Owner scope 2026-11-12 (Windows x64 + Linux, unchanged cycle-3 caps, >=7200 s, SDK 1.5 comparator).
- Linux ARM64 (late-slope, 7200.12 s, 1964 cycles / 491 restarts, shutdown verified, analyzer-validated):
first rejection RSS at cycle 247 / 900.73 s, 358,682,624 > 358,678,528 (+4 KiB); peak 382.98 MiB vs
limit 342.06 MiB; late plateau medians 344.4/341.1/341.9 MiB, OLS -1.90 MiB/h. FAIL.
- Windows x64 on windev (same workload, working set as RSS, handle table as descriptors): first rejection RSS
at cycle 285 / 1130.76 s, 303,505,408 > 302,174,208; peak 319.49 vs limit 288.18 MiB; plateau ~314-315 MiB,
+0.93 MiB/h. FAIL. Also incomplete: the launcher's wall-clock outer deadline fired early after a +552 s
w32time step and ended the run at test-phase 7014 s (rejection had already latched); launchers now use a
Stopwatch.
- SDK 1.5.378.182 comparator (efb9197a publication, Windows, 7201 s): RSS peak +65.8 MiB (never over its cap),
plateau ~+48 MiB, but the +32 handle cap tripped at cycle 17 (max 674 vs 642).
- Managed flat on both SDK2 runs; growth is the generated-factory JIT burst (+40k/+44k methods) - Linux JIT
doublemapper ~180 MiB; Windows private bytes grew ~25 MiB of a ~132 MiB working-set rise.
- Retained: b61f3e05 cycle-1 fixture failure (#808 unplaced UNS folder), Windows/SDK1 smoke diagnostics.
No cap raised, no warm baseline, no failed run counted as passing.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
GalaxyAlarmPriority, the one table live Galaxy alarms (MxAccessSeverityMapper)
and AVEVA A&E history (EventMapper, #886) map the AVEVA priority through, ranked
a higher number as more severe. AVEVA ranks priority 1 as the most urgent, so a
priority-1 alarm read 250 and a priority-999 system event 1000.
It now uses AVEVA's default priority ranges, most urgent first:
AVEVA priority AVEVA severity bucket OPC UA Severity
1-250 1 Critical Critical 1000
251-500 2 High High 750
501-750 3 Medium Medium 500
751-999 4 Low Low 250
(history only) 0 not an alarm - 1
Out-of-range priorities clamp (<= 0 Critical, > 999 Low). Evidence, vd03
Runtime.dbo.Events read-only: priority 1/100 alarms carry AVEVA severity 1,
400/500 severity 2, 600/700 severity 3, and every priority-999 System.* /
User.* row severity 0 with IsAlarm 0.
- Live and history now use the same number. GalaxyDriver passes the table's
exact value as AlarmEventArgs.OpcUaSeverity. Before, it passed only the bucket,
which NativeAlarmProjector re-mapped through its generic 200/500/700/900
ladder, so live and history disagreed even before the direction fix.
- History: a row whose AVEVA severity is 0 (System.Start/Stop/Deploy,
User.Write...) is not an alarm and reads Severity 1, below every alarm band.
For alarms AVEVA's 1-4 severity is still never the value: the live feed
carries only the priority, so it cannot use it, and history must agree.
- Docs: docs/Historian.md gains "Galaxy alarm priority -> Severity
(lmxopcua#887)"; docs/AlarmTracking.md updated.
- Live re-run of the #801/#886 harness against vd03 HistorianGateway
ea9f21c (read-only, network-only): 224 of 233 cases identical to the #886
run, the 9 others differ only in Severity (priority-700 alarm rows 750 ->
500, 19 system/user events 1000 -> 1). Record: aveva-history-801.md
"#887 re-run", report aveva-gate-sdk2-887.json.gz.
Client-visible: every OPC UA client that ranks or filters Galaxy alarms by
Severity sees the order reversed (live and HistoryReadEvents). ScadaBridge has
no code that depends on the old values; its own mixed severity directions are
ScadaBridge#129.
