Completes 835fc08c, which fixed the harness's DisposeAsync but left StopNodeBAsync
calling NodeB.DisposeAsync() directly — despite that commit's own message naming
it as carrying the same bug. Caught by re-reading the code rather than trusting
the commit message.
This one is not just about memory. IHost.DisposeAsync never invokes
IHostedService.StopAsync, so CoordinatedShutdown never runs and node B never
performs a graceful Cluster.Leave. Node A instead notices only when the failure
detector times the member out — a slower and materially different path from the
graceful leave the failover tests are written to exercise. The doc comment
asserted the opposite ("shuts down node B via DisposeAsync, which runs
CoordinatedShutdown"); it now says what actually happens and why.
Verified: the three failover suites that call it pass (5/5), and the project is
unchanged at 194 passed with only the AbCip fixture failure, which needs a docker
fixture that is not running.
Also checked the sibling harness while here: LocalDbPairHarness already stops each
host before disposing it, so it needed nothing.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
User-reported: the integration tests leave instances behind and take 18+ GB.
Measured before changing anything — a SINGLE Host.IntegrationTests process peaked
at 30,153 MB RSS across its ~200 tests.
The cause was not a leak and not parallelism. Referencing the Host drags in the
ASP.NET Core framework reference, which turns Server GC on by default: one heap
per core, tuned for throughput and deliberately reluctant to hand memory back.
Correct for a server, wrong for a test host. Workstation GC takes the same suite
from 30,153 MB to 6,717 MB — 78% less — with no change in runtime (45s).
Two hypotheses were measured and rejected on the way, recorded here so nobody
re-runs the experiment: capping xunit maxParallelThreads to 4 cost 29% wall-clock
and saved nothing (30,153 -> 31,115 MB, i.e. noise), and it turns out a single
process was doing all of it. Reverted.
The harness fix is real but was worth only 4% on its own: TwoNodeClusterHarness
disposed each node without stopping it first, and IHost.DisposeAsync only disposes
the service provider — it never invokes IHostedService.StopAsync, which is where
Akka.Hosting terminates the ActorSystem. So every test left two live ActorSystems,
remoting transports and dispatcher pools included, rooted for the process
lifetime. The class doc-comment asserted the opposite ("DisposeAsync, which runs
CoordinatedShutdown"); it was wrong, and StopNodeBAsync inherited the same bug on
a path failover tests depend on.
The payoff is larger than the memory number. Two failures I had classified as
known-flaky baseline noise now pass consistently:
RoslynVirtualTagEvaluatorTests.Evaluate_racing_ClearCompiledScripts_never_fails_with_disposed
and ContinuousHistorizationRecorderTests.Retry_after_writer_failure_eventually_acks.
They were never flaky — they were starved. Full solution: peak summed test-process
RSS 15.2 GB, zero new failures, two fewer.
Claude-Session: https://claude.ai/code/session_01GASWkNEi68FSCtvr6rLoEW
Integration-sweep follow-up #6 (SQL leg). The sweep flagged 7 heavy 2-node
deploy/failover/reconnect E2E tests as 'time out under amd64-emulated SQL' and
suggested raising deadlines. That diagnosis was WRONG: run against real
(native-amd64) SQL on the Docker host, they still hang — even at 4x deadline.
Root cause is that these tests had never actually run against real SQL, only the
EF in-memory provider, which does NOT enforce foreign keys.
Two distinct defects, both hidden by in-memory:
1. Missing FK-parent seed (5 tests). The deploy records a NodeDeploymentState row
per node, FK'd to ClusterNode (FK_NodeDeploymentState_ClusterNode_NodeId). The
tests never seeded ServerCluster + ClusterNode, so on SQL each node-state INSERT
throws and the deploy never seals (waits the full deadline regardless of size).
Added TwoNodeClusterHarness.SeedDefaultClusterAsync (seeds a ServerCluster +
a ClusterNode per node; Warm/NodeCount=2 to satisfy the SQL CHECK constraint
CK_ServerCluster_RedundancyMode_NodeCount AND the ClusterEnabledNodeCountMismatch
validator). Called it in DeployHappyPath x2 / Failover / FleetDiagnostics /
EquipmentNamespace; fixed DriverReconnect to seed node B + NodeCount=2.
2. Stale EquipmentId (EquipmentNamespace test; pre-existing on BOTH providers).
Seeded EquipmentId='eq-1', but the later-added EquipmentIdNotDerived validator
requires EquipmentId == DeriveEquipmentId(EquipmentUuid) ('EQ-'+first 12 hex).
Fixed the seed to a canonical id + matching UUID. Only surfaced once the FK fix
let the deploy reach equipment validation.
Also added OTOPCUA_HARNESS_SQL_HOST / OTOPCUA_HARNESS_LDAP_HOST overrides so the
harness fixtures can run on the native-amd64 Docker host (avoids arm64 mssql
emulation entirely).
