpinned. The lab places images by exporting them from the workstation, because a sealed scenario cannot reach a registry -- and that loses the digest, since a repo digest only exists for an image a registry served. Measured: the load says 'Loaded image ID:' rather than 'Loaded image:', and the image lands dangling. So a tag is refused by the host and a digest is unusable in the lab. There is currently no declaration the lab can raise that exercises the container shape, which matters because the container shape IS the substrate -- every bootstrap step past the runtime is one. The resolution is a registry inside the scenario, and that is not a workaround: 0048 already names an OCI registry as substrate and every node after the first pulls from the mesh's own. It also removes the lab's export-and-push mechanism rather than repairing it. 0046 now carries a pointer, since its own consequence is where the collision was predicted -- half of it is closed and the other half turned out to be harder than 'not solved here' suggested. |
||
|
jschoubben
|
9dc57b4712 |
Graduate 005; record what 0058 answered in 008
Continuing the sweep. Both were answered by records that did not cite them, which is the same pattern 003 showed -- an effort stays active because the decision that resolved it was reached from another direction. 005 graduates. Three of its four questions are answered: provider modules do not group (0044), the ~50 modules that co-change with nothing stay as they are, and 'group or leave' was never the right pair -- 0054 reframes it as authority versus package. Worth noting the debt runs the other way too: this effort's measurement, that reachability is the ONLY place modules genuinely co-change, is what 0054 rests on and why connectivity is a context while nothing else needed one. Its fourth question moves rather than closes. Whether applications leave the monorepo before or after they group is a sequencing question, so it belongs to 009-migration. 008 stays active, with its central question marked answered: the coordinator converges nodes on a declaration rather than dispatching stages (0058). The three-silo split survives with the third redefined. What 0058 explicitly does NOT answer is how a change becomes a pipeline reliably -- detection is upstream of everything it changed and remains the fragile input. |
||
|
jschoubben
|
0a37d751e2 |
Graduate research 003; file the rescue that does not exist
A sweep of the nine active research efforts. 003 was answered five days ago and nobody closed it -- the decision it asked for was taken without citing it, which is how an effort stays `active` after being resolved. Its recommendation is what the mesh adopted, and the match is exact rather than approximate. "Run the daemons as containers, making Docker the supervisor for everything" is ADR 0057. Its warning that a mesh-native supervisor inherits fate-sharing "unless it sits outside the mesh's own process tree" is where ADR 0061 put the launcher. And its insistence that it cannot be all-or-nothing is why the host itself is the one thing an init starts. Its incidental finding does not graduate with it, so it is now issue 008: the automatic node rescue the documentation describes does not exist. No unit declares OnFailure=, nothing calls the rescue script on a timer. That is worse than having no rescue. A rescue nobody wrote is a gap somebody can see; a documented one that is absent is a gap nobody looks for, and the documentation is read exactly when a node has failed and somebody is deciding whether to intervene. The issue names two honest resolutions -- implement it, or delete the documentation and say a failed node needs a person -- and says the choice is scheduling rather than technical, since the new host's recovery is built and tested. It also says what would make the finding certain: it came from reading the repository, and confirming it on a running node is the difference between "no unit declares this" and "no unit in the source declares this". |
||
|
jschoubben
|
ba0d01788e |
0062: a host may be episodic; 0060's Android gap closed
0060 named the gap and did not close it: everywhere else an init runs the launcher at boot, and Android grants neither an init to register with nor anything worth supervising, because a supervisor would be killed alongside what it supervises. Closed by narrowing what is required rather than building something. A host is resident or episodic, and both are hosts. Being killed by the platform is disconnection, which 0036 already made ordinary -- and every mechanism an episodic host needs already exists because it was built for laptops that close. A partial host can join a mesh and cannot be the first node, since every bootstrap step is a shape it refuses. Its bundle says so. Two consequences that are easy to miss: last-heard-from means much less on an episodic host, so a healthy phone reads as a dead server unless the reader knows which kind it is; and a declaration may take a long time to land, which makes 0058's outstanding-versus-failed distinction load-bearing. Still open, and in that order: what an Android node is FOR, and only then how it is started. |
||
|
jschoubben
|
f1b1cd9aa0 |
Review: three ADRs no longer said what we had concluded
A sweep for claims overtaken by the last few days. Annotated rather than rewritten, following the pattern already in 0049 -- what changed and why is the useful part, and an accepted record should not quietly become something else. 0057's init section was wrong on all three of its claims. It said the host needs FOUR things from an init; 0061 reduced that to one. It said every machine the mesh targets already has systemd; Alpine does not, and it is the intended first node. It said there is no second init to abstract over; there is now, and the answer is still not an abstraction -- it is a four-line file per system. What survives is the part that was always right: an init is not a dependency in 0041's sense, because it is not installed, it is what the machine already is. 0048 named Docker as the container runtime. It is now docker or podman, detected rather than chosen -- because adoption keeps what a machine already has, so naming one contradicted a rule already decided. That row is the only one of the five that names two, and the record now says why. 0060 claimed the bundle is portable across operating systems. Its mechanism is; its contents are not -- package names, unit names, service names all differ, so an Arch host embeds an Arch bundle. That was my error, and it is the exact confusion behind the question that found it. The design layer had the same drift: 07 and 09 said "Docker" where they meant a container runtime, 09 said systemd restarts the host after an upgrade when the launcher does, and both install snippets assumed Arch. They now show Alpine and Arch side by side, which makes the point better than prose did -- step 1 differs per system, step 2 never does. Checked and NOT changed: 0047's "the vocabulary grows by one shape" is a claim about the rate, not the count, and is still true. 0037 lists docker among tools the host manages, which it does. 0041 says nothing about either. |
||
|
jschoubben
|
66df0eb53e |
0061: the launcher supervises; init is asked for one thing
The record said an init is asked for two things -- start at boot and restart on exit -- which was half a change. It moved the give-up logic out of unit files and left the restart in one, so init still decided when the host came back. The launcher no longer execs the host. It supervises it, so restarting is ours too, and init is asked only to run it at boot. There is an OpenRC script beside the systemd unit now. Records the cost honestly: not exec'ing means the launcher must trap the shutdown signal and pass it down, because a supervisor that exits while its child runs leaves the host to be killed rather than to stop. And records what the implementation found: the counter counts consecutive FAILURES, not starts. Counting starts meant a host that upgraded itself three times rolled itself back, having worked perfectly every time -- because a clean exit IS the upgrade path. That is now the second time a clean exit has been mishandled, so it is called out as the thing to check. |
||
|
jschoubben
|
c557f99cba |
Record what testing podman actually showed
0060 said the container runtime was a separate decision. It is now made, and the reasoning is worth keeping because it is the opposite answer to the same question one paragraph earlier. Abstracting service managers is lossy -- systemd and OpenRC are different models and LoadState has no equivalent. Container runtimes converged on one CLI deliberately, so almost nothing is lost: checked against podman 6.1.0, run, rm -f and docker's own template syntax for state and labels all work unchanged. Only the probe differs. So: a two-entry lookup, not an interface. The difference that is NOT in the CLI is the one that would have shipped silently. Podman accepts --restart unless-stopped, records it, and has no daemon to act on it -- containers do not return after a reboot unless podman-restart.service is enabled, which by default it is not. Every command reports success and the effect does not happen. That belongs in the declaration rather than the host: a node using podman is told to enable the unit. Which is what made the service shape's missing 'boot' field visible, and it is now built. |
||
|
jschoubben
|
e1ad39b500 |
Per-OS hosts, and an init asked for only start and restart
0060 -- the host is built per operating system. systemd and pacman are the Arch host's implementation, not abstractions the mesh has to grow. They are not independent choices: a machine has pacman because it is Arch, and the package manager, service manager and packaging format arrive together as one decision somebody made at install time. Rejected abstracting them, and the reason is correctness rather than effort. The service applier reads LoadState to tell "not installed" apart from "stopped", which is what stops it reporting absence as success. An interface spanning systemd and OpenRC degrades to what both express, and the lowest common denominator is exactly where that fault lives. Almost all of it is shared -- the vocabulary, store, apply loop, read-back discipline, refusal model, bundle and link are portable. Two appliers differ. And delivery was already per-OS, since a .pkg.tar.zst is an Arch artifact, so this is the seam that already existed. Android is the interesting case rather than Debian: no service manager, no package installation, usually no root. Such a host implements file, directory and action and refuses the rest -- the same refusal a host already gives an unknown type, with a different reason. Those three are the portable floor. The container runtime is deliberately left open: it is not an OS split, since Arch runs docker or podman. 0061 -- the init is asked for start-at-boot and restart-on-exit, and nothing else. Both are expressible in OpenRC, runit, s6 and an Android init.rc. Counting failed starts and rolling back moves into a launcher, because that is the one piece which must work when the host does not, and a script with a counter can be tested where OnFailure= can only be hoped for. Supersedes 0059, keeping its reasoning in full. The checker found all six places citing 0059 and refused the commit until they named the replacement. |
||
|
jschoubben
|
dcc4b8339c |
Say who consumes the broker and who writes the registry
Left implicit by the previous commit, which said the owning context writes without saying what does the consuming. The control plane is the consumer, and there is one of it. Seven contexts but one deployable, so it is one process dispatching internally rather than seven consumers racing -- which matters because the as-is records two consumers accidentally sharing a queue and silently splitting the traffic, each getting half of what it expected. With one consumer that cannot arise. The broker is also the buffer while the control plane is down: nodes keep publishing, messages queue, the control plane drains them on return. That is what makes a single control plane tolerable -- an outage delays the mesh's knowledge rather than losing it. One consequence named because it will otherwise be discovered: an unbounded queue grows until the broker's disk is full, and the broker is the component every node depends on. The bound is per queue and undecided -- dropping the oldest health report is obviously right, dropping the oldest declaration acknowledgement is not. |
||
|
jschoubben
|
19997d56c3 |
Approve 0057-0059, with four corrections from review
Not approved as drafted -- four things came out of checking them against each other, and one was a bug that would have broken every upgrade. The bug: 0059 specified Restart=on-failure while 0057 has the host restart onto a new binary by exiting CLEANLY. on-failure does not restart a process that exited zero, so every upgraded node would have been left stopped, having successfully upgraded. Found by reading the two records against each other rather than by either alone. Now Restart=always in all three places that mention it. The host cannot run in a container, and the reason is decisive rather than stylistic: step 0 of the substrate bootstrap installs the container runtime, so a host inside a container would need the thing it exists to install. It would also break 0041 -- copy it onto a machine and run it stops being true when the machine must already have a runtime. Everything above tier 0 is a container; the host is not. That split is the tier boundary, not an inconsistency. systemd is named rather than abstracted. An init is not a dependency in 0041's sense: 0041 is about what must be installed before the host works, and an init is not installed, it is what the machine already is. The unit file is the only systemd-specific artefact and it belongs to the package, so a machine with a different supervisor ships a different package. The mesh is a watchdog, and my first draft was half an answer. Recovery must be local -- nothing dials a node, and a host that cannot start cannot report. But detection is the mesh's, and a local supervisor structurally cannot do it: it sees one process failing and cannot tell a broken machine from a broken release. Only something watching every node can, and that distinction decides whether the response is "fix this machine" or "stop shipping this version". So a host rollout is staged -- a few nodes, wait for heartbeats, continue or stop on silence. Local rollback still needed, because the canary nodes break and because a node offline during the rollout gets the declaration later with no batch around it. The first declaration is the overlay and nothing else. Forced, because a node's address and peers are assigned rather than chosen. But also the way back in: a node reachable over the overlay can be fixed by hand if a later declaration breaks it, and a large first declaration risks a node that is broken and unreachable at once. Also stated plainly, because it reads as a contradiction: nodes reach each other over the overlay and every node consumes from the broker; what 0039 forbids is an inbound CONTROL surface, not reachability. And in 06: no node holds a credential to any control-plane store, for reads or writes. Four ADRs already say this separately and none of them said it in one place. Nodes state over the broker; the owning context writes. With a note that most high-frequency writes are observability's, not the registry's -- routing logs into the registry would be the shared-schema mistake arriving through a door marked performance. |
||
|
jschoubben
|
605c9fd441 |
Changes are pushed, not polled; and a stuck host rolls itself back
Two corrections and one new decision, all from Jochen catching things. Pushed, not polled. I described updates as landing "on the next reconcile", which reads as polling and is not the design. A declaration arrives as a message on a link that is already open; the host applies it then. Polling over an existing connection would be slower to land AND constant traffic to learn nothing. The timer is for drift and nothing else, and it cannot be replaced by an event for a definitional reason: drift is change the mesh did not make -- somebody edited a managed file, a distribution upgrade replaced a config -- so nothing will ever publish a message about it. Only looking finds it. Separated the heartbeat from the reconcile timer, which I had been conflating. They point in opposite directions and answer different questions: the timer looks at the machine and asks whether it still matches; the heartbeat reports upward and is what makes silence mean something. A node with nothing to do sends nothing, and without a heartbeat that is indistinguishable from a node that stopped. 0059 -- a host that cannot start is rolled back by the service manager. I had left this open on the grounds that recovery meant the host judging its own health. That objection does not survive being asked properly: a keepalive is something else judging the host. The watchdog must be local, because nothing dials a node and a host that cannot start cannot report -- so it is the service manager, which is already there. The failure it prevents is sharper than "the node is down": a host that will not start looks exactly like a machine somebody switched off, which is the one condition this design has deliberately decided not to alarm on. So a bad release reaches every node, each goes quiet, and the mesh reports a fleet of sleeping laptops. Confirmed means started and completed one reconcile -- deliberately not "the link is up", or a laptop on a train would roll itself back. The rollback is a script shipped by the package, not a host subcommand, because a binary that will not start cannot be its own recovery. It rolls back once: a second failure means the machine is the problem, not the binary. Also refined the records checker, which produced a false positive: a proposed record may extend another proposed one, because decisions are drafted in chains and the alternative is marking things accepted to satisfy a check. An accepted document resting on a proposed record still fails, and that was verified. 0057, 0058 and 0059 are all proposed. |
||
|
jschoubben
|
aeea2a9f9a |
Resolve the host lifecycle's open items, and say how the host is delivered
The upgrade question turned out to be a delivery question, so 0058 answers both. Today's third silo runs once per node and sends each one a command to install and start. That is where the as-is records a package install that 404ed from every mirror while the job went green, an image pull failure that did not fail the deploy, and a verify stage that was built and never scheduled because it was missing from a list. The shape underneath all of those is that the thing reporting success was not the thing doing the work. Meanwhile ADR 0037 has given every node a component that applies state, reads back and reports -- so two mechanisms now change a node and only one checks its work. 0058: a pipeline ends when the declaration is updated. Deploy stops sending commands to nodes and becomes one write. The host applies it on its next reconcile, and the host cannot report success it did not verify. The verify stage disappears as a stage, which is the point -- verification stops being a step that can be left off a list. A pipeline result now means "the declaration is updated, and here is which nodes have applied it". It does not wait for every node, because a node may be legitimately switched off for a week. Outstanding is reported separately from failed, since conflating them is how the old system produced a stall with no error anywhere. The host is delivered by exactly this path and needs no new resource type: a `file` writes the package manager's config pointing at the mesh's repository, a `package` names the version. Added a step I had missed -- before exiting for a restart, the host runs the new binary once. A package can install something that does not execute here, and that turns "the node never came back" into "the apply failed and said why". Six open items resolved: re-enrolment is decided when the token is issued and revokes the previous identity; the mesh keeps a recovery copy of what each node reports it owns, which un-strands the orphans; last-contact is reported with no threshold, because a laptop off for three weeks is doing nothing wrong; adoption always completes but a failed line makes a node ineligible for assignment; a briefing is a structured document whose outcome is computed from its lines; and the token is printed once and carried by hand, which is the property that makes it worth anything. Still open and named: automatic rollback of a host version that will not start. 0057 and 0058 are both proposed. |
||
|
jschoubben
|
2204b01909 |
Design the node lifecycle end to end
The host was described as a component and never as something that runs for years on a machine somebody else also uses. 09 covers every state a machine can be in and every transition between them. Four states: unmanaged, hosted, enrolled, disconnected. Only the last two are nodes, and they are the same node in two situations. `hosted` -- the host installed but never told which mesh it belongs to -- had no name before and is where a machine sits between the two adoption commands. Things that were unclear and now are not: The first node walks the same path in an unusual order: reconcile from the bundle, the control plane it just raised issues a token, enrol against it. Its specialness lasts two commands. A side effect worth having -- enrolment is exercised on node one, rather than being written and first used on node two. Enrolment reports profile and inventory BEFORE the control plane decides anything. The profile is the input to that decision, not a diagnostic; the control plane cannot decide what a machine should run without knowing what it can run. Rebooting mid-apply is safe by construction. The store records each resource after it worked, so a host that dies half way through comes back and applies the rest. The rule that stops the host lying about what it did also makes it crash-safe. Retiring splits in two. Graceful is a final empty declaration. A node that is gone will reconcile its last declaration forever -- the honest consequence of making disconnection ordinary. The answer is not to make the host expire but that the node holds nothing that outlives revocation: every grant is a per-node credential revoked at the provider. A lost node keeps running and stops being able to reach anything. Said plainly rather than implying the mesh can switch a machine off, which it cannot and should not. Losing the store is quiet and permanent, so it gets its own section. The host re-enrols and re-applies fine; what does not come back is removal, because resources it no longer has a record of become unowned and sit there indefinitely. Also corrects 0057, which said the mesh must not upgrade the host at all. That conflated two acts. Replacing the binary is safe -- Unix keeps the running inode. Stopping the unit is not. So the host may apply a package naming itself, and restarts by finishing its apply and exiting cleanly, letting the supervisor start it on the new binary. It never asks the service manager to restart it. That makes a fleet-wide host upgrade an ordinary declaration, which the first draft gave up on. 0057 remains proposed. |
||
|
jschoubben
|
3ab11c96ef |
Say what the host process is: a root service, installed as a package
The design described what the host does and never what it is at runtime. The words daemon, long-running, interval, poll and heartbeat appeared nowhere in it or in the relevant decisions. What exists is a command that runs and exits; what the design needs is a process holding a link. Nobody had written down that those differ, so several questions had no answer. 0057 settles them. It runs on every node -- the host is what makes a machine managed, so a machine without one is not a node. Root, because no useful subset of the job is unprivileged. A systemd unit, because something must survive a reboot to hold the link. It never manages its own unit. The temptation is obvious and it ends with a host stopping itself half way through an apply, leaving a machine with nothing running to fix it. The installation owns the host; the host owns everything else. Installed as a package, with a tarball as the floor. The package carries the unit file, the state directory and an upgrade path, which a bare binary does not. But the mesh's package repository is hosted on the mesh, so any route that needs the mesh to install the thing that joins the mesh is a circle -- the tarball is the path that must never acquire a dependency. Reconciles on start, on a declaration, on a timer and on reconnect. The timer is the one easy to leave out, and without it `owned` reports what the host applied rather than what is there -- ADR 0035 violated by omission. The records checker caught this commit on its first attempt: 05 listed 0057 in its frontmatter while 0057 is still proposed, and a to-be document may not rest on an unaccepted record. The section now says so in the body instead. |
||
|
jschoubben
|
2330d74c1b |
The host's vocabulary is complete; 05 and 07 said otherwise
All six shapes are built. 07 still said the last three did not exist, and 05 still described stage 2 as having built three of six. Records what the lab still cannot do, because that is now the only thing between here and an end-to-end substrate bootstrap: a sealed scenario cannot fetch an image and its machines carry no container runtime, so package, container and action were verified against a real machine instead. |
||
|
jschoubben
|
03874f3fe2 |
Add a structural check over HQ's own records
Nothing in this repository was verified by anything but reading, which is how a superseded decision stayed live in the constitution and in the to-be README at the same time. Both were found by a person looking, and nothing stopped a third. Five checks: links resolve; `decisions:`/`extends:` name records that exist and are accepted; a governing document citing a superseded record must name its replacement in the same paragraph; supersession is symmetric; filename number matches heading number. Each was made to fail before it was made to pass. The live-citation check was verified against a reconstruction of the actual incident -- the to-be README citing ADR 0017 as live guidance -- and reports it with file and line. It found one thing nobody had noticed: ADR 0018 never declared that it superseded 0011, though 0011 has named 0018 as its superseder since August. Fixed. Deliberately not checked, and said so in the README: 02-DECISIONS and 01-RESEARCH may cite superseded records freely, because a decision record discusses history and research records what was observed. 00-as-is may rest on one, per 0056. Flagging those would put noise on correct documents, and a check that cries wolf gets suppressed -- which costs more than not having it. Two bugs found by running it: the frontmatter reader iterated an inline list as characters, and the as-is exemption was missing entirely. |
||
|
jschoubben
|
e1f4c7d9e0 |
Approve 0054-0056, apply them, and fix the two smaller findings
0003 is now superseded by 0056. Nothing is left proposed. Applied: - 06 corrected from ten contexts to seven plus the api, each row now stating why it passes the more-than-one-node test. work, knowledge and stream are named as mesh-hosted rather than dropped; `ai` folds into config; `record` is deferred explicitly rather than listed. Its frontmatter now cites 0055. - how-we-build §4 amended per 0054, and the derived page republished by playbook 05. The sync found the drift the playbook exists to catch: the published §4 and the source did not say the same thing. The source said "four accidents, not four boundaries"; the published page said "one intent expressed four times", and only the published page carried the scope caveat. Same rule, two texts, already diverging. Verified the republish by reading back -- the new rule is present and the old section's body returns nothing -- rather than trusting the success message. The two smaller findings: - 0051 separated the transport identity from the declaring authority. It said the token carries "an address" and "the identity to expect" without saying what the node dials. It dials the broker, so pinning only that would make the control plane's authority transitive and let a compromised broker forge declarations -- which, since the host applies whatever the link delivers, is the whole machine. The token now carries four things, and declarations are signed and verified per declaration. Cost recorded: rotating the signing identity is fleet-wide. - 0026 no longer restates 0022's rule about generated views. 0022's own words are "prose does not restate status; one place, and two is one too many", which is what 0026 was doing to it. |
||
|
jschoubben
|
f49d177a31 |
Draft three records for the contradictions the review found
0054 -- things that change together share an authority, not a package. The constitution instructs agents to group "how a node is reachable" into one module, citing superseded ADR 0017; ADR 0044 says there is no networking thing to install. Since the constitution is injected where work is decided, the superseded rule is the one actually steering work. The observation behind it was right -- research 005 measured that reachability is the only place modules genuinely change together -- but the conclusion was wrong: tight coupling means a shared authority, not one artifact. wireguard and traefik deploy to different node sets, so the merged module would be assigned where half is unwanted. Requires amending how-we-build and republishing the derived page. 0055 -- the control plane is the node-coordinating contexts. Three context lists were in circulation (0015 says nine, 06 says ten, the README said eight) and none was decided. Research 006 said explicitly that the change "belongs in a new record -- not written here", and the design used the list anyway. Reconciling them shows `stream` and `ai` were dropped with no reasoning at all. Applying 06's own test -- needs to know about more than one node -- gives seven contexts plus the api, with work, knowledge and stream as hosted applications and `ai` folded into config as an ordinary grant. The record defers rather than lists. The cost is stated rather than reassured away: a board composing across the boundary reads more than one interface. That was raised before as "only moves the problem up a layer", and the answer is that 0045 already requires surfaces to read interfaces rather than stores -- what changes is the count, not the kind of work. 0056 -- the authority is the control plane, not a database. Every clause of 0003 has been decided against in four separate records and it is still accepted and cited as live. The error underneath is the same category error 0054 corrects: "source of truth" named a storage location when it meant an authority, and once the store is the answer, shared schemas follow. The half that was right -- the repository defines what exists, the mesh defines what runs where -- survives untouched. Best consequence: the cache mode disappears, so a node that has not heard from the mesh is no longer indistinguishable from one that has. 0003 is left accepted until 0056 is. |
||
|
jschoubben
|
ef5dd0751b |
Approve 0049-0053; drop a to-be item superseded by ADR 0044
The 'domain grouping' item cited ADR 0017 as live guidance. 0044 superseded it -- there is no domain module to group into, so there is no domain list to settle. |
||
|
jschoubben
|
ccbbfa9c8a |
One node runs the control plane, and nothing takes over
Closes the two open questions in 06 and 08, which turned out to be one question: how many control planes run, and what happens when the hub is down. Both were drifting toward redundancy by default -- a standby plane, a second hub, an election to pick between them. That is not one feature but a property every layer must then honour, and each layer gets it wrong independently. Not wanted, and not needed. A handful of machines with one node hosting the registry is not a distributed system. The argument for why this is sound rather than merely cheap is that the design already tolerates it by construction. ADR 0036 makes reachability state rather than class; the host reconciles from its own store (0043) and never needed to ask anybody to hold the state it was last given. So the control plane being down is not a new failure mode -- it is every node in the ordinary disconnected situation at once. What is lost is change, not operation. No node holds a contended role: the control plane is assigned like any other module, and the overlay hub is declared (0050). No promotion, no quorum, no fencing, no split brain, no replicated store, and no "which node is authoritative" recurring at every layer. Two consequences stated plainly rather than buried. The control-plane node is a single point of failure -- deliberate, and said out loud so it stays deliberate. And recovery is restore rather than failover, which makes backup the availability story rather than hygiene. The sharpest one is the clock: the control plane owns certificate issuance (0049), so an outage outlasting a renewal window expires every public name. That bounds how long recovery may take, and nothing measures it today. |
||
|
jschoubben
|
4e80820e2f |
Design connectivity in full: overlay, resolution, exposure, filtering, certificates
Written as one document because the five are one design. They share inputs, they must agree, and every one of them today is computed in a different place by a different module from a different copy of the same facts. The through-line is that none of the five can be answered by a machine alone, so all five are decided centrally and delivered as `file` resources. That costs no new host vocabulary and removes both remaining direct database connections from nodes -- wireguard and traefik are the only two, and both are connectivity. Three decisions fall out, all proposed: 0050 -- reachability is declared, not inferred from an address. The RFC1918 regex is wrong for carrier-grade NAT (100.64/10 tests as public, so an endpoint is written to an address nothing can reach), wrong for IPv6, and wrong for a routable address behind a closed firewall. The lab needing TEST-NET-3 to satisfy the regex is the same bug from the other side. Also kills hub election by address prefix, which fails silently and makes renumbering an outage. 0051 -- the enrolment token carries where the mesh is and how to recognise it. Closes two circles with one mechanism: verifying the mesh needed the CA, and obtaining the CA meant trusting whoever handed it over; and a node had to reach the mesh before it could resolve any mesh name. An address plus a fingerprint, carried out of band, resolves both -- and closes the CA question 0049 deferred. 0052 -- a filter rule names its source. `scope:` is declared in five manifests, is part of no rule type, and is referenced by no code, so those manifests appear to restrict ports and restrict nothing. Removed rather than implemented; the general fix is refusing unknown keys, which the host already does and manifests do not. Also corrects two claims in 0049 asserting wireguard was already handled. Research 006 says both modules still reach upward; neither is. |
||
|
jschoubben
|
8d9282d86b |
Resolve the ingress gap: a route is a grant
ADR 0048 named ingress as an unclosed hole -- nothing said what terminates TLS, how a public name reaches a container, or which tier owned it. Resolving it needed no new concepts, which is why it survived: nobody had applied the rules already written to it. Ingress is not substrate. The control plane does not need a route to start, and no node needs one to reach it -- the node dials out and has no listening control surface. It grants itself a route afterwards, like a bucket. A route is an instantiation edge under ADR 0044. The direction mirrors a database -- the consumer supplies a target and receives a name rather than credentials -- but it is the same edge. The substantive finding is that exposure is three facts at two scopes: name resolution and certificate issuance need to know which node is publicly reachable, and only the proxy mapping is a single machine's business. That is why it belongs to the connectivity context, and why Traefik doing all three on the node is wrong. Which matters beyond tidiness: research 006 counted traefik as one of two modules opening a direct Postgres connection, reading nodes and mesh_ca. That violates 0037, 0045 and 0039 at once, and is why every node permanently holds a credential to the control plane's database. Deriving the config centrally and delivering it as `file` resources removes it, costs zero new host vocabulary, and closes the set 0039 identified -- wireguard was the other. Left open deliberately: the mesh's internal CA is the other thing traefik reads, and it belongs to the link's mutual authority, not to exposure. Conflating the two is what made the gap hard to see. Also fixes an inconsistency from the previous commit: 06 still claimed the virtual host was raised from the bundle. Proposed, not accepted -- for review. |
||
|
jschoubben
|
4d19e93900 |
Name the substrate's actual products
