C1: SqlDriver's Initialize/Read failure paths no longer interpolate the ADO.NET
provider's ex.Message into LastError, the thrown message, or host status — only
ex.GetType().Name reaches the operator surface; the full exception goes to the log
sink via the structured logger's exception parameter. The driver is dialect-agnostic
over an arbitrary DbProviderFactory, so it cannot rely on any provider's message
discipline (Microsoft.Data.SqlClient's parser echoes an unrecognised keyword
lower-cased, which a value-based redactor misses).
C1a: replace the vacuous credential test (SQLite's "unable to open" message never
contains the connection string, so it passed regardless) with one that fabricates a
DbException whose own .Message carries a credential token and drives it through the
Initialize liveness-failure path via the injectable factory seam — red against the
pre-fix ex.Message interpolation.
I1: HandlePollError now ignores the DbException class ReadAsync already classified,
so a subscribed-read outage is classified (and logged) once, not twice with the
worse message.
I2: the poll's Healthy verdict routes through a shared SetHealthUnlessFaulted guard
(the one Degrade already used), so a late in-flight poll cannot un-fault a driver a
concurrent ReinitializeAsync just faulted.
I3: DiscoverAsync warns when materializing an omitted-type tag as String, the only
operator signal for the declared-vs-published type mismatch on a numeric column.
M1: BuildTagTable wraps the per-entry parse so a stray non-JsonException throw skips
the tag instead of stranding health at Initializing (it runs before Initialize's
try/catch). M2 left as a documented TODO to avoid touching SqlPollReader.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
SqlTagConfigModel + SqlRowSelectorModel mirror SqlEquipmentTagParser's
accept/reject boundary byte-for-byte on the fields they own: model/type
serialize as NAME strings (never numbers), KeyValue requires
table+keyColumn+keyValue+valueColumn, WideRow requires table+columnName+a
selector (where-pair OR topByTimestamp), Query is rejected, and unknown keys
(top-level and nested rowSelector) survive load->save. Registered in
TagConfigEditorMap + TagConfigValidator keyed off SqlDriver.DriverTypeName
(DriverTypeNames.Sql is deferred to Task 11). A minimal placeholder
SqlTagConfigEditor.razor lands the map entry now; Task 20 fleshes out the UI.
The load-bearing test rounds editor output back through
SqlEquipmentTagParser.TryParse (editor-output <=> parser-input agreement).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Its doc said it pinned "the real provider's cancellation path", but the
token is cancelled before ReadAsync so the OCE is raised at the reader's
SemaphoreSlim.WaitAsync gate, before any SqlConnection opens — it never
exercises Microsoft.Data.SqlClient's in-flight command cancellation.
Softened (option (a)) to state accurately what it proves: an already-
cancelled token propagates as OCE and is never swallowed into a Bad
snapshot or an unreachable-database verdict. The in-flight/deadline path
is covered by SqlBlackholeTimeoutTests.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Adds the frozen-peer / BadTimeout gate for the Sql driver — the single
highest-value integration test. Mirrors the S7 R2-01 blackhole gate:
docker pause a DEDICATED mssql mid-poll, assert the next read surfaces
BadTimeout within the client-side operationTimeout (≈3s) and well below
the server-side CommandTimeout backstop (30s), the driver degrades
(Degraded + host Stopped), and docker unpause recovers to Healthy/Running.
- Docker/docker-compose.yml: disposable `otopcua-sql-blackhole` mssql on
:14333 (never the shared :14330 ConfigDb server) + one-shot seed.
- Docker/seed.sql: the two sample tables.
- SqlBlackholeTimeoutTests.cs: env-gated (SQL_BLACKHOLE_ENDPOINT); pauses
ONLY the hard-coded container name; skips loudly if the endpoint is the
shared port 14330; bounds its own wait so a wedged impl fails, not hangs.
Offline: clean skip (no socket, no docker shell-out).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Locks the driver's injection guarantee against the SQLite fixture. An
authored VALUE ('; DROP TABLE TagValues; --) binds as a DbParameter and can
only ever be a key that matches no row (BadNoData); the seed table survives.
A hostile IDENTIFIER in table/column position is dialect-quoted into a
single nonexistent identifier — inert — and the payload never executes.
Scope, stated plainly: design §8.1 also specifies a catalog gate (validate
an authored identifier against INFORMATION_SCHEMA, reject an unknown one as
BadNodeIdUnknown). That gate does NOT exist in the driver yet and no task
here builds it, so a hostile identifier is not rejected up front — it is
quoted, the query fails after the connection opened, and the tag Bad-codes
as a query failure (BadCommunicationError), not BadNodeIdUnknown. These
tests assert what the code actually guarantees — the payload is inert and
the table intact — rather than a catalog gate that isn't there. The gate is
a tracked follow-up.