Tests updated because they pinned the inverted table:
- Driver.Galaxy.Tests MxAccessSeverityMapperTests (rewritten: edges 1/250/251/
500/501/750/751/999, 0/-100/int.MinValue/1000/int.MaxValue clamps,
monotonic-decreasing sweep, equals GalaxyAlarmPriority for -2..1001)
- Driver.Galaxy.Tests GatewayGalaxyAlarmFeedTests (750 and 600 -> Medium)
- Driver.Historian.Gateway.Tests EventMapperTests (band table, priority 999
-> 250, 1 -> 1000; system events -> 1)
- OpcUaServer.IntegrationTests GalaxyEquipmentEventHistoryWireTests (700 ->
500, 100 -> 1000, System.Stop -> 1; the fake now stores AVEVA's real 0-4
severity)
New tests:
- Runtime.Tests NativeAlarmRestatementEndToEndTests
.Galaxy_live_condition_severity_is_the_AVEVA_priority_band (8 rows: the real
feed -> real GalaxyDriver -> host -> projector; the same rows EventMapperTests
pins on the history path)
- EventMapperTests.Aveva_severity_zero_marks_an_event_that_is_not_an_alarm
(5 rows, incl. the priority-999 alarm positive control)
- GalaxyDriverAlarmSourceTests: OpcUaSeverity passes through
Mutation check, 8 mutants, all killed: the pre-#887 inverted table; the
driver dropping OpcUaSeverity; no not-an-alarm rule; each of the three band
edges off by one; history ignoring priority; a missing AVEVA severity read
as 0.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
A notifier folder read one event source, its NodeId identifier (a RawPath or an EquipmentId). AVEVA stores a
Galaxy object's A&E history under the object's tag_name (Source_Object), so AVEVA's own alarm and event
history reached no production folder. Its lower-case property keys were also missed, so every row read
Severity 1 with no ConditionName.
Sources. A folder's event history now reads its own source plus the Galaxy object of every native condition
that fires on it live: those parented under it and those WireAlarmNotifiers wired to it. Live and history use
the same membership. The object comes from the condition's Galaxy alarm reference, its raw tag's attributeRef,
up to the first '.':
- The composer reads it with the driver's own parser (GalaxyAttributeRef, shared with GalaxyDriver), only for
a GalaxyMxGateway tag with an alarm (RawTagPlan.NativeAlarmReference, part of plan equality).
- The applier binds it after materialising the condition (IOpcUaAddressSpaceSink.BindNativeAlarmHistory,
forwarded by the deferred sink). A rebuild or the condition's removal forgets it.
Each object is read with the #882 rule: its own events and <Object>.<Attribute> rows, never a prefix sibling.
Fan-out. historiangw's ReadEvents filter is a single predicate, so there is one request per source; the wire
contract is unchanged. EventSourceFanOut merges the reads as one source:
- order is (time, source position, arrival), reversed for a reverse read; ties are never re-sorted by EventId,
so the boundary skip stays stable;
- duplicates by EventId are dropped and the merge is cut to the cap;
- a truncated source bounds the page at the time it stopped.
Paging, continuation points (bound to the whole source set), the walked-back reverse read and the tie-cluster
over-fetch are unchanged. A folder with one source is read exactly as before.
ACL. Every source is read inside the folder's own HistoryRead, after AuthorizeHistoryRead (#826). A denied
folder reads nothing and fails closed.
Mapping (EventMapper, property names now case-insensitive):
- Severity: an AVEVA row's priority goes through GalaxyAlarmPriority, the table live Galaxy alarms use
(MxAccessSeverityMapper now delegates to it), so history matches live. Otherwise severity (OtOpcUa's own
rows, which AVEVA stores lower-cased), clamped. AVEVA's own 1-4 severity is ignored. The table ranks a
higher priority as more severe, which inverts AVEVA's convention: filed as #887.
- An AVEVA row whose source_conditionvariable names a native condition that fires on the folder takes that
condition's SourceNode and live ConditionName. Otherwise ConditionName is alarmname or the condition
variable's attribute.
- alarm_id/alarmid map to AlarmId.
Tests:
- GalaxyEquipmentEventHistoryWireTests (4, real SDK session, production Gateway backend, AVEVA-shaped fake
gateway): equipment EqA backed by objects Oven_007 + OvenDoor_07. Covers object and attribute events from
both objects, the prefix sibling excluded, cross-object and in-object ties, forward paging at 1-4 a page,
reverse whole and paged, SourceNode/ConditionName/Severity, the raw device folder, ACL deny with no backend
call plus a granted positive control, and a rebuild forgetting the bindings.
- EventSourceFanOutTests (8), EventMapper AVEVA cases (11), GalaxyAttributeRefTests (16), a composer test and
an applier ordering test.