The design layer described every service by role and never once by name: Postgres appeared in zero design documents. That was over-application of the research rule "never identify the mesh it observed", which is about node names and domains, not software. Two things were actually broken by it. substrate.lock pins images by digest and a digest belongs to a named image, so the bundle could not be written from the design. And a reader could not tell a settled choice from an unexamined one -- "a relational store" reads identically either way. ADR 0048 names them: PostgreSQL, LavinMQ, MinIO, an OCI registry, Docker. The argument for each is continuity, which is a real argument -- replacing a substrate service migrates the mesh's own state. Role and product are now both written, because the design depends on the protocol while the installer needs the product. Also separates two questions the substrate doc had merged: being substrate and being in the bundle. Only Postgres must precede the control plane; the rest are substrate by role and ordinary by delivery. Whether the bus joins it is left open, because it turns on the control plane's internal shape. Names the forge as Gitea, and records ingress/Traefik as an unclosed gap rather than a naming one -- nothing says what terminates TLS or which tier owns it. Fixes a miscount: the host's bootstrap vocabulary is six shapes, not five. |
||
|
jschoubben
|
c631cbd07c |
The bootstrap starts a step earlier than recorded
Asked whether postgres has to be installed, and the answer exposed a missing step. The store is a container, so something must run containers before anything else happens — and a container runtime is a PACKAGE, not a container. Step 0 is where several threads meet. It is what the host's capability detection already reports, and the first use of that report by something other than a person. It is adopted rather than installed when the machine already has a runtime with configuration somebody chose. And it is a package, needing the machine's own package manager and a network, both of which ADR 0046 permits. So the host's bootstrap vocabulary is six shapes: package, container, file, directory, service, action. Stage 2 built three of them. The node host design now names which three remain and why the lab cannot yet exercise them — a sealed scenario fetches nothing and its machines carry no container runtime, which is lab-installation work rather than a constraint on the design, because production machines have a network. |
||
|
jschoubben
|
93470f6162 |
ADR 0047 — the bundle may carry actions the link may not
The bootstrap's sharpest open question, and the framing was wrong. "State on this machine" was being read as the filesystem and the service manager. A service running on this machine IS part of this machine — writing a file and creating a database in a local store differ in mechanism, not in scope. The real question was underneath: must the host learn what a database is? It must not. Giving it a `database` resource type means tier 0 knows Postgres, then a bucket, then a virtual host — the host acquiring the substrate's vocabulary one service at a time, which is what ADR 0037 exists to stop. So the bundle declares an ACTION and the host runs it and verifies it. What a database means stays with the module that provides one; the host knows only how to run a declared action against something local and check the result. Its vocabulary grows by one shape rather than by one resource type per service. Actions are permitted in the bundle and forbidden over the link, and the asymmetry is deliberate. A bundle arrives WITH the binary: anyone able to put a hostile action in it could equally have put it in the host itself, so refusing actions there buys nothing and costs the bootstrap. The link is a separate party, reachable separately, and an action there is the unbounded blast radius ADR 0039 refuses. That decision stands unchanged. And ongoing provisioning is not the host's at all — the control plane does it once a mesh exists — so the asymmetry costs nothing. Which dissolves the earlier worry about one mechanism with a tier boundary inside it: there are two mechanisms, with different actors, scopes and trust models, and that is the answer rather than a compromise. Named rather than hidden: this is the escape hatch research 011 warned about, arbitrary code in the place hardest to remove later. It is bounded by being bundle-only and by every action having to declare how it verifies itself, and that boundary is the whole defence. |
||
|
jschoubben
|
5b3d0ebd4f |
ADR 0046 — the installer fetches what it pins
The blocking question was where a container image comes from, and the version that blocked assumed the machine might have no network. That assumption came from the LAB: a scenario is a closed address space by design, which is what lets two scenarios hold the same addresses without meeting. Production is not sealed — a machine being adopted has a network, and one that does not is a machine where very little works anyway. So substrate.lock carries references, not payload: an image name and a digest, fetched at apply time. A first node pulls from upstream because no mesh registry exists yet; every node after that pulls from the mesh's own. The lab is the exception and places images itself, the way it already places the host binary — a property of a test environment, and letting it dictate the production design would be the tail wagging the dog. Pinned by DIGEST rather than tag. Reproducibility comes from pinning the identity of a thing, not from carrying its bytes, which is what makes fetching acceptable rather than a compromise. ADR 0041 survives untouched, which was the point. "Copy it onto a machine and run it" stays literally true — one binary, a few megabytes, which then fetches what it was told to. Carrying images would have quietly redefined the property that decision rests on. Costs accepted and named: an apply can now fail because something is unreachable, which a self-contained artifact could not, so it must fail legibly — naming what it could not fetch and from where. And the lab needs a way to place images into a machine that also has no container runtime, both of which are lab-installation concerns and neither solved here. Research 012's build-time-versus-apply-time reframing narrows accordingly: it still holds for what a tailored installer contains, and no longer has to hold for images. |
||
|
jschoubben
|
0531d6fc38 |
ADRs 0044 and 0045 — the module design, closed; 011 graduates
011 opened asking what a graph deletes and found the graph already existed. The work became design, worked through twenty cases and one provider in full. Two decisions close it. 0044 — a module declares presence, instantiation and exclusion. Two kinds of edge because a game wanting a database is not a game wanting postgres to exist: one creates something per consumer, carries credentials back, can be revoked, and leaves the provider holding state. Names are concrete unless providers are genuinely substitutable — `terminal` passes, `database` fails, and the adapter is what creates an interface. Where there is no contract there is a tag, which describes and does not bind. Exclusion is a third relation and is not derivable. A node provides names too, which makes capability checking stop being a separate mechanism and makes the host's detection an input to resolution. Constraints, never placement. Scope decides which provider and the binding is written down and sticky, in a place that follows the scope. And there are three entities, not two — the assignment carries what belongs to neither end, which is what node-agnostic modules ran out of. 0045 — a context owns its store, exclusively. No shared writes and no read roles on another context's store, because reading couples you to its layout just as firmly and invisibly. The unit is the CONTEXT, not the process: a board showing the mesh's own data is the mesh showing its own data. Asking or subscribing is derived from ADR 0036 rather than chosen. And it is the first clear list of what the design removes: grant kinds, table ownership, cross-context migration ordering, and a class of permission modelling. 0017 is superseded rather than narrowed — its text unchanged, its status changed. Folders assert relationships where edges record them, and the domain module goes with it. Left explicitly undecided in both: what a resolver delegates rather than reimplements, how many instances a module should have, and what a provider hands back. |
||
|
jschoubben
|
60aea14935 |
Define the substrate, and answer 006's four-or-five conditionally
Same gap as the control plane: load-bearing and unpinned. The substrate is what the control plane CONSUMES AND CANNOT GRANT ITSELF. Every module needing a database asks provisioning for one; the control plane needs one too and cannot ask itself, because it is not running yet. That circularity is not an awkwardness to work around — it is the definition, and anything on the wrong side of it must be raised by the bundle the host carries. Which answers 006's open question in the honest form rather than with a number. The identity provider is substrate only if the control plane DELEGATES authentication — then it cannot serve anybody before the provider exists and cannot grant itself a client. If it authenticates natively, the provider is an ordinary hosted service. So the count follows from a decision not yet taken, and asserting four was asserting that decision. The test also rules out the tempting wrong answer: an identity provider, a mail server and an analytics service are all infrastructure by any ordinary reading, and none are substrate, because the control plane starts and runs without them. Important is not the test. Records why the bundle is pinned by hand — it is applied when no mesh exists, so nothing can resolve a version or ask a registry — and why it must be self-contained, which makes it an artifact built on a machine with a network for a machine that may have none. |
||
|
jschoubben
|
148395ca54 |
Define the control plane, which was used 79 times and defined nowhere
Nineteen files, seventy-nine mentions, no definition. That is how-we-build §5 failing on this repository's own vocabulary — ubiquitous language is checked, not assumed. The definition, and it is not arbitrary: the control plane is everything that needs to know about MORE THAN ONE NODE. It follows from ADR 0037, which has the host applying rather than deciding precisely because deciding needs knowledge the machine does not have. So the line falls exactly there — writing a file is the host's, choosing which nodes run the store is the control plane's, and anything a single machine could answer alone does not belong here at all. That last consequence is worth having: putting a single-machine concern in tier 2 is a mistake the tier rule will NOT catch, because the dependency direction stays correct. Also states what it is not — not the thing that changes machines, not a surface, not the substrate, and not privileged on a node beyond what the declaration vocabulary allows. And the property that makes tier 2 unlike the others: it is itself a consumer, with the same requirements as any module, which is the circularity the bundle exists to resolve rather than hide. Scoped deliberately: this defines the term and does not design the contexts inside it. Ten is the skeleton's claim rather than a settled list, and research 006 still asks whether the record belongs here or in the substrate. |
||
|
jschoubben
|
a4ab3e15c2 |
011: one interface, many contexts — and the constraint that hides in it
The objection is right: if every context runs its own service with its own interface, the board is coupled to N interfaces instead of N schemas, something has to compose them, and composition is logic — which tier 3 says a surface does not hold. That moves the problem up a layer rather than solving it. The skeleton already answers it, and the previous entry talked past it. `work` and `knowledge` are not separate services; they are contexts INSIDE the control plane, alongside the record, inventory and delivery — and `api` is listed there as the one interface every surface speaks to. So the board speaks to one interface. Behind it the contexts stay separate, integrating through the record, but they are one tier, one repository, one deployable — and coupling within a tier is not what the tier rule forbids. The problem does move up a layer, and the layer it moves to already exists and has this as its job. The caveat is load-bearing and now recorded as an open question: this holds only while the contexts are not separate deployables. The moment one becomes its own service with its own interface, the board is back to N clients, something must compose them, and the composition has nowhere to live that tier 3 permits. That is a real constraint on how far the control plane may be split, and it is worth knowing before splitting rather than after. |
||
|
jschoubben
|
a4a25ca7e3 |
011: one surface over several contexts is normal
The board visualises the mesh, the work engine, the knowledge base and more, and the alternative — a web application per context — is worse for everyone using it. Composing several sources into one view is what a surface IS, so this is not a compromise with the ownership rule. What changes is only where it reads from: each context's interface rather than each context's store. Most of that already exists — 56 of 126 modules carry a tool surface, more than carry a service. And the unified board is what keeps those interfaces honest. A view that cannot be built from a context's interface proves the interface inadequate, discovered where it is cheap to notice rather than the first time something else needs the same data and quietly reaches for the store instead. If composing many calls proves too slow, the answer is a projection the board owns and keeps current from events, not access to somebody else's tables. |
||
|
jschoubben
|
fa62c7f0e4 |
011: correct the rule — contexts, not processes
An earlier version argued a dashboard reading a dozen stores was caught by exclusive ownership, because a dashboard is a surface and surfaces speak to an interface. Wrong, and it drew the line in the wrong place. The mesh's own board showing nodes, modules and deployments is not a separate context reaching across a boundary — it is the mesh showing its own data. Requiring it to go through an interface to reach facts its own context owns is ceremony. The rule is that a CONTEXT is granted what it exclusively owns. Everything inside it — service, surface, tools — reads that store freely. What is forbidden is a different context reading it. Which is what the consumer count already showed: the problem was never surfaces, it was three other contexts keeping their tables in the mesh's database. |
||
|
jschoubben
|
e71d532c2e |
011: request or subscription is derived, not chosen
Asked what the distinction actually is, and the SQL half needed correcting first: under exclusive ownership SQL runs against your own database and nothing else, whatever transport a query might travel over. Both options are the mesh's own channel and both ride the broker, so the transport is not the distinction. The distinction is where the answer lives when you need it. A request asks at the moment and waits — always current, costs a round trip, cannot answer when the other side is down. A subscription keeps a local copy — instant, works offline, as current as the last event received, and you must handle what you missed. What decides is not taste. ADR 0036 makes disconnection an ordinary situation rather than an exception, so anything that must keep working while disconnected CANNOT use a request: there is nobody to ask. And the converse — anything where a stale answer is worse than no answer cannot use a subscription. A display can lag; a decision about whether a grant is still valid cannot. So an apparently open question turns out to be derived from a decision already taken. What stays open is narrower: what a consumer does about the events it missed while disconnected — replay from a point, ask once for a full picture and resume, or rebuild. The question every projection has. Also recorded: separate databases are required in the new design, and the shared registry is a leftover rather than a pattern. |
||
|
jschoubben
|
7e83723b7b |
011: rewrite the question table, which had gone stale silently
Several edits to the overview matched nothing and returned success, so the question table still carried answers superseded two or three exchanges ago — "when two modules provide one name, who chooses" was still open in the table while answered in the file it pointed at, and nothing recorded the instantiation edge, instance counts, grants, bootstrap provisioning, the tool audience, or the registry consumer check. That is the fault this repository catalogues, committed by the thing cataloguing it: a string replacement that found no match, reported nothing, and left the document claiming a state it did not have. Rewritten from what the documents actually say rather than patched again. Nine questions settled, fourteen live, and the split is now visible instead of implied. |
||
|
jschoubben
|
afcc355744 |
011: checked the registry's real consumers, and the question was the wrong shape
The exclusive-ownership rule turned on whether every reader of the mesh registry could be served another way. Eighteen consumers open a direct connection. Four groups, and only one is work. The owner and its machinery keep reading, because they own it. The node appliers are already resolved — ADR 0037 stops the host querying the mesh database, decided for tier reasons with nothing to do with this. The bulk are FOREIGN TENANTS. The work engine holds ten of its own tables in the registry's database, the knowledge base two, pipeline logs one. Thirteen foreign tables across three contexts, which is how-we-build §4's shared schema counted. So the question was the wrong shape: the problem is not readers needing a new route to data, it is tenants needing to move out. Tasks, agents and teams have nothing to do with nodes and modules and are co-located by history. Give that context its own database and its dependency on the registry shrinks to one table. A handful of genuine cross-context reads remain, small enough to enumerate rather than estimate. The rule holds. Left open: whether those reads want an interface or events. Asking which nodes exist at the moment you need to know is a request; reacting when a node appears is a subscription, and some consumers want both. |
||
|
jschoubben
|
6b1aab6a1e |
011: the dashboard case, and why exclusive ownership is the tier rule
Raised as the hardest test of the rule: a board showing nodes, modules, pipelines, agents and tasks wants to read a dozen stores, and under exclusive ownership it can read none of them. It survives, and not by luck. The board is a SURFACE, and surfaces already may not do this — the skeleton puts `api/` in the control plane as the one interface every surface speaks to, and tier 3 as thin, no logic. A board reading stores directly is a surface reaching past the context that owns the data, which the tier rule forbids for reasons that have nothing to do with databases. So it is not a counter-example; it is an instance the rule catches. And the two rules turn out to be one rule seen from two sides: exclusive ownership is the tier rule expressed in terms of storage. The general shape for anything needing to see across many things: consume the record and own your own view. A reporting context builds a projection from events and reads its own store, never anybody else's. The cost said plainly rather than buried: a projection is more work than a join, and it lags. A board queries the mesh's own database directly today — ordinary, working — and this rule makes that a migration rather than a preference. The reason to pay it is §4's already-measured cost, not elegance. |
||
|
jschoubben
|
aa767d17a8 |
011: a module is granted only what it exclusively owns
Reconsidered by the operator — maybe shared databases should not be allowed at all — and the stricter version is better and goes further than the schemas it replaces. No shared writes, and no read-only role on another module's database either. Reading another context's tables couples you to its layout exactly as firmly as writing them does, and the coupling is harder to see because nothing breaks until the owner changes a column. That is how-we-build §4 taken at its word rather than at its letter. The permissive version — a per-consumer schema, revocable, with cross-context joins possible but deliberate — kept the letter and left the temptation. A boundary that is merely inconvenient to cross is a boundary that gets crossed. The cost is cross-module reporting, and it is the point rather than a regrettable side effect: anything wanting to know what several modules hold consumes their events or calls their interface. That is §4's whole argument, and the mesh already has both mechanisms. What gets harder is precisely the thing that was making work belonging to one context keep having to be implemented in another. And it is the first clear instance of what this effort has been hunting — what the design DELETES rather than adds. Grant kinds collapse to one: an exclusive resource. With them go the question of who owns which table, the guessing at revocation time, cross-module migration ordering, and a class of permission modelling a shared store would otherwise need. One thing it does not answer, recorded because it could make the rule unworkable: the mesh's own registry is read directly by many things today, and under this rule they consume events or call tools instead. Achievable in principle. Whether EVERY current consumer can be served that way is unchecked, and should be before this becomes a decision. |
||
|
jschoubben
|
4c8515507a |
011: where a binding lives follows the scope, and a grant is not always a whole resource
Two questions asked directly, and the second collides with a rule in force. One module on two nodes sharing a database corrects something stated flatly: the binding is not "recorded on the assignment". Where it is written down FOLLOWS THE SCOPE. A shared grant belongs to the module and every assignment references the same one — which is the answer for two nodes wanting one database between them. A per-instance grant belongs to the assignment. Same relation, two homes, and which home is what makes two instances share something or not. Several modules adding their own tables to one database is three needs wearing one sentence, and a provider offers KINDS of grant rather than one: a database for a consumer whose tables are nobody else's business, a read-only role for one that needs to see what another holds, and a SCHEMA within a shared database for the case actually asked about. Loose tables in a shared database is what how-we-build §4 warns against in as many words — several domains sharing one forty-five-table schema, which is why work belonging to one context keeps having to be implemented in another. Not a style objection; the observed cost, already paid. A per-consumer schema keeps what the request wants and drops what §4 objects to. Same database, same connection, same backup, and a cross-schema read remains physically possible when genuinely needed. What it adds is ownership: migrations touch one namespace, two modules cannot collide over a table name, and revoking drops the schema rather than guessing which tables belonged to whom. So the fault §4 names is still possible and no longer accidental — a cross-context join becomes something somebody deliberately writes rather than the path of least resistance. And revocation becomes answerable, which the whole-database version never was. |
||
|
jschoubben
|
fa9889536c |
011: tools have a different audience, migrations cross the edge, provisioning is early
Three additions, and the third kills an assumption. Tools are the most common content in the catalogue — 56 of 126 modules, more than carry a service — and they survive the split without fitting either half. A tool is not an artifact and not node state; it is a contract the mesh publishes on a module's behalf, and what consumes it is an AGENT rather than another module. That is a second audience the design has not described. Whether it is one relation with two audiences or two relations is cheap to decide now and expensive later. A migration belongs to the CONSUMER and runs on the PROVIDER. A game's migrations run against the database the store granted it: owned by the consumer, hosted inside something it does not control, ordered after the provisioning edge because there is nothing to migrate until the grant exists, and scoped to that grant. Ownership crosses the edge, which nothing in provides and requires expresses — and it gives a consumer's own install an internal order, provisioned then migrated then started, that depends on an edge rather than on its contents. And provisioning is EARLY, not late. The assumption worth killing is that it is something the control plane does for consumers once a mesh is running. The mesh's own registry database is provisioned before there is a mesh, and so is its virtual host on the broker: the store runs from the carried bundle, a database is created in it, the mesh's own schema is applied, and only then does a control plane exist. Steps two and three happen before there is a mesh to do them, so provisioning is part of the bootstrap and part of what the bundle has to express. Which strains ADR 0043. The host applies declared state ON THIS MACHINE, and a database inside a running store is not a file or a unit. At bootstrap it is at least local — the store is on the same machine. Afterwards a consumer on one node provisioned from a store on another is the ordinary case and reaching it is not the host's job. The same operation is local at bootstrap and remote later, which is either two mechanisms or one with a tier boundary crossing inside it. Currently the sharpest unresolved thing in the effort. |
||
|
jschoubben
|
a3c7e7e1f1 |
011: providing is a facet, and the assignment is a third thing
Any hosted service can be a factory — an identity provider grants clients, an analytics service grants a tracking identity, a mail server grants mailboxes, an application platform grants a project that is several of those at once. Providing is a FACET a module may have, not a kind of module it is, which is the same conclusion this effort reached about services and applications arriving from the other direction. So `provider` stops being a category too. Two relational stores from different vendors both grant "a database" and are the sharpest possible test of the substitutability rule. They fail it completely — different protocol, dialect, driver, client library compiled into the consumer — so `database` stays a tag, now with two real providers rather than a thought experiment. The assignment is a third entity, recorded because the operator tried the alternative: modules were once node-agnostic and it did not survive. Several of a provider's properties belong to neither end — where its state lives, how it is reached, tuning derived from the machine's hardware, which instance serves a given consumer. Not the catalogue, because they differ per node; not the node, because they are about this module. A design with only modules and nodes has nowhere to put them, which is what node-agnostic ran out of. The current system already stores environment values per module AND per node, arriving the same way. Which answers the question asked directly: two nodes both run a store, so which serves a consumer? Neither obvious answer. Not the consumer naming a node — that is placement in the consumer's manifest, a game edited because a database moved. Not the consumer not caring — for presence it genuinely does not, for instantiation it cares permanently. What the consumer knows is the SCOPE of its own need: one instance shared across every instance of itself, or one each. That decides, and needs no node named. Then the mesh binds, and the binding is recorded on the assignment and is sticky — a resolver that re-derives which store serves a consumer will one day derive a different answer and relocate a database. |
||
|
jschoubben
|
13c6068874 |
011: the provider shape generalises, and two things differ inside it