No implementation change: the reader already binds correctly (Task 7); this
suite pins it. Falsifiability-checked — forcing a real DROP before the
row-count assertion turns it red, so "table survives" is load-bearing.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Register SqlDriverBrowser as an IDriverBrowser beside the OpcUaClient/Galaxy
lines, project-reference Driver.Sql.Browser, and add SqlAddressPickerBody.razor:
a DriverBrowseTree schema/table/column picker (DriverOperator-gated) with a
column attribute side-panel, manual-entry model/selector fields retained, that
composes the per-tag TagConfig blob (KeyValue / WideRow) matched to
SqlEquipmentTagParser. Pasted ad-hoc connection strings are session-only and
never persisted into the composed blob or driver config.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
The AdminUI "Test Connect" probe for the Sql driver: open a connection, run
the dialect's LivenessSql (SELECT 1) under a linked-CTS deadline, return
green with latency or red with a reason. Implements IDriverProbe; never
throws (malformed JSON, missing connectionStringRef, unprovisioned
connection string, a provider open failure, timeout, cancellation all
become red results).
Parses the SAME factory DTO with the SAME JsonStringEnumConverter as
SqlDriverFactoryExtensions (R2-11 factory parity, mirroring
ModbusDriverProbe) and resolves connectionStringRef the same way, so a
config that Test-Connects is the config that Deploys.
Credential hygiene: the resolved connection string never reaches the result
message. A provider exception can embed the data source in its OWN message
(the real SqlException shape), so the catch-all names the exception TYPE
only, never ex.Message. The regression test's fake connection embeds the
connection string in its thrown message specifically so surfacing ex.Message
would leak it — verified load-bearing by breaking the guard and watching it
go red.
ForTest injects factory+dialect for the offline SQLite fixture path.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
SqlDriverFactoryExtensions turns a deployed DriverConfig blob into a live
SqlDriver, and SqlConnectionStringResolver resolves the authored
connectionStringRef NAME from Sql__ConnectionStrings__<ref> so credentials
never ride in a config blob (design §8.2).
Three behaviours worth naming:
- String-enum guard. JsonOptions carries JsonStringEnumConverter +
camelCase, so `provider` parses whether authored as a name or an ordinal
and always round-trips as the NAME — the systemic AdminUI defect where a
page serialised an enum numerically against a string-typed DTO.
- Config validation lands here because nothing below does it.
operationTimeout must be STRICTLY greater than commandTimeout (design
§8.3); the reader and the driver stay deliberately usable with the pair
inverted so the frozen-database tests can prove the client-side bound
fires. Non-positive timeouts / poll interval and maxConcurrentGroups < 1
are rejected too.
- Credential hygiene. The resolved connection string reaches the provider
and nothing else: messages name the ref, the environment variable, or
SqlDriver.Endpoint's credential-free server/database rendering. Both a
log-leak and an exception-leak test cover it.
DTO changes:
- Adds RawTags (List<RawTagEntry>), following the Modbus precedent — the
deploy artifact delivers a driver's tags this way and the DTO had no way
to receive them, so an authored Sql tag could never reach the driver.
- Deletes SqlProbeDto + the `probe` key. It had no consumer: SqlDriver was
specified without a background probe loop, and deliberately so — its
IHostConnectivityProbe state is a by-product of the Initialize liveness
check and of every poll, i.e. a statement about traffic that actually
happened rather than a synthetic ping. On-demand connectivity is Task
10's SqlDriverProbe. Shipping a config key that silently does nothing is
worse than not shipping it; UnmappedMemberHandling.Skip means a blob that
still carries `probe` keeps parsing.
allowWrites stays inert (SqlDriver implements no write capability) but an
authored `true` now WARNS rather than passing silently — the flag cannot do
harm, an operator who believes writes are enabled can.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Opens one transient DbConnection from a form-supplied driver-config blob and
hands ownership to SqlBrowseSession, which closes it on disposal. A database is
named either by connectionStringRef — resolved in the AdminUI process from
Sql__ConnectionStrings__<ref>, with an unresolvable ref naming the exact missing
variable — or by a pasted, session-only literal. The literal wins and the ref is
then not read; it cannot arrive from a persisted blob (SqlDriverConfigDto has no
such property), so its presence proves an operator typed it now.
Credential hygiene is the point of the type: no cached config field, nothing
persisted, and no connection text in any log line or exception. Live-probed:
Microsoft.Data.SqlClient and Microsoft.Data.Sqlite keep connection-string values
out of SqlException/SqliteException, but their connection-string PARSER echoes an
unrecognised keyword verbatim — and an unquoted ';' inside a password splits, so
the tail of the password is reported as a keyword (Password=Sup3r;SecretTail =>
"Keyword not supported: 'secrettail;connect timeout'."). The parser's message is
therefore never surfaced; every other failure additionally passes through a
substring redactor that drops the inner exception when it fires.
DriverType comes from SqlDriver.DriverTypeName, the driver's interim local
constant — DriverTypeNames.Sql is added by the driver-factory task.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Three review findings against SqlPollReader (967e5140).
1. WideRow's where-pair form emits no ORDER BY and no row limit, so a
selector that is not actually unique made the reader publish an
arbitrary, storage-order-dependent row — silently, unlike IndexByKey's
duplicate-key warning. Row selection moves out of MapMember into a new
SelectWideRow, which is called once per group and emits the same
rate-limited contract warning (naming table, selector and row count).
The row-limit half of the suggested fix was deliberately NOT taken: a
TOP 1 with no ORDER BY is not deterministic either, and it would
destroy the Rows.Count > 1 evidence the warning depends on, turning a
loud misconfiguration into a permanently silent one. Rationale is in
SelectWideRow's docs. No SQL changed, so no planner golden string moved.
2. The "a frozen database can never accumulate more than
maxConcurrentGroups connections" claim was asserted in the docs and
untested. ReadAsync_aTimedOutGroupKeepsItsConcurrencySlotUntilTheQuery-
TrulyFinishes wedges a group behind an EXCLUSIVE transaction, times it
out, and proves the slot is still held (a second poll on a cap of 1
times out on the gate without ever reaching the factory) and is handed
back only once the abandoned query truly finishes. Moving the release
from the work to the waiter fails this test and no other.
3. Task.Run thread-pool starvation: documented, not changed.
Microsoft.Data.SqlClient's async path parks no pool thread on a frozen
server, so the driver cannot starve itself; LongRunning would cost a
dedicated OS thread per group per poll forever to insure against a
provider this driver does not ship. RunGroupAsync now records the
decision and what a BadTimeout does and does not mean, for the
blackhole gate.
Also: the timeout warning now names the group's table, so a BadTimeout
storm identifies which source is frozen.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Walks schemas -> tables/views -> columns via ISqlDialect's catalog SQL, with
@schema/@table bound as parameters at every level. Column leaves carry the
dialect-mapped DriverDataType in the attribute side-panel (ViewOnly, read-only
v1).
The NodeId encoding deliberately departs from the design sketch's literal
`schema.table|column`: SQL Server permits `.` and `|` inside a quoted
identifier, so that form mis-parses (main.a.b|c reads equally as schema
`main`+table `a.b` and schema `main.a`+table `b`) and silently binds an
operator's tag to the wrong column. SqlBrowseNodeId encodes
`<kind>:<part>[|<part>...]` with `\`/`|` escaped and the kind prefix carrying
the arity; it is public because the picker body decodes it back.
The session owns the connection it is handed and closes it on dispose -- the
registry-held session is the only lifetime hook, so a non-owning session would
leak one pooled connection per reaped picker. Per-call work stays bounded by
the AdminUI's existing 20s linked CTS (BrowserSessionService.PerCallTimeout);
no second deadline is invented here.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
SqlDriver implements IDriver / ITagDiscovery / IReadable / ISubscribable /
IHostConnectivityProbe / IAsyncDisposable over the landed SqlPollReader,
mirroring ModbusDriver. IWritable is deliberately absent: v1 is read-only
structurally, and every discovered variable is SecurityClass=ViewOnly.
The shell classifies poll outcomes because nothing below it does. The reader
honours IReadable literally — it throws only when the database is unreachable
and Bad-codes everything else — so a frozen database returns all-BadTimeout
snapshots through a perfectly successful PollGroupEngine tick. Without
ObservePollOutcome the driver would report Healthy while every value was Bad.
Connection-class codes (BadTimeout / BadCommunicationError) degrade health and
report the host Stopped; authoring-class codes (unresolvable RawPath, absent
row, type mismatch) change nothing, so a tag typo never reports the database
down. It deliberately does NOT synthesise an exception to earn engine backoff:
the engine's exception path publishes nothing, which would cost clients the
Bad quality the reader went out of its way to produce.
Initialize builds the authored RawPath table first (pure, cannot fail; a
malformed TagConfig is logged and skipped) then verifies liveness over one
open-use-dispose connection bounded by wall clock as well as by token (the
R2-01 lesson) — failure records Faulted and rethrows so DriverInstanceActor
retries. The connection string is never logged: a credential-free
server/database Endpoint is the only rendering that reaches a log, LastError,
or the host status.
DriverTypeNames.Sql is NOT added here — DriverTypeNamesGuardTests asserts
bidirectional parity with registered factories, so the constant must land with
the factory (Task 11). SqlDriver.DriverTypeName carries the string until then.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Code-review Minor: the record typed ParameterNames/Parameters/Members/
SelectedColumns as IReadOnlyList<T> but was handed the planner's live List<T>
instances. SqlQueryPlan documents itself as safe to hold for plan caching, so
an aliased list is a real hazard once the reader starts reusing plans across
polls: IReadOnlyList<T> is downcastable, and one mutation would corrupt every
subsequent poll served by that plan.
Copying at the record boundary fixes it for every call site at once rather
than relying on each planner branch remembering to call AsReadOnly.
Pinned by a test that mutates the caller's lists after construction; verified
load-bearing (reverting the record turns it red).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Executes a poll pass: resolve refs -> SqlGroupPlanner.Plan -> one command per
group -> slice the result set back to per-tag DataValueSnapshots, positionally
aligned with the caller's reference list (design §3.2).
The frozen-peer contract (design §8.3, the R2-01 S7 lesson) is the core of it.
CommandTimeout is only the server-side backstop; the client-side bound is a
linked CancelAfter PLUS Task.WaitAsync, because a linked token bounds only a
provider that honours it and ADO.NET providers vary. The whole per-group
operation — waiting for a concurrency slot, OpenAsync, and the query — runs off
one operationTimeout budget, so ReadAsync returns on time regardless of what the
database is doing. A breach Bad-codes that group (BadTimeout); caller
cancellation propagates instead, since nobody consumes a torn-down poll.
Also: absent row (BadNoData) stays distinct from a NULL cell (Uncertain, or Bad
under nullIsBad); duplicate plan members are all fed (two tags on one keyValue
bind one parameter but stay two members); the concurrency slot is released by the
work, not the waiter, so a frozen database can never hold more than
maxConcurrentGroups connections; the whole call throws only when the connection
itself cannot open, which is what earns PollGroupEngine's backoff.
SqlStatusCodes is driver-local, matching every sibling driver's own table — there
is no shared helper in Core.Abstractions and drivers do not reference the OPC UA
SDK. Values were read off Opc.Ua.StatusCodes rather than copied from a sibling,
because two of those tables carry transcription errors.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Three findings from the code review of the dialect seam:
- QuoteIdentifier used char.IsControl, which only covers Unicode category Cc.
Bidi overrides, zero-width spaces, soft hyphen and BOM are category Cf and
passed through untouched. They cannot escape the brackets, so this is not an
injection bypass — but the method rejects control characters precisely to
protect the integrity of logged/audited statements, and Cf characters spoof
rendered text while comparing byte-different from the real catalog name
(Trojan Source, CVE-2021-42574). Same rule, same rationale, wider category.
- The XML docs on both files asserted that identifiers 'are sourced only from
catalog-validated (INFORMATION_SCHEMA) names' and that rejection is 'a
backstop, not the primary defence'. Neither is true yet: design §8.1 does
specify that gate, but nothing implements it and no task in the plan
schedules one, so an authored TagConfig table/column reaches QuoteIdentifier
unfiltered. The docs now say so, because under-scrutinising this method on
the strength of a filter that does not exist is exactly the failure mode.
- The control-character test only pinned NUL, so a regression to a
Contains('\0') check would still have passed. Pinned the whole Cc category
plus the new Cf rule; verified load-bearing by deleting the guard and
observing exactly the 5 new cases go red.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
The dialect owns the two things ADO.NET's System.Data.Common base types
cannot abstract: identifier quoting and the metadata-catalog SQL. It is
the driver's SQL-injection boundary — values always bind as DbParameters,
identifiers cannot, so the few that reach a command text are catalog-
sourced and pass through QuoteIdentifier.
