LocalDbSyncService observed only the call token, which ASP.NET Core cancels
at the END of Kestrel's graceful drain. A peer's sync stream never ends on
its own, so every graceful stop of the passive node waited out the host's
whole ShutdownTimeout (30 s on OtOpcUa QA).
The service now takes an optional IHostApplicationLifetime (DI supplies it
in any generic host): ApplicationStopping is linked into the stream's token
and ends the session cleanly, and a stream arriving once stopping has begun
is refused Unavailable. Wire-compatible; no proto/schema change.
Tests: open stream ends within 5 s of StopApplication, new stream refused
Unavailable, control without a lifetime still running; both halves
mutation-killed. 262/262. Source only, NOT published (still 0.3.7).
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
Re-ran the two design 9 checks that were blocked for OtOpcUa on 2026-07-24, against a
docker-dev rig from OtOpcUa master 412da27c (Health 0.3.0) with the dashboard in a
container on the rig network:
1. one Active per pair: exactly one /health/active 200 per mesh (MAIN central-2,
SITE-A site-a-1, SITE-B site-b-1), dashboard shows one ACTIVE + one STANDBY each;
2. leader chip on every group; network-partitioning site-b-2 from site-b-1 produced
two 1-member clusters, both Active, and the SPLIT BRAIN chip, which cleared once
site-b-2 was restarted (healing the link alone does not merge two Akka clusters,
the documented OtOpcUa recovery); stopping site-a-1 moved the SITE-A chip.
Evidence (screenshots, /health/active bodies, timeline) under
docs/plans/evidence/2026-11-12-overview-phase4/. Ledger task 20 -> completed (20/20).
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
ZbOtelCollector (otelcol-contrib 0.162.0, NSSM, NT SERVICE\ZbOtelCollector, every
listener on 127.0.0.1, file + Prometheus exporters, size-capped rotation) on DEV
vd03 and QA vq03/vq04. OtOpcUa, ScadaBridge Central + Site, MxAccessGateway and
HistorianGateway export OTLP to http://127.0.0.1:4317 via two service-env entries
each; rolling restarts, standby-first on the QA pairs. Overview has no exporter
binding and is unchanged. Config, install/switch/verify scripts, runbook and the
dated rollout record under infra/otel-collector/wonder/; evidence doc, GAPS,
index row and deployment.md updated.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
The Serilog OTLP log sink is not part of the OTel SDK and does not run
inside SuppressInstrumentationScope, so every log batch produced a root
opentelemetry.proto.collector.logs.v1.LogsService/Export gRPC client span.
AddZbTelemetry now drops telemetry's own transport from tracing:
- HttpClient instrumentation: FilterHttpRequestMessage rejects export calls
- gRPC client instrumentation (no Filter option): the start-time
EnrichWithHttpRequestMessage callback clears Recorded/IsAllDataRequested,
which also leaves the child HTTP activity unsampled.
An export call is any /opentelemetry.proto.collector.* path, or an
OTLP/HTTP /v1/{traces,metrics,logs} path on the configured endpoint's
host:port (ZbOtlpExportCallFilter). Every other outgoing call keeps its span.
Tests: predicate theory + an end-to-end run against a loopback OTLP
receiver with positive controls (app gRPC + HttpClient calls traced) and
negatives (the real log-sink export, /v1/logs, gRPC path over HttpClient);
mutation-checked three ways.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi