Profiled the remaining lmxopcua#789 misses on windev, a VMware VM, using per-thread CPU and context-switch accounting
plus ETW CPU sampling.
- One blocking hand-off costs ~26 us of latency and ~18 us of CPU there, about 10x macOS.
- An idle evaluation makes four hand-offs in a row: host writer -> worker reader -> evaluation thread -> host reader
-> dispatcher.
- 48% of host CPU at 10k/s is in the kernel: the per-RunAsync NtWriteFile, thread-pool wakes, reads and waits.
Spin-before-block was measured at every hand-off and is not kept. It meets 150 us idle p50 (118 us) only by keeping
about two host cores and four worker cores busy for one stream. At 10k/s it reads +250 points.
Harness: new "paced 10,000/s, batched" row. It uses the same rate and keys, with each 1 ms tick submitted as one
RunBatch call, which is the production seams' shape. It runs after the original four phases, so their rows stay
comparable. On windev it halves host CPU, to +36..+45 points against +71..+84 unbatched. That is still above 25.
Design doc:
- 12.8: the record.
- 14 #15: the owner's decision on the Windows budgets. Recommended: measure on bare-metal Windows, then state the two
budgets per platform, with the batched row as the gate. The alternative is a shared-memory transport.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
A folder event item under enforcement withholds an event that names no live
condition; flag off, the SDK's own type + SourceNode check applies. Positive
controls: the same event reaches the item flag-off and a condition event
reaches it flag-on. Mutation-checked: IsNotifierEventVisibleTo returning true
for a non-condition event fails the test.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
Adds GatewayGalaxyStreamAttachedAckLiveTests (skip-gated on MXGW_ENDPOINT +
GALAXY_MXGW_API_KEY, like the other gateway live tests). Proves what the unit
suite cannot: the request flag reaches a real gateway and the sentinel comes back.
- live session advertises stream-events-attached-ack; GatewayGalaxySubscriber reports it
- raw stream: sentinel is the first message when requested (family Unspecified,
worker_sequence 0, live_after 0) and never sent when not requested (control,
with a data event as the positive control)
- real GalaxyDriver + EventPump + GatewayGalaxySubscriber behind a pass-through
recorder: exactly one sentinel, ack observed before the first SubscribeBulk,
pump StreamAttached completed, sentinel never dispatched nor counted as
received (galaxy.events.stream_attached=1, received == non-sentinel events),
no missing-ack warning within the production 1 s budget, and the static
attribute TestMachine_002.Tagname delivers its initial value
- control: a subscriber that opts out gets no sentinel and the pump never
reports one, while the #865 fallback still delivers
Run 2026-11-12 against windev 10.100.0.48:5120 (mxaccessgw 447070a5): 4/4.
Mutation checks, each caught: subscriber not requesting the ack (sentinel
count 0); driver skipping the ack wait (wire order became stream | subscribe |
ack, 2/2 runs, so the race is real on a live gateway); pump not intercepting
the sentinel (received=2 vs 1 non-sentinel).
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
mxaccessgw#227 (Contracts/Client 0.4.0) adds an opt-in attach acknowledgement on
StreamEvents: a StreamAttached sentinel MxEvent sent first, only after the
stream's subscriber is registered on the session. This closes the #865 attach
race that waiting for "stream requested" only narrowed.
- Bump vendored MxGateway.Client/Contracts 0.3.0 -> 0.4.0 (nuget-packages/,
pulled from the Gitea feed; local-mxgw stays the mapped source).
- IGalaxySubscriber.SupportsStreamAttachedAck (default false, so fakes and older
seams keep the fallback); GatewayGalaxySubscriber reads it from the live
session's capabilities (stream-events-attached-ack) and passes
emitStreamAttached accordingly; TracedGalaxySubscriber forwards it.
- EventPump: records AttachAckExpected before the stream opens (published
before StreamRequested completes); intercepts ev.StreamAttached ahead of the
ReplayGap check - completes StreamAttached and continues: never dispatched,
never counted as received, never a resume cursor (worker_sequence 0). The
signal also completes when the loop ends so no waiter hangs. One pump = one
stream, so a reconnect's new pump has a fresh signal. New counter
galaxy.events.stream_attached.