Verified: the 5 SQL-backed classes go 11/11 green vs native SQL; in-memory default
unregressed (12/12). RoslynVirtualTagEvaluatorTests racing test is a pre-existing
flaky race (not SQL-backed; passes in isolation) — left as-is.
The Host.IntegrationTests opt-in real-LDAP mode (OTOPCUA_HARNESS_USE_LDAP=1)
used bitnami/openldap:2.6, gone since Bitnami deprecated their free image
catalog in 2025. Replaced it with GLAuth — the same LDAP server the rest of
the project already uses (docker-dev + the live OPC UA data-plane auth, both
against the shared GLAuth on :3893) — removing the lone OpenLDAP outlier and
the gone-image breakage.
- Added tests/.../Host.IntegrationTests/glauth/config.toml: baseDN dc=zb,dc=local,
a search-capable serviceaccount, dev users alice (full access) / bob (read-only),
and role groups with the same gidnumbers as scadaproj/infra/glauth so GroupToRole
maps identically.
- Compose ldap service -> glauth/glauth:latest, mounting the config, host :3894 -> :3893.
- Repointed the harness real-LDAP override from the OpenLDAP cn=admin/ldapadmin to
GLAuth's cn=serviceaccount/serviceaccount123.
Live-verified against a deployed container on :3894: the serviceaccount bind
searches and returns alice + her 5 memberOf groups; alice/bob bind with the
correct password; a wrong password yields Invalid credentials (49) — exactly the
OtOpcUaLdapAuthService flow (proven identical to the data-plane's shared-GLAuth path).
NB: real-LDAP mode is opt-in; the default Host run uses StubLdapAuthService, so
this never blocked the default suite. Integration-sweep follow-up #6 (LDAP leg);
the amd64-emulated-SQL deadline leg remains open.
The load-bearing integration half of Critical 1. Adds TwoNodeClusterHarness.HardKillNodeAAsync
(canonical in-process crash sim: shut down node A's remoting transport via
IRemoteActorRefProvider.Transport.Shutdown() — associations drop, heartbeats/gossip stop, NO
graceful Cluster.Leave, so node B's failure detector marks A Unreachable) + a
WaitForNodeBSoleDriverLeaderAsync helper (waits for a STABLE takeover: B sole Up member + driver
role-leader). HardKillFailoverTests crashes the oldest node and asserts the survivor takes over —
the exact scenario the graceful StopNodeBAsync path (Cluster.Leave, no downing decision) cannot
exercise.
Live-verified in-process (~35s: acceptable-heartbeat-pause 10s + stable-after 15s + convergence).
Negative control confirmed the test BITES: forcing explicit NoDowning
(akka.cluster.downing-provider-class = "") makes it time out with the node stuck Up-but-Unreachable.
FINDING surfaced by the negative control: removing Critical 1's typed
ClusterOptions.SplitBrainResolver did NOT break failover, because Akka.Cluster.Hosting's
WithClustering applies SplitBrainResolverOption.Default when the option is null — enabling the SBR
downing provider that reads the akka.conf keep-oldest block (present on master BEFORE Critical 1).
So the cluster was NOT running NoDowning before Critical 1; the typed option reinforces the akka.conf
strategy rather than being the sole activator, and Critical 1's 'HOCON inert / NoDowning' premise is
inaccurate. The test guards the failover OUTCOME regardless of activation path. Comments corrected to
reflect this; flagged for review.
A thin gateway over the admin-operations cluster singleton so CI/scripts can trigger a
deployment without the Blazor button. Forwards to the same IAdminOperationsClient.
StartDeploymentAsync; mounted on admin-role nodes. Auth is a fixed-time X-Api-Key check
against Security:DeployApiKey (orthogonal to the cookie-only web auth); AllowAnonymous so the
auth fallback doesn't 401 it, self-disabling (503) until the key is set. Outcome->status:
202/200/409/422. Unit tests for the key check + outcome mapping; HTTP E2E (real auth + real
deploy via the 2-node harness). Documented in docs/security.md.
Standardize the control-plane admin role VALUES on the canonical six
(ZB.MOM.WW.Auth CanonicalRole). OtOpcUa uses four:
ConfigViewer -> Viewer
ConfigEditor -> Designer
FleetAdmin -> Administrator
DriverOperator -> Operator (appsettings-only string role)
This is a rename, not a permission change: enforcement semantics are
preserved (whoever could deploy/administer/operate before still can).
- AdminRole enum members renamed (persisted as string names via
HasConversion<string>); RoleGrants.razor dropdown default updated.
- EF DATA migration CanonicalizeAdminRoles rewrites existing
LdapGroupRoleMapping.Role rows old->new (Up) and back (Down); schema /
model snapshot byte-identical (no pending model changes).
- Enforcement role STRINGS canonicalized:
* Security policies keep their NAMES ("DriverOperator"/"FleetAdmin")
but require canonical roles: RequireRole("Operator","Administrator")
and RequireRole("Administrator").
* Deployments.razor [Authorize(Roles="Administrator,Designer")].
* DevStub now grants "Administrator"; LdapOptions/doc-comment examples
canonicalized.