QuoteIdentifier brackets one identifier part and doubles embedded ] (the
only metacharacter that can terminate a bracketed identifier early);
rejects null/empty/whitespace, >128 chars (T-SQL sysname ceiling), and any
control character (incl. NUL) as ArgumentException, without echoing the
untrusted value into the message. MapColumnType folds the design §3.7
families case-insensitively and never throws — an unrecognised family
falls back to String so a browse over an exotic column still renders;
timestamp/rowversion and time are deliberately NOT mapped to DateTime.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Per-plan copy-paste command (subagent-driven-development in a git worktree),
plus the slash-command and executing-plans/resume alternatives, and a note that
the four independent plans can run in concurrent worktrees.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
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.
Add a living progress tracker (docs/plans/2026-07-24-driver-expansion-tracking.md)
for the driver-expansion program's Waves 0-2, linked from the program design doc's
document map. One row per deliverable pointing at design + implementation plan + status.
Add executable implementation plans (plan .md + co-located .tasks.json) for the four
Wave 1/2 drivers, each authored by writing-plans methodology off its design doc and
mirroring the relevant existing driver:
- Modbus RTU 11 tasks (extends Driver.Modbus; RTU-over-TCP, direct-serial descoped)
- SQL poll 22 tasks (ISqlDialect seam, SQL Server only in v1; SQLite + central-SQL fixtures)
- MTConnect Agent 23 tasks (browse free via Wave-0 seam; TrakHound-vs-hand-rolled = Task 0)
- MQTT/Sparkplug B 27 tasks (P1 plain MQTT 0-14 then P2 Sparkplug 15-26; MQTTnet-5 pinning spike = Task 0)
Wave 0 (universal browser) remains merged (056887d6) with its live gate open (#468).
All four new plans are design-only -> plan-ready; nothing built yet.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
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).
0.2.1 makes the shared ActiveNodeHealthCheck return Unhealthy (503), not
Degraded (200), for a node that carries the role but is not the role leader.
That is what the shared health spec always specified, what docs/ServiceHosting.md,
docs/v2/Architecture-v2.md and the 2026-05-26 alignment design all describe
("200 only on the Admin-role leader; 503 elsewhere"), and what
docker-dev/traefik-dynamic.yml + scripts/install/traefik-dynamic.yml assume when
they use /health/active as the load-balancer probe.
The implementation returned Degraded, which MapZbHealth maps to 200, so BOTH
central nodes always passed the probe and Traefik load-balanced the admin UI
across the pair instead of pinning it to the leader. The docs were right; the
code was wrong. Closes the Traefik half of #494.
No OtOpcUa code change — package pin only.
Live-verified on docker-dev, with the MAIN pair genuinely formed (memberCount 2,
both members naming one leader):
central-1 (leader) /health/active -> 200 traefik UP
central-2 (standby) /health/active -> 503 traefik DOWN
Failover drill (README step 2), which had never actually exercised anything:
stopping central-1 flipped Traefik to central-2 within 20 s and
http://localhost:9200/ answered 200 continuously across the swap — no outage.
Restarting central-1 reclaimed the role-leader and Traefik flipped back.
README corrected on two counts: the drill now states that exactly ONE backend is
UP by design, and step 3 no longer claims central-2 keeps leadership — RoleLeader
is the lowest-address member, so central-1 deterministically reclaims it. Added
the cold-boot caveat that starting both nodes at once can leave each self-forming
a 1-member cluster, in which case both are their own leader and both answer 200.
STILL OPEN in #494: driver-only nodes. The check is scoped to the admin role and
returns Healthy for any node that lacks it, so all four site nodes answer 200 and
the real per-Cluster redundancy Primary (oldest Up driver member, IsDriverPrimary)
is still not what this tier reports.
Picks up the additive per-entry `data` object in the canonical health JSON and
the cluster-view data the shared AkkaClusterHealthCheck now publishes (leader /
selfAddress / selfRoles / memberCount / unreachableCount).
No OtOpcUa code change is needed: the "akka" ready/active check is already the
shared ZB.MOM.WW.Health.Akka.AkkaClusterHealthCheck (Health/HealthEndpoints.cs),
so /health/ready starts carrying entries["akka"].data.leader on the bump alone.
Payloads from every other check are byte-identical — data is emitted only when a
check publishes some.
Prereq for the family overview dashboard (scadaproj
docs/plans/2026-07-22-overview-dashboard-impl-plan.md, Phase 2).
Verified: Host 21/21, AdminUI 692/692, Host builds 0 warnings.
Live rig check (admin node serves data.leader) deferred to the dashboard's
acceptance run.
The simultaneous-cold-start split-brain, carried as a Phase 7 "candidate follow-up,
not shipped", is now closed by the opt-in Cluster:BootstrapGuard (279d1d0f). Documents
the guard in CLAUDE.md's bootstrap section and marks the Phase 7 finding closed.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Closes the residual Phase-7 gap: when BOTH nodes of a 2-node pair cold-start at the
same instant, each self-first seed runs FirstSeedNodeProcess, times out waiting for
the other, and forms its own 1-node cluster — two Primaries in one pair (the Phase 6
live gate reproduced this reliably on docker). The prior mitigation was operational
(staggered start / compose depends_on), which does not exist on production hardware.
The guard (dark switch Cluster:BootstrapGuard:Enabled, default OFF) prevents the split
without giving up cold-start-alone:
- The lower-address node is the preferred founder: self-first, forms immediately.
- The higher node probes its partner's Akka port (TCP connect) up to PartnerProbeSeconds:
reachable => peer-first (join the founder, never race it); unreachable => self-first
(partner is dead, form alone). The order is decided BEFORE the single JoinSeedNodes,
from an explicit reachability signal — never re-formed mid-handshake (the retired
SelfFormAfter failure mode). When on, Akka gets no config seeds (BuildClusterOptions)
and ClusterBootstrapCoordinator drives the join.
Review-driven hardening: case-insensitive tie-break (a hostname-casing mismatch would
reopen the split); fail-fast validation of the timing knobs; the residual "founder dies
in the probe->join window" hang is documented and made operator-visible (warning + a
restart recovers). Real-ActorSystem coordinator tests cover the load-bearing higher-node
cold-start-alone case; 147/147 Cluster tests pass.
Live-gated on docker-dev (site-a = enablement demo, serialization removed, guard on;
site-b keeps depends_on-serialization + guard off as the A/B control): simultaneous
start -> site-a-1 founds, site-a-2 probes-reachable-joins -> 250/240, NO split;
higher-node-alone -> probes dead founder, self-forms -> 250; founder rejoin -> 240.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
All 6 failover drills pass on the docker-dev three-mesh rig:
- D1 MAIN graceful manual-failover via the AdminUI "Trigger failover" button —
Primary leaves, peer→250, restarts and rejoins with no split.
- D2 SITE-A auto-down failover, both directions; D3 SITE-B crash-the-oldest.
- D4 auto-down 1-vs-1 survival — every survivor stayed Up (closes the gate
deferred since Phase 0a).
- D5 self-first cold-start-alone — a node boots with its partner down and forms
its own mesh (gate b).
- D6 recovery — restarted nodes rejoin as Secondary (240).
Ships a one-page co-located operator runbook (OtOpcUa + ScadaBridge on shared
site VMs). Manual-failover button confirmed MAIN-only by construction; site
pairs (driver-only, no UI) fail over via auto-down. Carried Phase-6 finding —
simultaneous cold-start of both pair VMs can split — mitigated operationally in
the runbook (staggered start); a product-level join guard is a candidate
follow-up, not shipped.
The per-cluster mesh program (Phases 0a–7) is COMPLETE.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
The cluster-redundancy singleton's role scope is computed at AddAkka-configurator
time and must come from AkkaClusterOptions, never by resolving IClusterRoleInfo
there (it depends on the ActorSystem being built → stack overflow at boot). Caught
by the Phase 6 live gate; fixed in 3a4ed8dd.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Records the full live-gate run against master on the docker-dev three-mesh rig.
All 8 legs pass: three isolated 2-node meshes, one Primary per pair (250/240),
deploy sealed acks=6 over per-cluster ClusterClient, telemetry gRPC :4056 up +
reconnect, reconciler own-cluster-scoped, secrets pair-local, split validator
refuses Dps, and per-pair failover with auto-down 1-vs-1 survival (also closes the
deferred Phase 0a gate). Three defects were caught and fixed forward (3a4ed8dd,
792f28ec) — see the live-gate doc's Findings section. Task 9 marked completed.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Two docker-dev-only defects the Phase 6 live gate surfaced (both invisible to
`docker compose config`, which validates the file but never boots the app):
1. Driver-only site nodes crashed at boot with OptionsValidationException — LDAP
transport None + AllowInsecure false. Only central carried the Security__Ldap__*
block, but the site nodes serve OPC UA endpoints and validate LDAP options too, and
they override `environment` wholesale (YAML `<<:` doesn't deep-merge it). Extracted
the block to a shared &ldap-env anchor merged into every host node.
2. Site pairs split-brained at simultaneous startup (both nodes "JOINING itself",
250/250 instead of 250/240). Both are self-first seeds, so both run Akka's
FirstSeedNodeProcess; the partner didn't depend on its pair founder, so they raced
and each formed a 1-node cluster. Added an &akka-founder-healthcheck (bash /dev/tcp
probe of Akka 4053 — no nc/curl in the image) to site-a-1/site-b-1 and switched the
partners to depends_on service_healthy, so the founder forms first and the partner
JOINS it. (Carry the underlying simultaneous-cold-start property to Phase 7 — it is
inherent to symmetric self-first seeding, present for central since #459.)
Live gate record: docs/plans/2026-07-24-mesh-phase6-live-gate.md (legs 1-2 PASS —
three isolated meshes, one Primary per pair at 250/240).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Phase 6 wired the re-homed redundancy singleton in Program.cs as
`WithOtOpcUaClusterRedundancySingleton(sp.GetRequiredService<IClusterRoleInfo>())`
inside the AddAkka configurator lambda. That lambda runs WHILE the ActorSystem is
being built, but ClusterRoleInfo depends on the ActorSystem (it is a live
Cluster.State view) — so resolving it there recurses into ActorSystem construction
and every node dies at boot with StackOverflowException. It compiles cleanly (a
runtime DI cycle) and RedundancyStateSingletonRehomeTests passed a hand-built
FakeClusterRoleInfo straight into the extension, mocking around the composition-root
line that overflows. The docker-dev live gate caught it on first boot.
Derive the singleton's cluster-role scope from AkkaClusterOptions (pure config, no
ActorSystem) instead: BuildClusterRedundancySingletonOptions / the extension now take
AkkaClusterOptions and compute Role = Roles.FirstOrDefault(IsClusterRole) ?? driver —
identical to ClusterRoleInfo.ClusterRole, which derives from the same config. Program.cs
passes IOptions<AkkaClusterOptions>.Value, which has no ActorSystem dependency.
Live-verified: rebuilt image boots with zero stack-overflow lines, cluster singletons
form per mesh. 5/5 rehome tests pass; Host builds clean.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Redundancy.md: KNOWN-LIMITATION (per-Akka-cluster election) marked RESOLVED;
new 'Per-cluster meshes (Phase 6)' section + pair-local secrets note; manual-
failover mesh-scope caveat removed; auto-down 1-vs-1 note updated (every mesh is
now two nodes, drill deferred to Phase 7). IManualFailoverService + ClusterRedundancy
razor caveats updated to pair-local reality. Configuration.md documents the
cluster-{ClusterId} role, per-pair self-first seeds, SplitTopologyTransportValidator,
and the intentional Dps default. Program + design status tables + CLAUDE.md marked
Phase 6 DONE.
(Task 8 subagent completed the edits but died on an API error before committing;
reviewed and committed by the controller.)
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Rewrite docker-dev/docker-compose.yml from one shared six-node gossip ring
into three self-contained pairs (MAIN: central-1/2, SITE-A: site-a-1/2,
SITE-B: site-b-1/2), all still on the ActorSystem name `otopcua` and the one
shared docker network — separation is purely by per-pair self-first
Cluster:SeedNodes, faithful to how production splits over a WAN rather than
a firewall.
Per node:
- Cluster__Roles / OTOPCUA_ROLES gain a `cluster-<ClusterId>` role
(cluster-MAIN / cluster-SITE-A / cluster-SITE-B), the marker
SplitTopologyTransportValidator and ClusterRoleInfo key off.
- Cluster__SeedNodes now lists only the node's own pair (self first, partner
second); site nodes no longer seed off central-1.
- MeshTransport__Mode default flips Dps -> ClusterClient (still overridable
via OTOPCUA_MESH_MODE) on all six nodes — SplitTopologyTransportValidator
now refuses to boot a cluster-role node left on Dps.
- Telemetry__Mode=Grpc added on all six driver nodes; TelemetryDial__Mode=Grpc
added on central-1/central-2 only (the admin pair).
- ConfigSource/ConfigServe modes and MeshTransport contact points are
unchanged (already split-correct since Phases 3/4/2).
seed-clusters.sql needed no changes — the ClusterNode rows already carry the
correct ClusterId/Host/AkkaPort/GrpcPort per node. Validated with
`docker compose -f docker-dev/docker-compose.yml config`.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Code review found SplitTopologyTransportValidator read the wrong role view,
defeating its fail-loud purpose two ways:
(a) Roles-source divergence. AkkaClusterOptions.Roles binds Cluster:Roles
but falls back to OTOPCUA_ROLES (RoleParser.Parse) when it is empty. A
node configured with only OTOPCUA_ROLES=driver,cluster-SITE-A genuinely
carries the cluster role in Akka, yet the validator saw no cluster role
and passed silently on MeshTransport:Mode=Dps (the documented incident).