- AwaitCurrentEventStreamAsync: with the capability, waits on StreamAttached
within the same StreamRequestWait budget (1 s); a missing ack logs a
rate-limited Warning and proceeds instead of hanging. Without the capability
the old StreamRequested wait is the fallback.
Tests (StreamAttachedAckTests, 6): ack-gated subscribe gets its initial value;
capability absent uses the fallback; missing ack times out with a warning; the
sentinel is never dispatched; a dead stream releases ack waiters; reconnect
gives each stream a fresh signal. With the ack wait skipped, 3 of them fail.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
The rule fix itself landed in 1a08b4f7 (detail-0 failure rows route on
category); its test only exercised raw_detected_by = 3. Add per-source
coverage for every MxStatusSource value mxaccessgw's MapSource emits
(-1, 0..5, and an out-of-range value) across the failure categories, plus
the live-captured shape the old rule got wrong (OperationalError detected
by the responding automation object, mxaccessgw docs/AlarmProbeFindings.md)
and a RequestingLmx CommunicationError positive control. All 10 new cases
fail against the old 'detail 0 && raw_detected_by != 0 => comms' rule.
Comment in FromMxStatus now cites MapSource and says why NMX (2/3) is not
a transport signal. Item 1 (Warning/Pending) stays open pending a capture.
Co-Authored-By: Claude <noreply@anthropic.com>
OpcUaApplicationHost left ServerConfiguration.MaxRegistrationInterval at the
SDK default (30000) with a null RegistrationEndpoint, so SDK 2.0.1's
StandardServer armed its registration timer against opc.tcp://localhost:4840,
which on an OtOpcUa node is the server itself (no LDS). It answered
BadServiceUnsupported and the SDK retried forever, visible since #874.
Registration is now opt-in: ApplyDiscoveryRegistration sets
MaxRegistrationInterval = 0 (the value that keeps StandardServer from arming
the timer) unless OpcUa:DiscoveryServerUrl is set, in which case the server
registers against that URL every OpcUa:DiscoveryRegistrationIntervalMs
(default 30000) over SignAndEncrypt/Basic256Sha256, like the SDK's own
default endpoint. The validator requires an absolute opc.tcp:// URL and a
positive interval when the URL is set. Documented in Configuration.md and
OpcUaServer.md.
Co-Authored-By: Claude <noreply@anthropic.com>
- Host reader: the frame read and its handling are caught separately. Only the
read's I/O exceptions mean the pipe ended; anything HandleFrame's inline
completion work throws now kills that worker as ProtocolFault (outstanding
runs -> worker restarted / BadResourceUnavailable) instead of escaping a
dedicated thread (process crash) or being swallowed as a clean end (an
InvalidOperationException used to be, leaving the worker unread until a Crash
2 s later).
- Worker reader: the same catch-all around HandleFrame (exit ProtocolFault).
- Worker: Cancel() runs on the pool, so a script's blocking
ctx.CancellationToken callback cannot stop the worker reading its pipe.
- Host write-stall check: ScriptWorkerFrameWriter.WriteInProgressFor; the
heartbeat monitor kills a worker whose pipe write has been blocked longer than
max(10 s, silence limit) as KilledUnresponsive, which fails the write and
releases the producer (seam submit, deploy load, Cancel). Fixes the wrong
'heartbeat monitor ends it' comment on the pipe buffer quota.
- StopAsync sends Shutdown off the caller so its 1 s budget holds on a full pipe.
- WindowsBlockingPipeStream refuses the async surface (Stream's base versions
serialise a write behind a waiting read).
- Tests: PipeFaultTests (throwing completion callback -> ProtocolFault and a
BadResourceUnavailable sibling, shard serves again; a worker that stops reading
but heartbeats is killed within the bound and the blocked producer returns,
via the Hello.test_stall_reader_on_diag seam); BlockingPipeStreamTests dispose
with a write blocked on a full pipe, and with a read and a write both pending;
Category=ScriptWorker on the protocol, handshake and pipe test classes.
