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
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
- 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
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
- 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
§12.1 gains the 2026-11-11 windev (VMware VM) perf table: idle p50 581 us
(MISS), kill latency after T+G 8-16 ms (PASS), kill both shards 2.12-2.14 s
(MISS by ~6 %), worker RSS 160 MB compiled / 36 MB warm-cache (PASS under the
revised 200 MB budget), host CPU +131 % vs InProcess (MISS, untuned).
It also records the first Windows run of Category=ScriptWorker: 107/0/2 from
C:\build, all 13 WorkerSoftClockTests cases (5 freezing a real worker with
NtSuspendProcess) green with no code change. From a user-profile directory
the restricted-identity isolation test fails because the restricted token
cannot read the worker dll there (profile ACL), which is environment, not a
defect.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
- QueueFrozenReplay re-asks each item's verdict right before each entry's gate
hold (Server-object and folder cut paths): a grant revoked between the
pre-Lock verdicts and the replay no longer lets the replay through. A denied
entry is also removed from the cut's re-sent set.
- ReportEventWithoutBlocking logs and returns on a delivery that faulted
inline instead of booking it as a deferral episode (which ran the drain and
ClearStaleRefreshCuts under the event-report gate while Lock was held).
- OnConditionReportEvent isolates failures per Server-object item.
- Tests: revocation between cut and replay (folder + Server object), throwing
hop not booked as a deferral; production-overload sampling inside Lock.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
When the data starts inside an interval (no start bound) and its first value is
Uncertain, the frozen 1.5.378.182 AverageAggregateCalculator integrated the
uncovered span as a ramp from 0 and divided TimeAverage by the whole interval:
GetRegionsInValueSet (ignoreBadData, sloped) overwrites the BadNoData start
region's status with UncertainDataSubNormal in its one branch without a not-Bad
guard. Live AggregateFilter therefore disagreed with HistoryRead and Part 13
§5.4.3.6/§5.4.3.8 ("No Start Bound: Calculate Partial Interval"); QP: Total
188.96 vs 163.96, TimeAverage 9.448 vs 10.931. SDK 2.0.1 stock does the same.
Owner-approved 2026-11-11.
- New app-owned CoveredTimeAverageAggregateCalculator subclasses the baseline
and recomputes only TimeAverage/Total, restoring the Bad status of a region
that starts at a Bad point (only the start bound can). Only such an interval's
value changes; status, timestamps, type and bits are the baseline's. Stepped
and TreatUncertainAsBad=true never reached the defect. CompatibilityAggregate-
Factory routes exactly those two IDs; copied SDK files untouched.
- Golden test: 0 fixture vectors change (every data set has a usable value at
-10 s), asserted. New sweep over late-start variants of every data set (cut
+5 s / +25 s, first non-Bad sample Uncertain, all directions/stepping/configs/
incremental): 150 / 142 of 384 change, each pinned to a baseline-derived
oracle (untouched baseline's value with that sample Good, its status/bits on
the variant itself) - never our calculator. Plus a hand-computed case.
- Live-vs-HistoryRead parity (TimeWeightedAggregator) forward + reverse, no
masks, with discrimination against the pre-#821 baseline; QualityMix now
compares QP's TimeAverage/Total with the live oracle on both SQL backends;
booted-runtime facts test expects 163.96 / 10.93.
- README recorded decision, ServerAggregateConfiguration remarks, oracle
remarks, docs/Historian.md and .claude/rules/historian.md updated.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
PreStart ran the in-flight recovery query synchronously. With the ConfigDb
unreachable when the singleton started, the throw became an
ActorInitializationException, the default strategy stopped the actor, the
ClusterSingletonManager never re-created it, and every later deploy hung.
- PreStart now only subscribes and kicks off an async recovery load piped back
to the actor. A failure retries with exponential backoff (1 s doubling to
30 s, forever) and never stops the actor.
- Until the load succeeds the coordinator is in a Loading behaviour that
stashes every message (DispatchDeployment logged as queued, ApplyAck too);
they are replayed in order after the recovered state is applied. Recovery
decisions (re-dispatch / time out) now run inline before the replay instead
of via self-tells, and a duplicate dispatch of the in-flight deployment with
a seeded expected set is ignored so queued acks are never reset.