- Data-plane authorization (NodePermissions/NodeAcl/IPermissionEvaluator/
TriePermissionEvaluator/UserAuthorizationState) UNTOUCHED.
- New CanonicalAdminRolesTests pins canonical claim values end-to-end and
the real registered policies; existing role-string tests updated.
Adds <summary>, <param>, <typeparam>, and <inheritdoc/> tags to public
members surfaced by commentchecker — resolves 5,847 of 5,869 issues
(99.6%) across three /fixdocs passes.
Adds a real-infra mode for the integration test harness alongside the default
in-memory mode. Drops the previously-untested code paths (EF SqlServer
behaviors, real LDAP bind) under env-var control without breaking the
zero-infra default that CI runs.
- docker-compose.yml — minimal SQL 2022 (14331) + OpenLDAP (3894) stack
(ports chosen to coexist with docker-dev/ on 14330/3893)
- HarnessMode record reads OTOPCUA_HARNESS_USE_SQL=1 / USE_LDAP=1 from env
- SQL mode: per-harness unique DB OtOpcUa_Harness_{guid}, EnsureCreated
at startup, EnsureDeleted on dispose (best-effort)
- LDAP mode: drops StubLdapAuthService and configures real LdapAuthService
against the compose'd OpenLDAP via Authentication:Ldap:* config keys
- Microsoft.EntityFrameworkCore.SqlServer added to the test project
- README documents both modes + the macOS no-Docker caveat
Default in-memory mode unchanged — all 9 existing tests still pass.
Extends TwoNodeClusterHarness with three lifecycle primitives:
- StopNodeBAsync() — graceful CoordinatedShutdown (Cluster.Leave)
- RestartNodeBAsync() — rebuild node B on same Akka port + same in-memory DB
- WaitForClusterSizeAsync(n) — converge assertion helper
Adds three failover scenario tests:
- Stopping node B shrinks cluster to 1 Up member
- Restarted node B rejoins on the same Akka port
- Deployment started with B down seals with a single NodeDeploymentState
(validates ConfigPublishCoordinator.DiscoverDriverNodes snapshots
membership at dispatch time)
Closes follow-up F22. Integration test count: 6 → 9 (+3).
DeployHappyPathTests exercises the full deploy pipeline on the 2-node harness:
AdminOperationsActor → ConfigPublishCoordinator → DistributedPubSub →
DriverHostActor on both nodes → ApplyAck → coordinator seals. Verifies both
NodeDeploymentState rows reach Applied and Deployment.Status reaches Sealed.
Exposed + fixed two production bugs along the way:
1. Coordinator was publishing DispatchDeployment on the "deployments" topic but
never subscribed to anything — DriverHostActor ACKs published on the same
topic could not reach it. Added dedicated "deployment-acks" topic with
coordinator subscription in PreStart, and DriverHostActor publishes ACKs
there.
2. NodeId derivation used member.Address.Host only — two cluster members on a
shared loopback host (test harness, dev VMs) collided to one identity. The
coordinator's expected-ack set became {1} and the system sealed after only
half the nodes acked. Switched to host:port everywhere (ClusterRoleInfo +
coordinator) so loopback nodes stay distinct and production identities are
harmlessly more specific.
Tests: 95 v2 tests pass (was 93 + 2 deploy tests), 0 skipped.
Failover scenarios (design §8 cases 3-7: node-kill-mid-apply, split-brain,
restart-during-deploy) deferred — they need controlled node-down primitives
on the harness. Tracked as F22 (failover scenario test cases).
Builds TwoNodeClusterHarness: two in-process Host-equivalent nodes sharing
an in-memory ConfigDb. Forms a 2-member Akka cluster. ClusterFormationTests
proves both nodes see each other as admin+driver role members.
Fixes a real production bug uncovered while wiring the harness — Program.cs
ran two separate ActorSystems (one from AddOtOpcUaCluster.AkkaHostedService
with cluster HOCON, one from Akka.Hosting.AddAkka with bare HOCON). Cluster
singletons landed on the bare ActorSystem and could not actually form a
cluster ("Configuration does not contain `akka.cluster` node").
Consolidation:
- AddOtOpcUaCluster now only binds AkkaClusterOptions + registers IClusterRoleInfo
- New WithOtOpcUaClusterBootstrap pushes embedded HOCON + Remote/Cluster options
into Akka.Hosting's AkkaConfigurationBuilder
- AkkaHostedService.cs deleted — Akka.Hosting now owns the lifecycle
- Program.cs + harness call WithOtOpcUaClusterBootstrap inside AddAkka
Why not WebApplicationFactory<Program>? Program.cs reads OTOPCUA_ROLES from
process env (shared across in-process WAFs); the harness replays Program.cs's
DI graph from a clean WebApplicationBuilder per node with per-node config
overrides. Same production extensions, isolated config + Kestrel + Akka ports.
Tests: 93 v2 tests pass (was 91 + 2 new cluster formation), 0 skipped.
Joseph Doherty