Mutation checks: no stall kill -> stall test red; catch-all without the fault
-> callback test red (Crash).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Gate F (#761) against the VD03 gateway: Galaxy initialize (session open +
platform seed) took 4–5 s against Tier A's 5 s Initialize timeout, so 10 of 40
restarts failed, paid a 10 s retry and left a killed gateway worker.
DriverFactoryRegistry.Register takes optional capabilityDefaults per driver
type (GetCapabilityDefaults); DriverResilienceOptionsParser layers them over the
tier table and under the per-instance ResilienceConfig; the Host wires the
registry into DriverCapabilityInvokerFactory. Galaxy registers Initialize =
20 s (no retry, breaker 2), under the actor's 30 s pessimistic bound.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The host handed the SDK a provider-less telemetry context, so SDK-side
refusals — identity-token decryption (the SDK 2 RsaEncryptedSecret clock
window, #873), certificate validation, secure-channel faults — left no server
log line (found on windev during the #800 Windows validation). The hosted
service now creates one ServerSdkTelemetry context from its ILoggerFactory
(Warning and above, SDK categories, rate-limited 20/min per category) and
shares it between OtOpcUaSdkServer and the ApplicationInstance; released on
dispose. Test: an untrusted client's rejection reaches the captured log.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The fixed name lives on the SQL Server shared by every checkout and CI, so a
concurrent run's create or finally-drop could collide (failed once in the
local pre-push suite, 3/3 green on rerun; cause plausible, not proven).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Since #855 condition events are queued to folder event items under the node
manager Lock. At a full DiscardOldest queue the SDK's EventQueueHandler
dequeues the head of EventMonitoredItemQueue, a List.RemoveAt(0) shift of the
whole queue (~1 us/item at QueueSize 1000, ~10 us at 10000).
OtOpcUaSdkServer.CreateMonitoredItemQueueFactory now wraps the SDK factory in
RingEventQueueFactory: event queues are RingEventMonitoredItemQueue (O(1)
discard), data-change queues, durable requests and disposal go to the SDK
factory unchanged. A factory set through StandardServer.MonitoredItemQueueFactory
is used as-is.
The SDK queue is storage only (decompiled 2.0.1): DiscardOldest, overflow-event
positions and discarded/overflow counters live in EventQueueHandler, which is
unchanged. The ring reproduces QueueSize 0 initial + BadInternalError enqueue,
drop-oldest at Count == QueueSize, FIFO dequeue, SetQueueSize keep-newest /
keep-oldest, the newest-1000 duplicate window via IsEventRaisedFor, no-op
Dispose. RingEventMonitoredItemQueueParityTests drive both queues, and the SDK
EventQueueHandler over both factories, with the same random sequences.
FolderEventDeliveryPerfTests gains Ring_vs_sdk_event_queue (A/B in one process
via the internal OtOpcUaSdkServer.RingEventQueues switch).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Every raw detected_by value names an MXAccess layer (requesting/responding
LMX, NMX, automation object); the gateway's own failures arrive as
ProtocolStatus. The 'non-zero source = transport failure' rule made a failure
the responding object detected a comms error. Item 1 of #872 (Warning/Pending
categories) stays open pending a captured row.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The gateway fills Statuses on every data change (live OnDataChange, value-cache
replay and session snapshot all carry Quality too), and EventPump used the status
row whenever one was present, so Quality was never read: an off-scan, Bad or
Uncertain attribute with an OK status row was published Good. A failure status
row still wins; otherwise the quality byte grades the value (legacy parity:
quality >= 192 Good).
Live, read-only on VD03 (DEV): healthy ScanState and _TestArea attributes stay
Good (0x00000000) via quality 192; the stopped JoePlatform/JoeAppEngine ScanState
now publish BadCommunicationError (0x80050000) instead of generic Bad, with no
"unrecognised quality byte" warning.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- MeshPskProviderRefreshTests: GatedResolver.Entered was set by the priming
refresh and never reset, so Wait() returned before the held refresh reached
the store and the mid-refresh call-count snapshot raced its increment
(expected 1, was 2). Hold() now resets it.
