lmxopcua/docs/Telemetry.md

# Live Telemetry Transport (v2)

## Overview

The AdminUI's live observability panels (`/alerts`, `/script-log`, `/hosts` driver table, driver
resilience status) are fed by four node→central event channels. Historically all four rode
DistributedPubSub (DPS) on the shared Akka mesh — the same gossip ring that carries redundancy
state and the command-plane topics. Per-cluster mesh **Phase 5** adds a second transport for these
four channels only: one gRPC server-streaming contract, selected per node/central pair by
`Telemetry:Mode` / `TelemetryDial:Mode` (`Dps` default | `Grpc`).

This is the same motivation as [Phase 2's `MeshTransport`](Configuration.md#meshtransport-centralnode-command-transport)
and [Phase 3's `ConfigSource`/`ConfigServe`](Configuration.md#configsource--configserve-config-fetch-and-cache):
DPS only works when central and the node share a gossip ring, and [Phase 6](plans/2026-07-21-per-cluster-mesh-design.md)
splits the fleet into one Akka mesh per application `Cluster`. Once that split lands, a driver node
in a site's mesh is no longer a cluster member of central's mesh, so DPS can no longer reach it —
telemetry needs its own out-of-band transport, same as commands (Phase 2) and config bytes (Phase 3).

## Direction: the node hosts, central dials

**Load-bearing and easy to get backwards:** each **driver** node hosts the telemetry gRPC server
(a dedicated Kestrel h2c listener); **central (admin role) is the client** and dials in. This
mirrors the inversion ScadaBridge already uses for the same problem (its `SiteStreamService`).

```
 driver node                          central (admin)
┌─────────────────────┐              ┌──────────────────────────┐
│ 4 publish seams      │              │ TelemetryDialSupervisor   │
│  -> node-local hub    │   Subscribe  │  - discovers ClusterNode   │
│  -> DPS (unchanged)   │◄─────────────│    rows (Host+GrpcPort)   │
│                       │   (gRPC      │  - one reconnecting       │
│ TelemetryStreamGrpc-  │    stream)   │    dialer per node        │
│  Service (h2c server) │─────────────►│  - feeds the SAME sinks   │
│  Telemetry:GrpcListen- │  TelemetryEvent │   the DPS bridges feed   │
│  Port                 │  (oneof, 4 kinds)│   today                  │
└─────────────────────┘              └──────────────────────────┘
```

A fused `admin,driver` node both hosts (as driver) and dials (as admin) — it dials itself plus its
redundant pair peer, the same way central dials site nodes.

**This same driver-role listener also hosts the mesh gRPC command-plane ingress.** Whenever
`Telemetry:GrpcListenPort > 0`, the node maps both `TelemetryStreamGrpcService` (this stream) and
`NodeCommandService` (central→node commands under `MeshTransport:Mode=Grpc` — see
[`docs/Configuration.md` § `MeshTransport`](Configuration.md#meshtransport-centralnode-command-transport))
on the one port. They share it rather than each claiming its own, since both are dialled by the same
central admin nodes and a second cleartext h2c port would buy nothing but another firewall rule.
`NodeCommandService` is gated by its own `MeshCommandAuthInterceptor`/`GrpcPsk` bearer, independent
of this stream's `Telemetry:ApiKey`.

Central discovers driver-node telemetry endpoints from the `ClusterNode` table (`Host` +
`GrpcPort`, the latter added in Phase 1 for exactly this purpose) — the same DB-sourced discovery
Phase 1's ack set and Phase 2's gRPC command dial already use (see
[`docs/Configuration.md` § `MeshTransport`](Configuration.md#meshtransport-centralnode-command-transport)).
No shared gossip membership is required to find or dial a node.

## The dark switch

`Telemetry:Mode` (node/serve side) and `TelemetryDial:Mode` (central/dial side) each default to
`Dps` and can be independently set to `Grpc`. Both code paths are compiled into every binary —
flipping either flag is an appsettings/env change plus a restart, **not a redeploy or rebuild** —
the same discipline as `MeshTransport:Mode` and `ConfigSource:Mode`.

What actually changes per mode, precisely:

- **The node ALWAYS does both things, regardless of `Telemetry:Mode`.** Every one of the four
  publish seams emits into the node-local `ITelemetryLocalHub` **and** publishes to DPS,
  unconditionally. The node additionally **always** hosts the gRPC server whenever
  `Telemetry:GrpcListenPort > 0` — hosting is not mode-gated at all. `Telemetry:Mode` only affects
  central's own bookkeeping about which side it expects to be consulted (it is otherwise inert on
  the node).
- **Only central's ingest source switches**, driven by `TelemetryDial:Mode`:
  - `Dps` (default) — today's four DPS SignalR bridges (alert, script-log, driver-status,
    resilience) subscribe and feed the AdminUI sinks. Unchanged behavior.
  - `Grpc` — those four DPS bridges are **not spawned**. Instead the `TelemetryDialSupervisor`
    actor dials every driver node's gRPC stream and feeds the **identical** in-process sinks
    (`IInProcessBroadcaster<AlarmTransitionEvent>`, `IInProcessBroadcaster<ScriptLogEntry>`,
    `IDriverStatusSnapshotStore`, `IDriverResilienceStatusStore`).
- **The AdminUI panels themselves are untouched.** `/alerts`, `/script-log`, and the `/hosts`
  driver table all read from the same sinks in both modes — only the sinks' upstream feed swaps.

