TelemetryStreamAuthInterceptor, ConfigServeAuthInterceptor and
LocalDbSyncAuthInterceptor all decided "is a key configured?" with
string.IsNullOrEmpty. A whitespace-only configured value therefore passed
that guard and fell through to the constant-time comparison — where it is
satisfied by a whitespace-only bearer token. The gate would be open to
anyone who guessed a blank credential.
This backports the check the two NEW mesh interceptors already use:
MeshCommandAuthInterceptor (and its central-side clone) validate with
string.IsNullOrWhiteSpace, so the four/five planes now answer the same way
to the same misconfiguration. Nothing else changes — the denial status
codes (PermissionDenied / Unauthenticated / PermissionDenied respectively)
and every denial message are untouched, and no key material is logged.
No deployment configures a whitespace key: the docker-dev rig and
production both set real keys, so this is a fail-closed hardening of an
unreachable-today misconfiguration rather than a live exposure.
Tests: each of the three test files gains a ..._WithBlankKeyConfigured_IsDenied
fact mirroring MeshCommandAuthInterceptorTests.ExecuteMethod_WithBlankPskConfigured_IsDenied
(configured key " ", presented "Bearer "), verified red before the fix.
ConfigServeAuthInterceptor and LocalDbSyncAuthInterceptor also gain the
HasExactlyOnePublicConstructor pin their two siblings already had —
Grpc.AspNetCore silently stops invoking an interceptor with more than one
public constructor, which would disable auth without failing anything.
Claude-Session: https://claude.ai/code/session_014WNM4vjoVksyyBraTXSZE1
MeshGrpcBoundaryTests replaces the MeshClusterClientBoundaryTests deleted in step 1. Three
self-seeded single-member clusters sharing the ActorSystem name, joined ONLY by gRPC, driving
the REAL CentralCommunicationActor -> MeshGrpcCommandDialer -> Kestrel h2c ->
MeshCommandAuthInterceptor -> NodeCommandGrpcService -> NodeCommunicationActor -> EventStream:
1. No_cross_mesh_Akka_association_forms_between_central_and_either_site — the program's own
point, kept as a test rather than a live-gate note: each cluster stays single-member and
none ever holds another's address as a member.
2. One_dispatch_through_central_reaches_a_node_in_every_cluster_mesh — the fan, with each
site holding its OWN per-cluster PSK, so a dial that selected the wrong cluster's key
would be refused at the far side.
3. Falsifiability_control_a_dial_carrying_the_wrong_cluster_key_delivers_nothing — without
it, (2) cannot distinguish "the gRPC boundary carried it" from "something else in the
process did". The same node then accepts the right key, so the silence is the auth gate
and not a dead endpoint.
All three run fully offline — no Docker fixture — and pass.
Comment sweep across every remaining clusterclient|receptionist hit in src/ and tests/:
the program label becomes "mesh gRPC command plane", and every comment that described the
deleted transport as live now describes what actually runs (Dps = pair-local default,
Grpc = cross-mesh). mesh_transport.proto's header/service comments are reworded; no message,
field or enum is touched — the contract is locked. MeshCommActorPathTests is reshaped rather
than retired: the two paths still exist and still matter, but what a mismatch costs is now a
misleading Unavailable from the gRPC ingress rather than a silently dropped receptionist send.
The only surviving mentions of the old names are the ones that must survive: the fail-closed
rules that name "ClusterClient" as a retired value refused at boot, their tests, and this
test class's own account of what it replaces.
The known-cluster gate published only under Mode=Grpc, but the
interceptor it feeds is registered on every admin node and 1B maps
CentralCommandService by LISTENER PORT — neither consults the mode. A
dark-switch rollback to ClusterClient or Dps would therefore have left
that endpoint serving with the gate permanently inactive, silently
restoring both the timing oracle and the pre-auth secret-store/audit
flood, as a side effect of a rollback nobody would connect to
authentication. The gate is central's cluster inventory, not transport
state, so it is now decoupled from the mode entirely:
- SetKnownClusters + InvalidateAll run on EVERY successful contact load
in EVERY mode, and the contact-refresh timer starts in all three.