- FailedCredentialLockoutTests all-bad burst: the 15 waiters could queue
before a slot holder ran the increment after BeginAsync (authenticated 0
vs 5). The wait now covers both conditions; the security assertions are
unchanged.
Both looped 20x locally, 0 failures.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The gateway copies MXSTATUS_PROXY.detail verbatim, so a failure row's detail is
always an MxStatusDetail code. FromMxStatus routed any detail <= 0xFF through the
OPC DA quality table: MX_E_PlatformCommunicationError (2) on a stopped platform's
ScanState logged "Unrecognised MXAccess quality byte 0x02" and published generic
Bad (VD03 live capture), MX_E_InvalidObjectId (8) became BadNotConnected,
MX_E_Overflow (20) BadCommunicationError, and a detail in the Uncertain band
published Uncertain for a failure.
Every non-zero detail now maps per the gateway.md vocabulary (timeout,
communication, configuration, type, range, access, initialising), then by
category; the category-fallback warning logs once per (detail, category).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- The filter-group registry holds groups weakly (the filter->group table is
the only strong reference), sweeps dead entries amortised and counts only
live shapes against the 4096 cap; filters past the cap are counted on the
new meter otopcua.opcua.condition_events.filter_groups_unshared.
- A group validates its where clause too (restored, never-validated filters
evaluate like validated ones).
- Test: value-equal filters cost one projection per locale set per report.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- Filter-group buckets include the where clause's (value-based) hash, so
per-source filters sharing a select list are not all scanned under Lock.
- Locale keys are cached per session array.
- Tests: a notifier reached twice in one walk (Line above two wired
equipment) gets each event once on both paths (fails without the
per-report set); queued events re-projected by a modify are identical on
both paths; locale keys unambiguous and where clauses hash by value.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- A filter group owns a deep clone of its first filter, validated once and
evaluated under the group's lock: no item-owned filter is mutated or raced
(Validate mutates a select clause).
- The projection is computed in the SDK's order: where clause first, fields
only for an item that will queue.
- Unambiguous locale key; counter for filters given a private group past the
registry cap; DeliveryCache getter returns its own allocation.
- UseProjectedEventDelivery=false also asks the role and grant checks per
item, so the equivalence test compares against the pre-optimisation path.
- Equivalence test: same select list under three where clauses, a filter
modified mid-run, a session with other locales, an ungranted anonymous
Server-object item, per-rig counter; plus a small overflowing queue with
DiscardOldest true and false.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Condition events delivered to many event items (folder and Server-object)
no longer pay the SDK's per-item QueueEvent(IFilterTarget) work under the
node-manager Lock: the role-permission verdict is shared per (session,
identity), the #826 grant check per identity per hop, and the where clause +
select-clause projection per (value-equal EventFilter, locales); each item
gets its own EventFieldList over the shared read-only fields. The SDK's
QueueEvent body is reproduced under the item's own lock (three private SDK
members, guard-tested); anything outside that shape falls back to the SDK.
The per-report duplicate check replaces the SDK's O(queue) scan.
ProjectedEventDeliveryWireTests compares both paths over the wire.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
M1 IsPoison classified every SQL Server transient as poison:
Microsoft.Data.SqlClient never sets DbException.IsTransient, so a deadlock
(1205), timeout (-2 / Win32 258), dropped connection or AG failover was
counted from memory and the node's row left Applying. Poison is now an
ALLOWLIST: a DbUpdateException (not concurrency) carrying a SqlException
whose number is 8152, 2628, 547, 515, 2627 or 2601. Everything else
retries, bounded by the deadline deferral cap. IsPoison is public for tests.
Acks counted only from memory (poison branch and the deferral cap) are kept
in _memoryOnlyAcks and re-sent by PreRestart alongside _unpersisted, so a
restart before the terminal write does not lose them.
M2 Revisions adopted without an outcome-reporting apply are unconfirmed, so a
same-revision dispatch re-applies instead of short-circuiting a clean
Applied over an empty address space: TryRecoverFromStale (restores only the
revision marker) and RestoreApplied when the artifact read fails (null
blob, LoadArtifactBlob swallows the error) or the restore throws.