- New ConfigPublishCoordinatorReadiness (admin-node DI singleton) +
"config-publish" ready-tier health check: Degraded while loading (with the
attempt count and last error), Healthy when ready or not hosted here.
Tests: ConfigPublishCoordinatorDbOutageTests (DB down at start -> actor alive,
Loading reported, deploy queued, DB returns -> load succeeds, queued deploy
dispatched and sealed; queued ack applied to the recovered in-flight set;
NotHosted after stop) and ConfigPublishCoordinatorHealthCheckTests.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
HeidenhainSession.Leaf's `case Lsv2Tool t` switch expression mixed long, int and
double? arms, so C# unified the natural type to double? and Tool/Number (Int64),
Tool/Index and Tool/AxisCode (Int32) were published as Double. Every arm in that
switch (and in the alarm-list, axes and feature-acquire switches in the same file)
is now boxed to object explicitly so no numeric arm can widen another.
New HeidenhainLeafRuntimeTypeTests asserts value.GetType() for every catalog value
leaf against its declared DataType (scalar + array), plus PLC/DATA binding coercion
for Int16/Int32/Int64/Boolean. Verified red on the old code (exactly the three Tool
leaves) and green with the fix.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Condition events reaching one of our notifiers no longer go through the SDK
MonitoredNode2 channel: the walk serves the attached items itself, per item and
per event (audit guard, uncached INodeManager3 verdict, item gate/cut,
QueueEvent), as the Server-object path does. No notifier of ours defers any
more; the #838 deferral queue and the #854 cut stay as the safety net and
folder items share the cut (RefreshDirectItems, flag on and off). An event
whose walk suspends is marked deferred before it suspends.
Tests produce deferral with a notifier-stall probe instead of holding a folder
item's SDK lock (which would now block the reporter). New
FolderConditionRefreshWireTests; design note in docs/plans.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
TwinCATDriverProbe ran a first-phase TCP check against the AmsNetId's
first four octets on the AMS port (801/851). An AmsNetId is not an IP and
the AMS port is not a TCP port, so Test Connect timed out against every
real runtime (Z35474: dialled 5.118.56.93, an unrelated public host).
The probe now opens no socket of its own: it takes the driver's path,
AdsClient.Connect + ReadStateAsync through the local AMS router, bounded
by the probe budget (Task.Run + WaitAsync), and classifies failures as
no local AMS router / no route or target unreachable / timeout / other.
The degrade-to-green 'TCP reachability only' outcome is gone - with no
TCP check nothing was contacted, so a missing router is reported red.
Tests: the regression test listens on the AmsNetId-derived endpoint and
asserts no connection arrives (fails on the old probe); classification
and budget tests drive the handshake through an internal seam.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The owner revised the per-shard worker idle RSS budget from 80 MB to 200 MB
+ scripts on 2026-11-11: compiles run in the worker, so Roslyn's code and
metadata stay resident (measured 159-187 MB) and no GC setting moves it.
The perf harness now reports both RSS rows against the 200 MB constant with
a PASS/MISS mark; A-3 records what the measurement showed.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The frozen 1.5.378.182 StatusAggregateCalculator counts Uncertain regions as
Good under TreatUncertainAsBad=false, so live AggregateFilter PercentGood /
DurationGood disagreed with HistoryRead (Part 13 §5.4.3.31/§5.4.3.33, Mantis
0006166) whenever Uncertain data was present. Owner-approved 2026-11-11.
- New app-owned GoodOnlyStatusAggregateCalculator subclasses the baseline and
replaces only the PercentGood/DurationGood computation (Good regions only);
CompatibilityAggregateFactory routes exactly those two IDs to it. The copied
SDK files are untouched; the other 35 aggregates are byte-identical.
- Golden test: the 64 changed vectors (38 PercentGood + 26 DurationGood) are
exempted exactly — those whose baseline output differs from the same case's
baseline TreatUncertainAsBad=true twin — and pinned to that twin (baseline
evidence, not regenerated from the implementation); plus a full Good-only
pin over all 352 PercentGood/DurationGood vectors and a hand-computed case.
- Live-vs-history parity over Good/Uncertain/Bad mixes (no masks), with a
discrimination check against the pre-#809 baseline; the booted-runtime facts
test now expects 8000 ms / 40 %.