- Mutation-checked, 17 mutants, all killed. Covered: parent-only membership, no scope, a case-sensitive
condition match, no condition rewrite, no truncation bound, EventId tie order, no sort, no cap, no stop at
the bound, unreversed source order, no dedupe, no '.' boundary, priority ignored, case-sensitive keys, no
applier bind, no deferred forward, and a composer that ignores the driver type. The forget-on-rematerialise
guard survives as belt-and-braces: the rebuild clear covers it.
- Live, read-only, no WinRM, vd03 HistorianGateway ea9f21c: equipment folder EQ-886-ZipsenLeft (objects
DevWinPlatform + Left_001). It read 0 events before and 92 now, equal to the unfiltered read's rows for both
objects. 1/7-event pages and reverse whole/5-event pages agree, and 73 rows resolve to their live condition.
diff_runs vs the #882 run: 221/226 identical. The 5 that differ are expected: DevWinPlatform Severity
1 -> 750/1000, and #881 capabilities.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
SDK 2.0.1's UserNameIdentityTokenHandler encrypts a password longer than 64 bytes
(a private RsaEncryptedSecretPasswordThreshold, no setting) as an RsaEncryptedSecret
with EncryptionAlgorithm=null. SDK 1.5.378 (.106 and .182) decrypts only the legacy
format under RSA policies; its SecurityPolicies.Decrypt answers a null algorithm with
BadIdentityTokenInvalid ("Unexpected encryption algorithm") before any key check.
Every production API key is 65 chars, so no SDK 2 OtOpcUa client could use one
against QA (b8b4c18f), a rollback target or an older third-party server.
LegacyRsaUserNameIdentityTokenHandler (Driver.OpcUaClient.Session, plugged in through
the public UserIdentity(IUserIdentityTokenHandler) ctor) always uses the legacy
format, RSA-OAEP(len|password|serverNonce) under the server certificate, under an RSA
policy. That is what SDK 2 sends for passwords of 64 bytes or less and what SDK 1.5
clients always send. The legacy format has no size limit (the plaintext is
length-framed over as many RSA blocks as it needs), and an SDK 2 server reads both
formats. Policy None, the ECC policies and a missing certificate go to the SDK
handler unchanged.
OpcUaUserIdentityFactory.UserName is now the one UserName identity the driver, the
AdminUI browse and Client.Shared (CLI/UI) present. Client.Shared references .Session
for it.
Tests: handler unit tests (all three RSA policies, the 1.5-required algorithm URI,
SDK 2 server-side decrypt, the session Copy() keeps the type, None unchanged, and an
SDK control pinning the 64/65-byte switch); a Client.Shared identity test; and an
SDK2<->SDK2 integration test where a 65-char key is admitted in both formats and a
wrong 65-char key reaches the key check (BadIdentityTokenRejected). The #800 Windows
gate's SDK 2 CLI 65-char case now gates against the 1.5 server too.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
HistorianGateway answers a per-source ReadEvents (Source_Object = DevWinPlatform) with the
object's own events (SourceName DevWinPlatform) AND its attribute alarms, which it reports as
<Object>.<Attribute> (Source_ProcessVariable, e.g. DevWinPlatform.CPULoadAvg). The Gateway
backend's defensive client-side source filter compared SourceName for equality, so it dropped
the attribute alarms: 70 of the 72 DevWinPlatform events on vd03.
The filter (and the past-the-cap "more events" signal) now uses EventSourceMatch: the source
itself, or <source>.<anything> on the '.' boundary, ordinal. DevWinPlatform never matches
DevWinPlatform2 or DevWinPlatform2.X. This needs no wire change and still makes one request.
Each event keeps its own SourceName. SourceNode stays the condition its stored alarm id names,
which is null for AVEVA rows. The attribute events ride the notifier's own read, so the
per-folder HistoryRead verdict (#826) governs them: a denied folder answers
BadUserAccessDenied before any backend call. The SqlServer and QuestDB backends still match
exactly; they hold only OtOpcUa's own events.
Tests:
- GatewayAttributeEventSourceTests (16): the boundary theory, attribute-in/sibling-out through
the data source, the drain signal, and an exact attribute-named read.
- AttributeEventHistoryWireTests (2, over a real SDK session with the production Gateway
backend on an AVEVA-shaped deterministic gateway): object + attribute events with their own
SourceName, a prefix sibling excluded, paging, and the ACL-on deny with no backend call plus a
granted positive control.
- Mutation-checked: exact match fails both wire tests; dropping the '.' boundary fails the
wire sibling check and 6 unit tests.