L1 The re-drive comment documents the CreatedAtUtc recompute after a second
restart / failover.
L2 _currentRevisionDegraded renamed _currentRevisionUnconfirmed; the stale
TryBootFromCache remark corrected.
L3 docs/Redundancy.md: a Direct node running from cache that still cannot
reach the ConfigDb acks a same-revision re-dispatch Failed
(PartiallyFailed, not Sealed); Stale recovery and failed restores are
unconfirmed too.
Test harness: SqlExceptionFactory builds real SqlExceptions (reflection on
SqlError / SqlErrorCollection / SqlException.CreateException);
ConstraintEnforcingDbFactory now throws DbUpdateException over SQL 2628
(over-length), 547 (poison) and 1205 (transient), plus FailNextContexts.
Tests (all fail on the previous code): poison allowlist theory (1205, 40613,
10054, 4060 not poison) + timeout/concurrency/providerless not poison; a
deadlocked ack write is retried and recorded; a memory-counted ack survives a
coordinator restart; Stale-recovered and failed-restore revisions re-apply on
re-dispatch.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
A pending local ack can outlive its own delivery (sources that send no
Acknowledge back; a Clear now only carries it through), so a later failed
re-ack — or a stale Galaxy confirmation timing out after the alarm re-occurred
and was acked again — could revert an ack it does not own. MarkAcknowledged
now issues a sequence (0 when the ack changed nothing), carried on
RouteAlarmAck and echoed on AlarmAckFailed; TryRevertAcknowledge reverts only
on a matching sequence.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
FolderEventDeliveryPerfTests (Category=Perf, skipped unless OTOPCUA_EVENT_PERF=1)
reports per-alarm-report node-manager Lock hold and end-to-end delivery latency
for one equipment folder with N = 1/10/50/200 event items, ACL on/off, folder
subscribers publishing or stalled at a full queue; the latency of an i=2253
subscriber while folder subscribers stall; and a QueueSize sensitivity sweep.
Compiles against the pre-#855 tree for a before/after comparison.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Adds an escalated-stream-fault test seam and a test: the live pump's fault
escalates (positive controls before and after), a superseded pump's late
fault after a reopen does not; the absence is asserted only once both
pumps' callbacks are known to have run (HasEnded). Verified to fail with
the filter disabled. The filter's comment now says the identity check does
the work and warns about a future session bump while Healthy.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
NativeAlarmProjector dropped the pending local ack on EVERY driver transition,
so an operator ack followed by the alarm returning to normal (a Clear carrying
the optimistic ack) inside the confirmation window made a later "not
confirmed" report a no-op — the condition stayed inactive+acked while Galaxy
showed UNACK_RTN. Only a transition that states the ack (Acknowledge,
ActiveAcknowledged, or an un-acking Raise/Retrigger/InactiveUnacknowledged)
now settles it; a Clear / same-message Raise / Unspecified update leaves it
revertible, and the revert projects inactive+unacked.
Tests: projector + host end-to-end (ack, Clear, confirmation fails -> inactive
+ unacked); tightened the confirmed-ack positive controls; pinned that the
confirmation receives the same request instance as AcknowledgeAsync.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
1 PersistNodeAck wrote a Failed ack's FailureReason unbounded; the column is
nvarchar(2048), so on SQL Server the write failed forever, the ack retried
forever, the deadline deferred forever and every later deploy was refused.
- One shared bound, Configuration.FailureReasons.Truncate (MaxLength 2048),
used by the coordinator and DriverHostActor.UpsertNodeDeploymentState
(exception messages and degraded reasons).
- A non-transient DbUpdateException (not concurrency, no transient
DbException / timeout / socket / IO underneath) is POISON: the ack is
counted from memory with an Error, never retried.
- Deadline deferral on unpersisted acks is capped at 10 x the retry ceiling;
past it the outcome is decided from the acks held in memory.