- SdkAggregateOracle: Run/RunLive = the registered factory; RunSdkAsIs = the
untouched baseline; its Part13GoodCalculator duplicate is removed.
- README (recorded decision), ServerAggregateConfiguration remarks,
docs/Historian.md and .claude/rules/historian.md updated.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Owner decision 2026-11-11: /fleet/diagnostics "Faulted devices" counts DISTINCT
devices across the nodes of the selected node's cluster, Faulted state only.
- DriverHostActor: FaultedDevices = HostState.Faulted only (Stopped no longer
counted); new StoppedDevices + FaultedHostNames. Counting extracted to the
testable DiagnosticsFromHealth.
- DriverInstanceDiagnostics: StoppedDevices (default 0) and FaultedHostNames
(default null) are additive with defaults, so a pre-#786 sender's snapshot
still deserializes over Akka's default JSON serializer (pinned by a test).
- FleetDiagnosticsAggregation (pure): identity key = (driver-instance string
id, HostName), case-insensitive; a single-host driver / nameless pre-#786
count is combined by MAX across nodes, never summed. Cluster peers come from
enabled ClusterNode rows of the selected node's ClusterId.
- FleetDiagnostics.razor: tile fetches the cluster peers and shows the
distinct count + "n of m nodes answered"; per-driver table gains a Stopped
devices column.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Owner decision 2026-11-11: certificateValidation.validateServerHostName stays
opt-in (unset = off). The AdminUI toggle moves from the "Subscriptions, limits &
discovery" panel to the Security panel beside mode/policy, with help text on the
NAT/IP caveat, IPv6 zone ids and the loose first-label match. Binding unchanged
(tri-state, "Default (off)" leaves the key absent). OpcUaClient.md states the
decision; drivers rule updated; source-scan test pins placement + caveats.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Owner decision 2026-11-11 (Java parity): a genuinely empty observation on a
String/Reference or unauthored item is Good "" instead of BadNoCommunication.
Non-string items (numeric, bool, DateTime, arrays) keep BadNoCommunication -
the Agent supplied no value; BadTypeMismatch would wrongly blame the authored
type. The UNAVAILABLE sentinel is unchanged, and a structured observation with
no text stays BadNotSupported (structured check now precedes the empty check).
Parity matrix, quality-mapping section and drivers rule updated.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
With Reconnect.ReplayOnSessionLost=false a recreated gateway session carries none of
the old advises, so every pre-reconnect subscription went silently deaf: nodes kept
their last value as Good and health returned to Healthy (only a Warning log).
ReplayAsync now records the stranded subscriptions and publishes BadNoCommunication
(StatusCodeMap.BadNoCommunication, 0x80310000; value null — no driver value cache)
for every tag of each, straight to OnDataChange so LastSuccessfulRead is not stamped.
GetHealth reports Degraded naming the setting while any stranded subscription is
still registered; it clears on UnsubscribeAsync (the host's re-subscribe path) or
runtime teardown, and pruned ids never pin it. A subscribe after the reconnect is an
ordinary new-session subscription. With nothing else flowing, the comms-loss
watchdog re-initializes the driver after 90 s (a delayed replay).
Tests: Runtime/StrandedSubscriptionTests (6). Docs: drivers/Galaxy.md, which also
no longer claims the replay uses the gateway's ReplaySubscriptions RPC.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
BLOCKER: the reconnect replay chose work from the registry entries' sessions,
but ONE registry entry holds every probe handle and any probe publish on the
new session (a seed in the reopen→replay gap, or the straddle retry) re-stamps
it as current — the replay then skipped the probe watcher and the old
platforms stayed bound to handle 0 (host status frozen); a publish after the
snapshot made it re-subscribe everything (double advise, leaked handle).
ReplayAsync now excludes the probe entry from its snapshot and calls the new
PerPlatformProbeWatcher.ResubscribeStaleAsync on every reconnect, which
re-subscribes exactly the platforms whose ProbeHandle.Session predates the
current session. The replay-gap path keeps ResubscribeAsync (all platforms).