- Live AVEVA harness re-run against vd03 HistorianGateway ea9f21c (read-only, no WinRM):
folder DevWinPlatform 2 -> 72 events (whole, 72 1-event pages, reverse); the new #882 check
passes; 221 of 225 shared cases are identical to the #801 SDK2 baseline, and the 4 that
differ are exactly the DevWinPlatform event cases. Recorded in aveva-history-801.md.
Follow-up #886: production folders name RawPaths/EquipmentIds, not Galaxy objects, plus the AVEVA
severity/condition property mapping.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
Both builds served HistoryRead from the real AVEVA Historian via the DEV
HistorianGateway (vd03, ea9f21c): 225 cases / 3,855 rows / events, 0 differences.
Raw also matches direct SQL on Runtime.dbo.History (6/6) and histgw-cli; simple
aggregates match a recomputation from raw (336/336). Modified, HistoryUpdate,
seven refused aggregates and HistoryServerCapabilities answer as before. No write
reached the historian. Findings (pre-existing, not SDK): all HistoryServerCapabilities
booleans false; AVEVA attribute-alarm events unreachable per source.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
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
Measured on windev (3 alternated sessions) and macOS: a pool resume in place
of the dispatcher, and an inline-on-reader upper bound, with the harness as
committed and with its paced continuation made synchronous.
- pool: 10k/s host CPU unchanged, idle p50 and idle CPU worse (macOS idle CPU
x3 from pool spin); only the closed loop gains. Not kept.
- inline (unsafe for every seam; the ceiling for a VirtualTag-only opt-in):
~8 points of a core at 10k/s raw and ~50 us idle p50 on windev with a cheap
continuation, nothing with the harness as committed. Both budgets still
missed by a wide margin, so the opt-in was not built.
No code change. Records the remaining attribution (harness pool hop ~15
points raw; reader wake + syscalls; worker side untouched).
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
§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
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
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
#816 finding: the SDK checks that the server certificate names the host only
when a session is created with checkDomain:true. The driver, the AdminUI
browser and Client.Shared all called the 8-argument
DefaultSessionFactory.CreateAsync, which forwards checkDomain:false — so no
host-name check ever ran (why a trusted NAT cert for *.invalid browsed fine
over an IP in #813). When enabled the SDK (Session.OpenAsync ->
ValidateDomains -> FindDomain) compares ConfiguredEndpoint.EndpointUrl.IdnHost
— the CONFIGURED host, because OpcUaEndpointSelection/EndpointSelector
rewrote the advertised one — against the cert's DNS/IP SANs; secured
sessions only, and the auto-accept handler accepts a mismatch.
Policy: certificateValidation.validateServerHostName, tri-state. Unset =
OpcUaCertificateValidationOptions.DefaultValidateServerHostName = false, the
pre-#816 behaviour, so nothing changes on upgrade; true = validate the
configured host; false = explicit NAT/port-publish opt-out. Threaded into the
driver, the AdminUI browser (same JSON) and the AdminUI form (a select that
keeps "unset" unset), Client.Shared ConnectionSettings.ValidateServerHostName
and the CLI --validate-host-name. A driver refusal names the opt-out.
#818: the NAT fixture's /app/pki is on the opc-plc-nat-pki volume (verified:
same thumbprint across --force-recreate), its PUBLIC cert is committed as
Docker/nat/opc-plc-nat-server.der, and NatPublishedEndpointBrowseTests run
every secure browse on a throwaway PKI (internal browser seam): refusal with
nothing trusted, positive with the cert trusted (unset/false host check), and
a BadCertificateHostNameInvalid refusal with the check on. The trusted cert is
OPCUA_NAT_SERVER_CERT or the committed one (skip on thumbprint mismatch);
OPCUA_NAT_SERVER_CERT_TRUSTED is retired. CI exports its throwaway NAT
container's public cert. ServerHostNameValidationTests pin the driver against
the plain opc-plc. MIN_EXECUTED not raised (needs a hosted run): +6 in
v2-integration (440 -> 446), documented in scripts/ci/README.md.
docs/drivers/OpcUaClient.md, docs/plans/2026-05-28-driver-browsers-design.md and
docs/Client.CLI.md: one OpcUaEndpointUrlRewriter for driver, AdminUI browser and
Client CLI/UI; configured host+port win, advertised path kept for opc.tcp; policy/mode
selection and certificate validation unchanged; NAT test recipe against the #813 fixture.
Joseph Doherty