2 A node restarting in the outcome window boots the cached revision (cache is
written before the outcome) and never acks. Cache-booted revisions
(BootstrapFromCache / TryBootFromCache) are now marked unconfirmed, so a
re-dispatch re-applies and acks the real outcome instead of a clean
short-circuit. docs/Redundancy.md corrected (replay only on the ConfigDb boot
path).
3 A lost dispatch re-driven near its deadline now gets a FULL deadline (a node
needs its whole outcome timeout + margin = ApplyDeadline, so any shorter
remainder timed out nodes that applied). No dispatched-at column: a second
restart recomputes from CreatedAtUtc and closes it instead (documented).
4 docs/Redundancy.md upgrade note: do not deploy while the central pair is mid
rolling-upgrade across this change.
Tests (all fail on the previous code): oversized Failed reason truncated (on a
SaveChanges interceptor that enforces the column like SQL Server); rejected ack
write counted from memory; bounded deadline deferral; full deadline on a
near-deadline re-drive; cache-booted revision re-applies on re-dispatch;
oversized degraded reason truncated on the node's own row.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
ReconnectSupervisor swallowed a transport failure reported while Replaying,
so a restarted event stream refused/ended a round trip after the replay
began still let the replay go Healthy over a dead stream (and the next
advise's re-report restarted at attempt 1 with no backoff). A failure
reported during Replaying is now remembered and turns that attempt's
success into a retry with backoff. ReplayAsync's synchronous dead-pump
throw yields to a teardown's cancellation.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- Start the ack confirmation with Task.Run + ConfigureAwait(false) so it never
depends on the ReceiveAsync handler's ActorTaskScheduler (the PostStop
abandon-and-revert report must run after the actor stopped).
- DriverInstanceActorAckConfirmationTests: a confirmation that faults after the
handler returned is reported; one pending at actor stop is reported; a
confirmed ack reports nothing (positive control).
- Deterministic tracker matching test (Raise/Clear/Retrigger/other alarm do not
confirm; case-insensitive match does).
- AdminUI min/max on the confirm-timeout field; Galaxy.md documents the
reconnect, UNACK_RTN-as-Clear and RPC-budget limits.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- ReplayAsync fails when its restarted pump's stream has already ended, so
the supervisor backs off and stays Degraded instead of going Healthy over
a dead stream (whose next advise re-report restarted at attempt 1 with no
backoff - a tight reconnect loop).
- ReconnectSupervisor.AttemptInProgress (reopen through replay); advises
wait for a stale pump's replacement only while an attempt is running, not
for a whole backoff (the state still reads Reopening/Replaying there).
- ReconnectSupervisor: _disposed is set under the state lock and re-checked
inside ReportTransportFailure, so a late report cannot start a recovery
loop that StopAsync already missed.
- EventPump.HasEnded is a flag set before StreamRequested is released, and
a stream that ends/faults on its first MoveNext is marked ended first.
- SubscriptionRegistry fast path: a miss now falls through to the locked
path (lookup + session + window check atomically) instead of a lock-free
re-lookup.
- Tests: replay over an immediately-ending stream fails; AttemptInProgress
across attempt/backoff; ReportTransportFailure after stop; the
reopen-window test uses a generous wait bound.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- RefreshDirectItemsWithCut keys its "decided off Lock" set per (item,
condition), so a condition one item's walk saw is still decided for another.
- The deferral warning no longer blames slow folder subscribers (no folder of
ours defers since #855).
- FolderEventFloodWireTests' removal test now asserts the folder item got
nothing and the Server object exactly the pre-stall event (it passed with
the post-stall liveness re-check deleted).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
mxaccessgw answers AcknowledgeAlarm OK both when the AckMsg write was
applied and when it was only sent (no OnWriteComplete; the reply then
carries a diagnostic). EvaluateReply treated both as success, so the H6
live gate saw a condition go Acknowledged while Galaxy never applied it.
- EvaluateReply returns Applied (no diagnostic) or Unconfirmed (diagnostic)
and logs the gateway diagnostic on every branch that carries one.
- GalaxyAckConfirmationTracker watches the alarm feed from before the RPC
and, for an unconfirmed reply, waits up to AlarmAck.ConfirmTimeoutMs
(default 10 s, 0 = old behaviour) for an Acknowledge/ActiveAcknowledged
transition; on timeout it refuses (BadTimeout + diagnostic).
- New optional IAlarmAcknowledgeConfirmation: DriverInstanceActor awaits it
after AcknowledgeAsync, off its mailbox and outside the resilience
pipeline; a failure is the existing #760 AlarmAckFailed revert.
- AdminUI Galaxy form carries the new knob; docs updated.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- StreamRequestWait 1 s (under Tier-A Read 2 s / Subscribe 5 s budgets) and
a stale pump is only waited on while the supervisor is reopening or
replaying (or a restart is underway); in a reconnect backoff the advise
proceeds at once. The no-stream log is a rate-limited Warning while
Healthy, Debug otherwise.
- A current-session pump whose loop has ended is not a stream: if the
supervisor already reports Healthy (its fault was swallowed while it was
recovering) the failure is re-reported to start a new cycle.
- Subscribe/Read no longer start the pump directly (only
AwaitCurrentEventStreamAsync does, honouring the replay restart flag);
the teardown refusal is an explicit ThrowIfRuntimeStopping.
- SubscriptionRegistry.ResolveOrPark fast path: the handle lookup and the
open-window check were not atomic, so a subscribe that registered the
handle and closed its window in between got its initial value dropped as
stale (found by a stress run: the first event of a bulk advise lost). It
re-looks-up before declaring stale. Deterministic test via a seam.
- Tests: first Read on a fresh driver, replay of a data subscription.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
M1 DriverHostActor wrote its own NodeDeploymentState row Applied/null as soon
as the revision advanced, up to ~105 s before the address-space outcome.
The coordinator's recovery counts any non-Applying row as an ack, so a
coordinator restart in that window sealed CLEAN while the node was about to
report degraded (#601). The row now stays Applying until ReleasePendingAck
(or the no-publish-actor path) writes the final Applied, clean or with the
degraded reason. A node restart in the window replays the orphan Applying
row; /fleet shows Applying until the outcome (docs/Redundancy.md).
M2 Acks whose persist failed are tracked in _unpersisted (newest per node).
HandleDeadline persists that backlog first and defers its decision while
any remain, so a node that acked is never timed out as silent; PreRestart
re-sends them to Self so a restart (timers lost) does not drop them.
L3 A successful persist cancels the node's retry timer and drops its older
pending ack; a retry only fires for the node's current pending ack.
L4 A dispatch queued during Loading is sent with a fresh deadline however long
the outage lasted (never closed "Dispatch was lost"); the inline re-drive of
a genuinely lost dispatch arms the REMAINING deadline.
L5 A recovered row with a blank RevisionHash (re-dispatch) or blank NodeId is
closed PartiallyFailed with a reason instead of failing every load.
L6 Readiness carries PendingAckPersists + LastPersistError; the config-publish
health check is Degraded while Ready with a backlog.
L7 Ack-persist retries back off (shared Backoff); stack trace on the first
failure only.
L8 Terminal writes (Sealed / PartiallyFailed / TimedOut / failure reason) only
apply to a row still Dispatching/AwaitingApplyAcks.
Tests (all fail on the previous code): ControlPlane DbOutage — deadline does
not time out an unpersisted ack; unpersisted ack survives a restart; stale
retry does not overwrite a newer ack; queued dispatch through an outage longer
than the deadline is sent; malformed row closed and coordinator carries on;
readiness reports the persist backlog. Runtime DegradedAck — own row stays
Applying until the outcome. Host — Ready + backlog is Degraded.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- Both plain ConditionRefresh paths (Server object and folder) skip a
condition a deploy removed or replaced between the off-Lock walk and the
queueing (the SDK base did not either; only the cut path checked).
- A failure delivering at one notifier hop is logged and the walk continues
to the remaining hops; a failure deciding the Server-object items withholds
only them.