The broker has all nine properties the store has. So do the object store and the image registry. A substrate service is a SERVICE PLUS A FACTORY, there are four of them, and the pattern generalises past the substrate: anything granting something per consumer has this shape. Two differences matter more than the similarity. The broker cannot be managed over the broker. ADR 0001 makes it the channel every node takes work from and ADR 0039 makes it the security boundary, so the module providing it is also the way modules are managed — a declaration cannot be delivered to it over itself. Nothing else has that property; the store is consumed by the control plane but is not how the control plane REACHES anything. This is what the carried bundle exists for: the broker is raised from what the host carries because there is no other way to raise it. A constraint on one module, not a general rule, and a schema with no way to say so hides it. And two modules of identical shape want opposite instance counts. The broker is one per mesh by decision. The store cannot be, because a node that must keep working while disconnected cannot depend on a database elsewhere. Which settles what cases.md left open: how many instances is NOT derivable from what a module is. It is a per-module decision, it has to be declared, and nothing in provides, requires or excludes says it. Revocation differs in consequence too. Dropping a database leaves data until something removes it — a leak, recoverable. Dropping a virtual host loses whatever was undelivered — silent, and not. Same relation, different blast radius, which argues for the provider deciding what revocation means rather than the mesh applying one rule. File renamed: it was never really about postgres. |
||
|
jschoubben
|
f160b28a71 |
011: postgres worked through, and "one kind of edge" was wrong
The tidy version said a module provides names and requires names and that is the only edge. Working postgres through completely disproves it. A small game wanting to store data does not require postgres to EXIST. It requires postgres to MAKE IT A DATABASE and hand back credentials. Those are different relations in every way that matters: one creates something per consumer, carries a payload back, can be revoked, and leaves the provider holding state about who was granted what. The other creates nothing. So: two kinds of edge, one graph. Instantiation implies presence; presence does not imply instantiation. The current system already had exactly this split — `dependencies` for presence, `requires: provision:` for instantiation, with the resolver deriving one from the other. analysis.md called that derivation a convenience. It is not: it is the correct relationship between two genuinely different relations, and the design had collapsed them. Postgres also turns out to be nine things, not one. A container. Persistent state where moving nodes is a migration rather than a reschedule. Configuration partly derived from the machine's hardware. A tool surface. A provisioner. Its own bookkeeping about what it granted, which is not the data it stores. An exposure decision per node it runs on. Credentials it generates, which means a provisioning edge carries a secret. And health that is not "the container is up". Four questions the worked example makes concrete rather than abstract. WHICH postgres, when there are two — a consumer of `terminal` does not care and a consumer of a database cares permanently. How many instances a module should have, which cannot be a global rule because one-per-mesh is wrong for a store a disconnected node needs and one-per-node is wrong for the mesh's own registry. What happens to a grant when its consumer is removed, where dropping is data loss and keeping is a leak. And whether a declaration is composed PER NODE from what that node reported — because tuning follows hardware the control plane cannot know, and the alternative is the host deciding, which ADR 0037 forbids. |
||
|
jschoubben
|
c9c2dfe686 |
011: what a feature is, and what it splits into
The operator wants features gone, and 006 left it open. Measured, and the answer is that nothing replaces them because they were never one concept. A feature is a kind of content a module carries, detected from its directory: twenty-one of them, each with a handler owning six stages — build, publish, install, configure, start, verify. The structural finding: EVERY handler implements EVERY stage. `configs` writes files onto a node, has nothing to build, and has a build stage. `npm` publishes to a registry, has nothing to start, and has a start stage. One interface spans build-time and apply-time, so every kind of content must implement both halves and most do nothing in one — and a stage that does nothing looks exactly like a stage that failed to do anything. They split four ways, across three tiers. Artifacts built once per version and published, where no node is involved — delivery. Resources that are desired state on a machine, which is what ADR 0043 already describes and the host already does — tier 0. Actions run once against something that is not this machine, like a migration against a database on another node — delivery, and seeds go entirely. And checks: the prerequisites are REQUIREMENTS IN DISGUISE, a module saying what must be true before it can be installed, which is what an edge in the graph says; the verifiers are the read-back the host already performs. So `feature` is one word for four things spanning three tiers, which is why the pipeline is hard to reason about. One property must survive the split, and it is the thing the current design got right: content is DETECTED, relationships are DECLARED. A module that says it has migrations and has none is a fault nobody sees until it matters — but what it requires and provides is not visible in a directory and has to be said. |
||
|
jschoubben
|
ae099482a9 |
011: twenty cases, and two axes nothing covers
Before settling a schema, what a module can actually be. Twenty kinds of thing, with the hard ones at the end because they are the point. The ordinary nine are unsurprising: a supervised service, a system package with configuration, an application a person launches, a command-line tool, a library that never runs, a one-shot task, a scheduled one, an adapter, and a standalone application whose only difference is where its source lives. The eleven that break a naive schema are where the work is. Something that is a service AND an application — a git forge is consumed as a remote and operated through a web interface, and neither reading is wrong. Something that provides and consumes, because provider and consumer are ends of edges rather than kinds of module. Something the mesh installs that then becomes a node CAPABILITY, which means a node's provides-list is partly derived from what is installed on it and not only detected. Something that must be adopted rather than installed. Something that is a set rather than a thing. Something with exactly one instance for the whole mesh, where assigning it twice is not redundancy but two meshes. Something that is not software at all — a firewall policy, a DNS record, pure desired state, which fits the host's declaration model exactly and an installable package model not at all. An agent. The host itself, which is not a module and needs a schema that can say so. And the things the mesh depends on and does not control, which are why a node can be perfectly configured and still not work. Nine axes come out of it. Two are covered by nothing anyone has proposed: HOW MANY INSTANCES a thing may have, and WHETHER TWO CAN COEXIST — `excludes` covers part of the second and nothing covers the first. And one question the cases sharpen: is "runs" a property or a kind? The axes say property — one schema with a field saying how it runs, `never` included. The alternative is several kinds of module with different schemas, which is the taxonomy this effort already rejected once for services and applications. |
||
|
jschoubben
|
f55ecc1a47 |
011: an abstract name needs providers that are actually substitutable
Two corrections from the operator, and the first improves the design rather than narrowing it. `database` is not an edge. The test it fails, and the test the proposal was missing: can a consumer be switched from one provider to another WITHOUT CHANGING? A module speaking Postgres does not speak MongoDB or SQL Server — different wire protocol, dialect, driver — so a consumer declaring `requires: database` and handed any of them breaks. The name promises what no provider can deliver, and the resolver would report a requirement satisfied that is not. `terminal` passes: anything that runs a command in a terminal works and the consumer never learns which it got. So the ADAPTER is what creates an interface. `ai-assistant` is legitimate exactly because adapters normalise what is behind it. Without one there is no interface, there is a category — and a category is a TAG. Tags describe, edges bind, and keeping them apart is what stops the catalogue acquiring a second kind of relationship that looks like a dependency and is not, which is what a folder named after a domain already was. And the domain module goes. A `networking` module gathering a firewall, a resolver and a proxy under one name came from an older shape and does not fit — there is no such thing to install. There is core infrastructure: concrete modules named individually, not flavourable, with no grouping module standing in front of them. Fixed three places where the revision left the old rule standing, including an example manifest still requiring `database` — the kind of contradiction that would have been read as the design rather than as a leftover. |