Under Dps this is a ConfigDb query every ContactRefreshSeconds that
did not run before; central always has a ConfigDb, and one indexed
read a minute is the cheaper half of the trade.
- The GATE set is built from ALL ClusterNode rows. The eligibility
filter moved out of the query's WHERE and into the loop, so the FAN
set is byte-for-byte what it was (enabled && !maintenance) while the
gate widens. This also dissolves the maintenance coupling flagged in
1A.2: a cluster whose nodes are all disabled or in maintenance can
still ack, and refusing it would make maintenance mode an
authentication switch. Central still refuses clusters it has never
heard of.
- HandleContactsLoaded keeps a documented order — targets adopted, gate
published, cache dropped — now pinned by test.
Guards: PreStart mirrors the null-dialer error for a null PSK provider
(silent absence is the worst outcome there — nothing malfunctions, so
nothing is investigated, while the gate never publishes and keys never
rotate); a failed contact load now says out loud that the gate is stale
and, before any load has succeeded, inactive; MeshGrpcCommandDialer
clamps a non-positive DispatchTimeoutSeconds to 10 with a construction
Warning, because the validator for that key is driver-role-scoped and a
pure-admin central reads it unvalidated; the dialer's local PSK refusal
was shortened so its remediation survives the actor's 200-char
truncation of remote-authored status detail; the portless-node warn-once
set is rebuilt each refresh, so a node that regains then loses a port
warns again — the regression is the occurrence worth seeing.
Docs/pins: dialer disposal comment no longer claims one-way; the two
interceptor-test comments claiming Matches has a default implementation
corrected (it deliberately has none); new pins for per-refresh
invalidation, the SetKnownClusters-before-InvalidateAll order, and one
channel per endpoint via a new CachedChannelCount.
The_known_cluster_set_is_not_published_by_the_other_two_modes is
replaced — its premise is obsolete — by three facts: every mode
publishes, gate set is a strict superset of the fan set, and the handler
ordering.
ControlPlane.Tests 230 -> 236, Host.Tests 92, Cluster.Tests 184,
0 warnings.
Phase 1A security register items 1 and 2, landed ahead of the command
fan-out because they are independently testable and the fan-out is the
first caller of TryGetPsk.
Item 1 — known-cluster gate. MeshPskProvider gains SetKnownClusters, an
atomic whole-set swap consulted by TryGetPsk BEFORE the config map and
before the secret store. A cluster id outside the set is refused in O(1)
with no store read and no audit write, which closes two things at once:
the pre-auth store/audit flood an unauthenticated caller could drive with
a forged x-otopcua-cluster header, and the timing oracle that defeated
CentralCommandAuthInterceptor's uniform denial text (unknown clusters are
never cached, so they paid a store read on every probe while known ones
answered from memory). Empty or never-published leaves the gate inactive
— central publishes the set from its DB contact refresh, and 'the DB told
me nothing' must not become a fleet-wide authentication outage during
admin boot.
Item 2 — bool Matches(clusterId, presented) on IMeshPskProvider, with
FixedTimeEquals inside the provider. CentralCommandAuthInterceptor now
calls it and never holds key material; its two provider-side denial logs
collapse into one (the provider answers both as false) and the operator
hint is restated there.
No production caller publishes a cluster set yet, so behaviour is
unchanged until the fan-out commit wires the contact refresh.
ControlPlane.Tests 190 -> 205, Host.Tests 91 -> 92.
`Security:Acl:Enabled` (default false) makes scoped `NodeAcl` grants decide
data-plane authorization at the three seams that have one, and gives HistoryRead
a check it never had. Read, Browse and Subscribe stay ungated — stage 6.
The ACL decision REPLACES the server-wide role check rather than chaining with
it. Chained, a grant would have to confer the role as well as pass the scope
check, and only Cluster-scoped grants confer a role — so every narrow grant
would have stayed inert and the flag would have looked inert too. The
load-bearing test is an ALLOW: a session carrying no roles at all writing
successfully on the strength of its grant. A deny-only test passes against a
build where the ACL is never consulted, because the role gate denies too.
Found en route: the principal never reached the session. The authenticator
passed it as OpcUaUserAuthResult.DisplayName, but RoleCarryingUserIdentity
derives DisplayName from the presented token — for a key-bearing client that is
the static username `apikey`, identical fleet-wide. So the trie's matching key
was dropped at the identity boundary, and the audit ClientUserId on a rejected
write recorded `apikey` for every consuming system. Principal and ApiKeyId are
now carried explicitly.
Beyond the stage as sketched:
- Revocation reaches an OPEN session. Every decision re-checks the
authenticating key against the current snapshot; without it a key revoked by a
deployment kept its authority until the client chose to reconnect, which for a
machine-to-machine consumer can be indefinitely.
- UserAuthorizationState is built per CALL, not per session. The per-session
shape exists so an LDAP session can re-resolve membership that changed out of
band; nothing can change out of band under API keys, so the generation stamp is
the whole invalidation signal and a per-call record picks it up for free.
IsStale/NeedsRefresh therefore cannot fire here and are documented as vestigial.
- The trie is built inside DataPlaneAuthorityState.Apply, under the same
generation as the snapshot, rather than by a second wiring at the publish
actor. An uninstalled trie denies everything — indistinguishable from working
enforcement. Deleting the install turns 13 of the 22 new tests red.
- The deploy log reports what the BUILD dropped instead of re-deriving it. The
old count agreed with the builder on one of its four drop reasons; a rule with
a blank ScopeId, a retired scope kind, or a path not starting at a realm
vanished with the log reading zero.
Both IApiKeyStore implementations widened to the full grant set. The narrowing
was redundant — ApiKeyInterimRoleMapper applies the same filter itself — and
became a trap once scope was enforced. Pinned by asserting the roles are
unchanged, with the two grants carrying different flags so it cannot pass by
accident.
Falsified by mutation: enforcement never engaging, the write seam ignoring the
ACL, the realm dropped from the scope, the revocation check removed, and the
trie never installed each turn the intended tests red and no others.
Full solution suite green.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
Makes the /raw device tree addressable by an ACL rule, and builds the
scope-path producer PermissionTrieBuilder has accepted since it was written
but which had no caller anywhere.
The plan for this stage said "add NodeHierarchyKind.Raw, whose path is the
RawPath segments Folder/Driver/Device/TagGroup/Tag". That shape is wrong.
RawFolder.ParentRawFolderId and TagGroup.ParentTagGroupId are both
self-referencing, so folders and tag groups each nest without bound and a real
RawPath is folder…/Driver/Device/group…/Tag. NodeScope was fixed-arity and
PermissionTrie.CollectMatches was five hardcoded steps, so the scope model
could not address a real raw tree at all. NodeScope is now {ClusterId, Kind,
Segments} and the walk is a segment loop. The trie STRUCTURE was already
variable-depth (NodeAclPath.Segments is a list, Children is a recursive
dictionary), so only the walker and the scope record had to change.
Four decisions worth keeping:
- The namespace level became the realm. Level 1 gated on NodeScope.NamespaceId,
but v3 has no Namespace entity — the snapshot was retired when the address
space went dual-namespace, and the identifier appears nowhere in production
outside NodeScope. So the level could only ever fail to match, which silently
made every sub-cluster grant unreachable while cluster grants kept working.
It is now a realm token derived from the scope's Kind, not carried on it, so
the walk and the producer cannot disagree about its spelling.
- Provenance comes from the granting row's own ScopeKind, recorded on TrieGrant
at build time. Depth-derived provenance could not name a raw level at all,
since the depth of "Device" varies with folder nesting.
- The trie comparer moved from OrdinalIgnoreCase to Ordinal. Raw sibling-name
uniqueness is ordinal, so Speed and speed can both exist under one device and
a grant on either would have matched both.
- The null-scopePaths "deterministic test mode" is deleted. It planted
unresolvable rows at a direct child of the trie root, a position the walker
cannot reach, so the grant looked placed and could never match — and every
test using that mode proved nothing about production placement. scopePaths is
now required, and a row that does not resolve is dropped with a diagnostic.
A scope resolves to a LIST of paths: a raw tag projected into N equipments is
reachable at N+1 NodeIds and a write through any of them lands on the same
driver reference, so a Tag-scoped grant covering only the raw node would be
bypassable through a UNS one. The lookup is keyed on (ScopeKind, ScopeId), not
ScopeId alone — ScopeId has no FK and no type discriminator, so an id collision
across entity types would hand a grant the wrong node's paths.
Also closed here, all found by review rather than asked for:
- NodeAcl rows had no deploy-time validation of any kind and were not reachable
from DraftValidator at all (DraftSnapshot carried no NodeAcls). They are now
gated for dangling ids, cross-cluster scopes, blank/contradictory ScopeId, the
retired FolderSegment kind, and undefined ordinals.
- ConfigDbApiKeyStore conferred roles from Cluster-scoped grants authored in ANY
cluster, because it filtered on scope kind with no ClusterId predicate while
ArtifactApiKeyStore read an already-cluster-filtered snapshot. On a fused
admin+driver node that was a cross-cluster over-grant. Its doc comment claimed
the two stores matched.
- The two AdminUI ACL banners still said enforcement was LDAP-group -> role
mapping, which stages 1/2 made false in both directions.
- The /apikeys page was never classified in PageAuthorizationGuardTests, so that
test has been red since the key-minting commit merged — I did not run
AdminUI.Tests before merging stage 4.
Falsified rather than assumed: 8 mutations against the trie tests, 5 against the
deploy gate (including one proving a cross-cluster row is not silently
mis-diagnosed as unresolved), a staged red->green for the over-grant regression,
and the realm gate, the comparer and the malformed-path guard each broken
individually to confirm only their own tests go red.
Node ACLs are still not ENFORCED — nothing constructs a NodeScope or calls the
evaluator yet. That is stages 5 and 6; this stage makes the model able to
express what those will enforce.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
A driver-only node has had no ConfigDb since per-cluster mesh Phase 4, so it
registered NullOpcUaUserAuthenticator and denied every UserName token. Its keys
now ride the deployment artifact instead.
- ConfigComposer snapshots ApiKeys alongside the NodeAcls it already carried.
Only the peppered HMAC ships, never a secret, and the pepper is node config
that is deliberately not in the artifact.
- DeploymentArtifact.ParseDataPlaneAuthority decodes both node-side. Grants are
scoped to the node's own cluster; keys are not, because ApiKey carries no
ClusterId — a key is a fleet-wide credential and the ACL rows are what bind it.
- OpcUaPublishActor installs them on DataPlaneAuthorityState on every apply,
BEFORE its plan.IsEmpty early-return: an artifact that only revokes a key
produces an empty address-space plan, so feeding after that return would leave
the revoked key working until an unrelated structural change redeployed.
- ArtifactApiKeyStore serves them to session activation; a node with a ConfigDb
keeps ConfigDbApiKeyStore, whose per-request read is what makes revocation
immediate there.
- OtOpcUaNodeManager gains DataPlaneAuthority — the cluster identity blocker #4
named. NodeScope.ClusterId is required and the evaluator fails closed on a
mismatch, so until now no scope it built could have evaluated. Nothing reads it
until stages 5/6.
The generation is a node-local COUNTER driven by the RevisionHash, not derived
from its bits. PermissionTrieCache treats the highest GenerationId as current, so
a generation taken from hash bytes would install a lower number about half the
time and the cache would go on serving the previous deployment's ACLs, silently.
Sessions are not stamped with it yet — that belongs with the first seam that
reads a UserAuthorizationState, in stage 5, rather than being built now and
consulted by nothing.
Live-gated on docker-dev against site-a-1 (driver-only, no ConfigDb): key
authenticates, wrong secret denied, both halves of the revocation asymmetry
observed, and a SITE-A grant confirmed not to reach site-b-1. Key minted with an
independent Python HMAC so the verifier was checked against something other than
itself. Rig left clean.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
Anonymous sessions used to short-circuit in HandleImpersonation — "let the SDK's
default validation stand" — so they always connected and always carried no roles.
Now admission and authority are separated:
admission -> OpcUa:AllowAnonymous, the HOST's decision (it owns token policies)
authority -> the authenticator, resolving the reserved principal "anonymous"
through the SAME grant path a key's principal goes through
So anonymous can be granted WRITE, not merely tolerated as read-only, and it
arrives at the downstream gates as the same RoleCarryingUserIdentity a key
produces — which is the real payoff: no gate carries an anonymous-shaped branch,
so there is no bypass path to get wrong.
AllowAnonymous defaults to TRUE on purpose. Defaulting it false would silently
deny every client without a key the moment this deploys, including ScadaBridge's
Data Connection Layer in another repo. Making keys mandatory is a rollout
decision, not an upgrade side effect. When false, the Anonymous token policy is
withheld so a client is refused at DISCOVERY rather than only at activation, and
activation is rejected independently in case a client attempts a token type the
endpoint never advertised.
Two fail-closed choices, both falsified:
- A failure to read anonymous's grants DENIES rather than admitting with no
roles. Admitting is the pre-#520 behaviour, and defaulting to it exactly when
the server cannot determine what anonymous may do stops being harmless once
reads are gated. Making it admit reds Anonymous_denies_when_grants_cannot_be_read.
- A key may not authenticate as the reserved "anonymous" principal, or a
credentialled client silently inherits whatever anonymous was granted.
AuthenticateAnonymousAsync is a DEFAULT interface implementation returning
admit-with-no-roles, i.e. exactly the pre-#520 behaviour, so
NullOpcUaUserAuthenticator and every test double keep working rather than being
forced to reimplement it — important because driver-only nodes run the Null one
and would otherwise black out. Note a default interface method is only reachable
through the interface-typed reference, which is how OpcUaApplicationHost holds it.
An existing test asserted the old fall-through and correctly went red; it is
rewritten to pin the new contract rather than deleted, with the reason recorded.
OpcUaServer 366, Host 42 green.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
Stage 2a. Clients present the static username "apikey" with an
otua_<keyId>_<secret> credential as the password; the key resolves to a
principal. LDAP registrations are deliberately RETAINED for AdminUI login —
only IOpcUaUserAuthenticator moves, which is what D2 asked for.
The load-bearing decision here is what roles a key gets. The two live gates
(WriteOperate on Value writes, AlarmAck on Part 9 methods) check a SERVER-WIDE
role set, so the cutover has to keep producing one or it would lock every client
out of writes long before per-node evaluation lands in stage 5.
ApiKeyInterimRoleMapper therefore maps ACL grants to roles, but ONLY
cluster-scoped ones. The obvious version — union the permission flags across all
of a principal's grants — would let a grant authored narrowly ("WriteOperate on
Equipment X") confer server-wide write authority, because nothing in the pipeline
can enforce scope yet. That fails in the dangerous direction: the operator's
narrow intent silently becomes broad. Restricting to cluster scope under-grants
instead, which is no worse than today (where every ACL row does nothing) and can
never exceed what was authored. Falsified: deleting the scope check reds exactly
A_narrower_scoped_write_grant_confers_nothing_yet and nothing else.
Verification is one indexed lookup plus one HMAC, with no network I/O at all.
That matters because the SDK raises ImpersonateUser synchronously while holding a
SessionManager-wide event lock, so this path does not stall one request — it
serializes every session activation server-wide. It is the main practical gain
over the LDAP bind it replaces.
Every rejection returns one opaque message so the server cannot be used as an
oracle for whether a key id exists or which half of a credential was wrong; a
test asserts the failure modes really are indistinguishable AND that a valid key
still succeeds, so it cannot pass by denying everything.
Caught while wiring: OtOpcUaServerHostedService takes IOpcUaUserAuthenticator as
a REQUIRED ctor parameter, so leaving a driver-only node unregistered is not a
quiet fallback to deny-all — it fails DI and the node does not boot.
NullOpcUaUserAuthenticator is now registered explicitly there. A driver-only node
has no ConfigDb since Phase 4 and so has no key source until the artifact-backed
store lands in stage 4; it denies UserName tokens until then, which is the safe
direction, and anonymous sessions are unaffected.
LdapOpcUaUserAuthenticator is now registered nowhere and is dead code. Deleting
it plus LdapOpcUaUserAuthenticatorTests / LdapAuthResilienceTests /
LdapGlauthOutageLiveTests is stage 2b, kept separate because those cover a
live-gated resilience feature (hard wall-clock bound, bounded concurrency,
directory-outage circuit) that exists precisely BECAUSE the LDAP bind ran inline
on session activation — a constraint the key path does not have. That deserves
its own commit, not a footnote in this one.
Anonymous-as-a-grantable-principal (D3) is also stage 2b.
Host 37, Security 89, OpcUaServer 364 green.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
`IHostConnectivityProbe` was a dead surface: eleven drivers implement it, `GetHostStatuses()`
had ZERO production call sites, and `OnHostStatusChanged` had no subscriber outside the Galaxy
driver's own aggregator. Per-host connectivity was computed by every driver and read by nobody.
The issue offered "build the publisher or delete it". The publisher as its entity doc described
it — driver nodes upserting `DriverHostStatus` rows — is not buildable: per-cluster mesh Phase 4
gates `AddOtOpcUaConfigDb` on the `admin` role, so a driver-only node has no ConfigDb connection
to write rows with. So the capability is kept and the transport changed.
`DriverHealthChanged.HostStatuses` now carries the probe result to `/hosts` as a Hosts column.
That channel already reached the page, already survives the mesh split via the Phase 5 gRPC
telemetry stream, and already replays a last-value snapshot on re-subscribe — so per-host state
re-primes after a reconnect without a durable store. Both halves of the interface finally do
what they are for: the event triggers a prompt publish, the pull is the source of truth.
The point of the column is the case the driver-level state chip structurally cannot express: a
multi-device driver stays aggregate-Healthy while ONE of its devices is unreachable.
Two traps, both pinned by tests that were falsified against the prod code:
- The host digest MUST be in the publish fingerprint. On a single-host-down transition every
other fingerprint component is unchanged, so the dedup would swallow exactly the publish
carrying the news — the trap that already bit the rediscovery signal. Removing it turns the
guard test red, verified.
- null (no probe) must stay distinct from empty (probe with no hosts). proto3 cannot tell an
absent repeated field from an empty one, hence the explicit `has_host_statuses` flag;
collapsing them would render every probe-less driver as one whose devices are all fine.
Dropped: the DriverHostStatus entity, enum, DbSet, model config and table (migration
DropDriverHostStatusTable — empty on every deployment, so the scaffolder's data-loss warning is
moot, and Down() recreates it exactly).
Found en route, NOT fixed here: `DriverInstanceResilienceStatus` is the identical defect — no
writer, no reader, only a DbSet declaration, while the live data rides the
`driver-resilience-status` telemetry channel. Its doc-comment now states that rather than
describing the sampler and AdminUI join that were never built. Filed as #524 rather than widening
this schema change beyond what was asked.