- Test: a condition removed during a plain refresh (folder + Server object).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Round 1 replaced a stale-session pump from any advise path and disposed the
old one in the background; that could start a second stream while the old
one still held the gateway's single-subscriber slot, and a kept
current-session pump could be dead with no fault re-reported.
- EnsureEventPumpStarted reuses again; only ReplayAsync restarts the pump.
RestartEventPumpForReplay keeps a current-session pump only while its loop
is running (EventPump.HasEnded), and flags the restart so no advise path
starts a competing pump in the dispose gap.
- AwaitCurrentEventStreamAsync (bounded by StreamRequestWait) starts a pump
only when none exists, WAITS for the replay while the pump is stale or
restarting, waits for StreamRequested, and returns the session it verified.
SubscribeAsync (per #861 attempt), read-via-subscribe and the probe hook
issue and fence the advise under that session, so a reconnect after the
check is caught as a straddle and retried.
- Read: a deadline (not caller) cancel during the wait falls through to
per-ref results.
- Tests: reopen-window probe test now asserts the advise waits for the
replay's single new stream; first Subscribe/Read wait for the stream; the
fencing tests that advise in the window use a short wait bound; the
production-path seed test uses an empty galaxy (a real probe advise now
starts a stream against the fake endpoint).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- QueueFrozenReplay's verdict refuses a condition a deploy removed or replaced
since the cut (Server-object and folder), so the cut path cannot replay a
ghost or judge a replaced scripted alarm on its old folder.
- DeliverToNotifierItems resolves the #826 decision's item-independent part
(live condition, notifier check, scopes) once per hop; only the grant and
SDK role checks stay per item.
- A deferred delivery's drain starts on the pool, never inline under Lock.
- Test: a condition removed between the cut and its replay (folder + Server
object).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Review findings on 02503a4a (ConfigPublishCoordinator):
1 (HIGH) A recovered deployment whose every expected ack was already in
NodeDeploymentState (a seal whose SaveChanges failed, a fused node's own row)
was never completed: inside the deadline it waited for acks that had landed,
past it HandleDeadline's "race" guard returned without sealing — stuck in
AwaitingApplyAcks, refusing every later deploy. ApplyRecovery now completes
it (Seal / PartiallyFailed), and HandleDeadline completes a full ack set
instead of returning.
2 (MED-HIGH) The past-deadline recovery of a populated expected set ran
HandleDeadline inline before the stash replay, discarding on-time acks queued
during the outage. It is now self-sent, behind the replay. The zero-rows
past-deadline branch stays inline on purpose: no ack can count there, and a
self-sent deadline would time out a queued dispatch of that same row right
after publishing it.
3 (MED) A failed PersistNodeAck threw, restarted the actor and dropped the ack
(driver-only nodes never write their own row and do not resend). Acks are now
persisted before being counted, and a failed persist is re-queued on a timer.
The class remarks document that the Loading stash is node-local (lost on a
hand-over; queued dispatches are recovered from their rows).
Minor: MarkReady/Become(Ready) only after the recovery applied; a throw while
applying is a failed load (backoff) rather than a tight restart loop; a queued
dispatch matching the recovered Dispatching row logs Info, not the "lost
dispatch" Warning.
Tests (ConfigPublishCoordinatorDbOutageTests): fully-acked recovery seals /
partially fails inside and past the deadline (4 cases); queued ack applied
before a past-deadline recovery; failed ack persist retried. All 6 fail on
02503a4a.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- EnsureEventPumpStarted replaces a pump created for an older registry
session (the reopen -> replay window still held the dead session's pump,
so an advise there lost its initial value and the replay skipped it);
RestartEventPumpForReplay keeps a pump already on the current session.
- SubscribeAsync, read-via-subscribe and ReplayAsync also await the
bounded StreamRequested wait before advising, not just the probes.
- TryEnsureEventPumpStarted tolerates a teardown flipping _runtimeStopping
mid-check; the probe hook honours an already-cancelled seed token.
- StreamRequested is documented as narrowing, not closing, the attach race.
- Tests: reopen-window probe test, EventPump.StreamRequested tests, and the
FixRound1 fresh-pump assertion now waits on a positive control.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Joseph Doherty