(b) Case. The Cluster:Roles bind path is verbatim while RoleParser.Parse
lowercases the OTOPCUA_ROLES path; 'Cluster-SITE-A' / 'Driver' / 'Admin'
slipped past the ordinal checks.
Resolve effective roles the same way the node does — Cluster:Roles, else
RoleParser.Parse(env OTOPCUA_ROLES) exactly as Program.cs — then normalize
both (trim + ToLowerInvariant) before IsClusterRole/driver/admin. The message
still names the ClusterId in the operator's original spelling. Exemption and
all other behaviour unchanged. Contained to the validator.
Tests: OTOPCUA_ROLES-only cluster role on Dps fails; mixed-case roles on Dps
fail; cluster role with neither admin nor driver needs only ClusterClient;
Cluster:Roles wins over a stray OTOPCUA_ROLES. 16/16 green (124/124 project).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Per-cluster mesh Phase 6 (Task 5). A node carrying a cluster-{ClusterId}
role (RoleParser.IsClusterRole) is in the split topology, where the
Phase 2/3/5 DPS dark-switch branches deliver nothing across the mesh
boundary. SplitTopologyTransportValidator fails host start unless such a
node uses the mesh-crossing transports: MeshTransport:Mode=ClusterClient
always; Telemetry:Mode=Grpc if it has the driver role; TelemetryDial:Mode
=Grpc if it has the admin role. It is a no-op for any node with no
cluster-role (legacy / single-mesh / test). Roles + the three modes are
cross-read from IConfiguration, mirroring ConfigSourceOptionsValidator.
Registered via AddValidatedOptions on MeshTransportOptions with
ValidateOnStart, alongside the sibling validators.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
The per-cluster contact map only gains a key when a row parses, so an
all-malformed cluster has no key at all and is caught by the absent-key arm;
the empty-set guard is defensive/unreachable for the DB producer. Comment-only
(Task 4 code-review minor follow-up).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Task 3 review: scoping only the rows/enabled DB queries by ClusterId left
the observed set (from Cluster.State.Members) filtered by the driver role
alone. In the mixed-mesh window — DB scoping live via a cluster role, but
the physical Akka mesh not yet split so central still sees foreign members
via gossip — a foreign-cluster driver member stays in observed while its row
is filtered out, hitting the RunningNodeHasNoRow branch, which logs at ERROR.
That traded the Warning-level false positive for a worse Error-level one.
Filter observed by the node's own cluster ROLE too. Extracted the projection
into a static FilterOwnClusterDriverMembers(members, ownClusterRole) so the
membership filter is unit-testable without seeding a live cluster (the actor
harness joins as admin only, so observed is always empty). Sourced at
registration from IClusterRoleInfo.ClusterRole alongside ClusterId. Null role
(legacy, no cluster role) keeps every driver member — reconcile-all, unchanged.
Also clarified the ownClusterId doc wording (non-null-and-non-empty).
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Review follow-ups on the redundancy re-home (Phase 6, Task 2):
- Program.cs: the hasDriver comment claimed IClusterRoleInfo "throws at host
start if this driver node carries no cluster role" — false since the fallback
landed. Rewrite it to describe the cluster-{ClusterId} scope with the driver-role
fallback for legacy single-mesh / not-yet-migrated nodes.
- ClusterRoleInfo.cs: the SubscriberActor comment described RedundancyStateActor as
the "admin-role singleton" — stale. Note it is now a cluster-{ClusterId}-scoped
singleton spawned on every driver node (LeaderChanged stays a no-op here).
- Add a host/registry-level test for the driver-role FALLBACK path (roles ["driver"],
no cluster role) closing the coverage asymmetry — it was proven only in the pure
helper. Asserts the node boots and registers RedundancyStateActorKey, i.e. the boot
the earlier throw-based version would have aborted.
The boot helper now supplies an in-memory IDbContextFactory and a fake IClusterRoleInfo:
Akka.Hosting invokes singleton props factories eagerly at StartAsync, and the admin
ClusterNodeAddressReconciler factory reads both (the latter for its per-cluster reconcile
scope, added by a sibling Phase 6 task) — without them the admin boot NREs.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Joseph Doherty