Claude-Session: https://claude.ai/code/session_015p7wGqy3YpZNCpDzTpGMKo
ClusterPrimaryHealthCheck was written three days' worth of debugging ago because
the shared ActiveNodeHealthCheck selected by RoleLeader and reported Healthy for
any node lacking the role. Health 0.3.0 fixes the shared check — oldest Up
member, role-preference scoping, Unhealthy when the node owns no active work —
so the private copy is now duplication rather than a workaround, and it is
deleted.
RolePreference [admin, driver] reproduces exactly what it did: a fused central
node answers for admin (where the singletons and the AdminUI are pinned), a
driver-only site node answers for driver.
SelectOldestUpMemberOfRole now delegates to the shared ClusterActiveNode instead
of re-implementing the age ordering. This is the point of the change rather than
a tidy-up: the redundancy snapshot drives the OPC UA ServiceLevel 250/240 split
while the health tier drives Traefik's admin routing, so two copies of "oldest Up
member of a role" would be two chances to advertise one node as authoritative
while gating the data plane on another. ControlPlane takes a ZB.MOM.WW.Health.Akka
reference for it; the layering trade-off is recorded at the PackageReference.
Behaviour note: a node carrying neither admin nor driver now answers 503 on the
active tier instead of 200. That case is reachable — RoleParser admits dev-only
and cluster-role-only nodes — and 503 is correct, since such a node owns no
active work and must not be in an active-tier pool. No docker-dev node is
affected: every rig node is admin+driver or driver-only.
Tests replaced in kind. The rule itself is now pinned in the library against a
real two-node cluster; what stays OtOpcUa's own decision is the wiring, so
ActiveTierRegistrationTests pins which check is on the active tag, that it is
registered on driver-only nodes too, and that it is not on the ready tier.
Verified: Host.Tests 25/25, ControlPlane redundancy 15/15.
Closes the remaining half of #494, and corrects the half fixed in 8dd9da7d.
The shared ActiveNodeHealthCheck(role: "admin") answered the wrong question twice
over. It returns Healthy for any node LACKING the role, so all four driver-only
site nodes called themselves active and no consumer could find the Primary of a
site Cluster. And it selects by RoleLeader - the lowest-ADDRESSED member - which
is not where Akka places singletons.
Replaced with ClusterPrimaryHealthCheck ("cluster-primary"). One rule: this node
is active iff it is the OLDEST Up member carrying its own active role, where the
active role is admin when the node has it and driver otherwise. That serves both
consumers correctly - a fused admin node answers for admin, so Traefik pins the
AdminUI to the node hosting the singletons; a driver-only site node answers for
driver, which is exactly SelectDriverPrimary, the same election behind
IsDriverPrimary and the OPC UA ServiceLevel 250/240 split. Per-mesh scoping is
free after Phase 6: ClusterState.Members already contains only this node's own
Cluster.
SelectDriverPrimary is generalised to SelectOldestUpMemberOfRole so the
age-ordering rule has one implementation. Two copies of "oldest Up member of a
role" would be two chances to silently disagree about who is in charge, and the
tier and the redundancy snapshot must never disagree.
THE ROLE-LEADER HALF MATTERED IN PRACTICE, not just in theory. On the rebuilt rig
the two orderings diverge right now:
akka leader (lowest address) = central-1
oldest admin member = central-2 <- hosts the singletons
8dd9da7d made the tier a real 503 but still selected by RoleLeader, so it would
have pinned Traefik to central-1 - the node NOT hosting the work. With this change
Traefik correctly routes to central-2.
Live-verified, exactly one 200 per mesh:
MAIN central-2 200 central-1 503 traefik: central-2 UP, central-1 DOWN
SITE-A site-a-1 200 site-a-2 503
SITE-B site-b-1 200 site-b-2 503
Tests: role-selection matrix and the startup-safe Degraded path in Host.Tests
(26 pass); parity between the tier's selector and the redundancy snapshot's, plus
role scoping, added to RedundancyPrimaryElectionTests, which forms a real two-node
cluster deliberately built so the oldest member is not the lowest-addressed one
(5 pass).
Joseph Doherty