Also: ClearRegistry (teardown) starts a new registry session, so a late
PurgeParked from a teardown-orphaned call cannot reach the next runtime's
parked values; the read path's unused pump local is gone (call kept).
Tests go through the driver's reopen/replay: the probe old-session-number test
is rewritten that way, plus the seed-in-the-gap scenario (only P and Q are
re-subscribed, R keeps its new-session handle, all three resolve). Both checked
red with the probe replay removed / re-subscribing all platforms.
Refs #859, #861.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
B-1: the subscribe-once read's cleanup unsubscribes what the registry holds at
removal (the replay may have rebound it to new-session handles), not its
original bindings; nothing when the entry still holds dead old-session
handles. Found while testing it: reads were attached to the pump INSTANCE, so
a read across a reconnect's pump restart never saw the replayed value — reads
now listen on a driver-level ReadDataChange raised for whichever pump is live.
I-A: PerPlatformProbeWatcher records the registry session with each probe
handle. Publishing goes through the new fenced TryUpsert and binds only
current-session handles; removal/stop unsubscribe and purge only those; a
seed reply that straddled a reconnect records nothing and is re-issued (max 3).
I-B: PurgeParked takes the session captured BEFORE the unsubscribe RPC and is
a no-op under the park gate if the session changed.
I-C: ReopenAsync starts a new registry session before AND after the gateway
session swap; ReplayAsync does so only when no reopen ran (injected subscriber
/ tests) and replays only subscriptions still holding older-session handles,
so a subscribe issued after the swap is live and not double-advised.
P-1: ReplayAsync restarts the pump before any early return (no entries used to
leave the dead pump in place). P-2: ReplayOnSessionLost=false logs a Warning;
misleading comment about gateway-side replay fixed.
M-a: TryRebind returns BindOutcome; UnknownId releases the handles just
advised for a subscription removed mid-replay. M-b: the replay/re-snapshot
pass reads the session once and stops if it changes. M-c: parked entries drop
their MxEvent when they leave the parked set. M-d (retry a straddled read)
not done.
Tests (TestWait) for each, each checked red against a mutation of its fix.
Refs #859, #861.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
I-2: unclaimed parked events are trimmed by window sequence. Each window gets a
monotonically increasing sequence at Begin, each parked event is stamped with
the newest one, and on every End events stamped before the oldest open
window's sequence are dropped (amortised O(1) via a park-order queue). Under
concurrent reads the open count rarely hit 0, so trailing events used to pile
up to the bound and push genuine early values into overflow.
I-1 / #861: SubscriptionRegistry.StartNewSession() — called FIRST in
ReplayAsync on every reconnect — clears held events, invalidates the reverse
map (subscriptions kept) and bumps a session number. New-session events whose
number an old binding still held now park until the replay's per-subscription
Rebind promotes them, instead of resolving to the wrong tag. TryRegister /
TryRebind refuse a stale session atomically under the park gate: a
SubscribeAsync whose reply straddled a reconnect is re-issued on the new
session (max 3 attempts), a subscribe-once read returns BadCommunicationError,
ResubscribeAllAsync stops if the session changes again. Old-session handle
numbers are never unsubscribed (Remove reports whether the entry belongs to
the current session; the read's late-reply path checks too), and an old
session's pump can no longer resolve through the map (fast path checks the
session).
I-3 (interim): a completed UnsubscribeBulk purges events parked for its
handles (driver + probe watcher). Full fix needs a worker sequence on the
SubscribeBulk reply (mxaccessgw, optional) — documented.
M-1: a resolved handle with a stranded parked list promotes it and resolves
the live event (never parks it). M-2: ResolveOrPark returns the ready events
to replay first, one gate acquisition per event while anything is held.
M-3: SubscribeAsync/read dispose the window before the orphan-release RPC.
M-5: ResolveOrPark remarks document the single-dispatching-pump invariant.
Tests (TestWait): overlapping windows don't accumulate, purge, stranded list,
session fence + map invalidation, stale-session registration refused,
end-to-end replay with swapped handle numbers and event-before-reply,
subscribe and read straddling a reconnect, read deadline closes its window.
Each was checked red against a mutation of the fix it pins.
Refs #859, #861.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Under load 40 refreshes could finish before any stall overlapped a report, failing the 'stress must have
run with deferred events' precondition. Keep refreshing (>=40, ceiling 90 s) until deferral is seen.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
EventQueueOverflowAclWireTests drives >4096 condition events at a folder notifier
item and a Server-object (i=2253) item with Security:Acl enforcement on, a
non-publishing client, and a second session reading a node-manager node throughout:
- denied (EqA-owned scripted) events interleaved with allowed native ones, under one
Lock hold and per-projection: no deadlock, reads stay bounded, deferral peak stays
within 2 x (MaxDeferredEventsPerCondition + 1), delivery <= item QueueSize, oldest
dropped / newest kept in report order, one EventQueueOverflowEventType published ahead
of the survivors (SDK DiscardOldest, Part 4 5.13.1.5), zero denied events delivered;
- a grant revoked (state.Apply -> revalidation, re-entering Lock) while a deferred tail
waits behind a full folder channel: neither the tail nor later events reach the
revoked session's folder or Server-object item; a still-granted session on the same
channel receives all of it.
Falsified: with enforcement off (and with the revocation removed) all three cases fail.
No production defect found.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
1. ReadAsync (subscribe-once) now opens the early-event window around its
SubscribeBulk + registration; an initial value that overtook the reply was
dropped and the read returned BadTimeout.
2. Register/Rebind write the reverse map and promote parked events in ONE
critical section under the park gate, and ResolveOrPark resolves under that
gate whenever anything is held (lock-free fast path only when nothing is),
so a later live event can no longer overtake the parked initial value. A
resolved handle that still holds a parked list queues behind it.
3. Held events are fenced by a session epoch: each new EventPump calls
StartNewEpoch(), discarding parked AND promoted-but-unreplayed events; an
older pump can neither park nor replay. Item-handle reuse across a reconnect
can no longer put an old-session value on a new tag.
4. ResubscribeAllAsync opens one window per subscription, not across the whole
replay.
5. Promoted-but-unreplayed (ready) events count toward MaxParkedEvents.
6. Begin/End of a window run under the park gate, so N->0 discard and 0->1
open are atomic (no inherited leftovers).
Documented that ParkedEventsReady may run under _registryGate/_syncLock and
relies on the channel's AllowSynchronousContinuations=false.
Tests use TestWait (#858); each new test was checked red against a mutation
of the fix it pins.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
mxaccessgw gives no reply-before-event ordering: the worker completes the STA
command via a RunContinuationsAsynchronously TCS, so the reply is framed on a
thread-pool continuation while the STA already pumps the OnDataChange Advise
queued (drained/framed by a separate loop), and the reply (unary Invoke) and
the event (StreamEvents) are separate HTTP/2 streams. The driver then also
registers handles concurrently with the pump's dispatch loop. MXAccess's
initial value could therefore be dropped as "stale after unsubscribe",
leaving the node at its pre-subscribe value until the next change.
SubscriptionRegistry gains an early-event window: every SubscribeBulk call
site (SubscribeAsync, ResubscribeAllAsync, PerPlatformProbeWatcher) opens
BeginPendingSubscribe() before the RPC and closes it after registering.
Inside the window unknown-handle events are parked (bounded, 100k); the
Register/Rebind that binds a handle promotes them, and the EventPump replays
them on its single dispatch thread before any later event for that handle
(woken by a sentinel when no further event arrives). Closing the last window
discards unclaimed events, so the stale drop is unchanged outside it.
New counters galaxy.events.early_{parked,replayed,overflow}.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Under heavy parallel load (14 concurrent test processes + 14 CPU burners, with
DOTNET_ThreadPool_ForceMinWorkerThreads=1 as the reproducer) the Galaxy tests
with 1-3 s fixed waits failed 16/70 runs. Every failure was a background loop
started with Task.Run (alarm feed, DeployWatcher, EventPump, ReconnectSupervisor)
not being SCHEDULED within the ceiling - e.g. DeployWatcher's stream had not even
been opened after 2 s. No production race: the code under test was checked for
each (alarm feed dispatch is synchronous in-order, DeployWatcher/ReconnectSupervisor
state is lock- or loop-local, subscribe registers before returning).
- New shared TestWait (UntilAsync / ForAsync / HealthyAsync): 30 s presence
ceiling, failure reports elapsed ms + an observed-state snapshot (#856 shape;
GalaxyDriverFixRound1Tests.WaitUntilAsync now delegates to it).
- Converted the short waits in GatewayGalaxyAlarmFeedTests, DeployWatcherTests,
GalaxyDriverSubscribeTests, GalaxyDriverRawPathKeyingTests (its WaitForAsync
never threw on timeout), EventPumpStreamFaultTests, EventPumpReplayGapTests,
ReconnectSupervisorTests; FixRound2 InitBudget 2 s -> 10 s (the parked fetch
never completes, so any finite bound keeps the verdict).
- Absence by short sleep replaced with positive controls: DeployWatcher bootstrap
and same-deploy-time (follow with a real change, assert exactly that one raise -
verified red when suppression is removed), subscribe failed-tag / unsubscribed
handle (follow with a live handle's event), faulted pump (wait for its
onStreamFault instead of 100 ms).
- Fixed sleeps that raced scheduling: subscribe-once read tests wait for the
read's registration (TrackedSubscriptionCountForTests) instead of 50 ms;
supervisor Dispose test waits for reopen entry instead of 50 ms; the
already-Healthy WaitForHealthy test asserts synchronous completion instead of a
wall-clock bound. FakeSubscriber.Map is now a ConcurrentDictionary.
Loaded run (14 parallel x 5, 14 burners, ForceMinWorkerThreads=1):
before 16/70 failed runs; after 0/350 (70 + 210 + 70). Without the thread-pool
override: 0/70 before and after on this host. Serial: 472 passed, 8 skipped.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
catthehacker/ubuntu:act-latest ships no iproute2, so fixture_host() fell back to 127.0.0.1 inside the
job container and every published-port fixture timed out (QuestDB is the first one probed over the
network). Read the default gateway from /proc/net/route in plain bash, warn when none is found, and log
the chosen host. Verified in that image on a user-defined network: gateway parsed, published port 200.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
The persistent NAT PKI (ae91e6b7) re-verified 2026-11-11: same thumbprint
1FBF8791... across --force-recreate on 10.100.0.35, matching the committed
opc-plc-nat-server.der; NAT class 7/7 against the shared stack and against a
CI-style throwaway container + docker-exec cert hand-off. MIN_EXECUTED stays
at 440: no hosted v2-integration run has reached the test step (#860).
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Found while load-verifying #856: ShutdownAsync_cancels_the_event_stream_and_releases_the_pump
timed out 7/700 under parallel load. The fake's StreamCancelled was set by a
CancellationToken registration scoped by `using` inside the async iterator.
On Cancel() the newer ReadAllAsync registration runs first and resumes the
iterator on a pool thread, whose `using` can dispose our registration while
Cancel() is still walking the callback list - so it never runs. Signal from
the iterator's finally (when the token is cancelled) instead, which also
proves the stream actually ended. 0/700 after.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
PerPlatformProbeWatcher records a platform's handle and publishes the probe
binding into the driver registry as one step under _syncLock, but
WatchedPlatforms read the handle map lock-free - so it could report a
platform one step ahead of the registry Upsert. A ScanState event raised in
that window was dispatched by the EventPump, missed ResolveSubscribers and
was dropped as "stale after unsubscribe", leaving the host status empty.
That is the #856 flake: A_fetch_completing_after_init_populates_host_statuses
waited for WatchedPlatforms, emitted ScanState, and timed out at 10 s on the
final Running wait. Reproduced 11/700 (28 parallel runners x 25 + 14 CPU
burners), every failure at that step with the binding present by the time of
diagnosis and hostStatuses=[] - i.e. the event was dropped. 0/1400 after.
WatchedPlatforms now reads under _syncLock (test seam only; the dispatch path
already checks the map after the registry, the safe order). New
PerPlatformProbeWatcherVisibilityTests spins a reader on another thread and
asserts the registry entry exists the moment the platform is visible
(old code: torn in 1691/2000 iterations). WaitUntilAsync gains a describe
overload; the #856 test's waits now report elapsed time and the observed
seed/binding/pump/host-status chain on timeout.
Claude-Session: https://claude.ai/code/session_01GT67AD99zQQqpLCVm3p7xk
Joseph Doherty