Because hosting is unconditional and both ingest paths are always compiled in, a fleet can be
flipped node-by-node and central-by-central with no coordination window where telemetry is lost —
the node is always serving, so central can switch to `Grpc` whenever it likes.

## The four migrated channels

| DPS topic (today) | Domain record | Central sink fed |
|---|---|---|
| `alerts` | `AlarmTransitionEvent` | `IInProcessBroadcaster<AlarmTransitionEvent>` (+ `AlertHub`); `/alerts` page; live/disconnected pill |
| `script-logs` | `ScriptLogEntry` | `IInProcessBroadcaster<ScriptLogEntry>` (+ `ScriptLogHub`); `/script-log` page; live pill |
| `driver-health` | `DriverHealthChanged` | `IDriverStatusSnapshotStore`; `/hosts` driver table |
| `driver-resilience-status` | `DriverResilienceStatusChanged` | `IDriverResilienceStatusStore` |

These four are carried as `oneof` event kinds on one `TelemetryStreamService.Subscribe` RPC — a
single stream per (central, driver-node) pair carries all four, rather than one stream per topic.
Proto field evolution is additive-only (never renumber/reuse a tag), locked by a contract test that
reflects over the `oneof` cases.

### `driver-health` is keyed per driver instance, not per Akka member

`DriverHealthChanged` carries no per-Akka-member identity — a snapshot names the driver instance it
describes, never the node that published it. `IDriverStatusSnapshotStore` therefore keys on the
driver instance, and the `/hosts` driver table renders **cluster-scoped** rows: one entry per
configured driver instance, grouped under its cluster, rather than one entry per (member, driver)
pair. A pair's two nodes running the same deployed driver contribute to the same row, with the most
recent snapshot winning.

That is the correct granularity for what the channel reports — a driver instance's transport health
is a property of the instance's configuration and its device, not of which member happens to host
it — but it does mean the panel cannot answer "which node saw this?". Adding that would mean
widening the wire record, not changing the page. The one sub-instance breakdown the channel *does*
carry is per-host connectivity, below.

### Per-host connectivity rides `driver-health` (Gitea #521)

`DriverHealthChanged.HostStatuses` carries each driver's `IHostConnectivityProbe.GetHostStatuses()`
result, rendered as the **Hosts** column on `/hosts`. It exists to show the one thing the driver-level
state chip structurally cannot: a **multi-device driver stays aggregate-`Healthy` while one of its
devices is unreachable**. A FOCAS or AbLegacy instance owning several PLCs previously hid that
entirely.

Three things to know before touching it:

- **This deliberately does NOT go through SQL.** A `DriverHostStatus` table existed, with an entity
  doc describing a publisher hosted service that upserted rows from each driver node. That publisher
  was never written, and **per-cluster mesh Phase 4 made it unbuildable as described** — `Program.cs`
  gates `AddOtOpcUaConfigDb` on the `admin` role, so a driver-only node has no ConfigDb connection to
  write rows with. The table was dropped (migration `DropDriverHostStatusTable`); it was empty on
  every deployment. This channel needs no DB and already survives the mesh split.
- **null ≠ empty, on the wire too.** Null means "the driver has no probe"; an empty list means "it has
  one that currently knows no hosts". proto3 cannot distinguish an absent repeated field from an empty
  one, so `DriverHealth.has_host_statuses` carries that bit explicitly. Collapsing the two would render
  every probe-less driver as one whose devices are all fine.
- **The host digest is part of the publish fingerprint, and must stay there.** `PublishHealthSnapshot`
  dedups on that fingerprint, and on a single-host-down transition *every other component is
  unchanged* — so without it the dedup swallows precisely the publish carrying the news. Same trap
  that bit the rediscovery signal; pinned by
  `DriverInstanceActorHostStatusTests.A_single_host_going_down_is_not_swallowed_by_the_unchanged_health_dedup`.
  The digest is a flattened, host-name-ordered **string** for the converse reason: a tuple holding an
  `IReadOnlyList` compares by reference and would re-publish on every 30 s heartbeat forever.

⚠️ **`DriverInstanceResilienceStatus` is the same defect, still open.** That table also has no writer
and no reader — the only reference in the repo is its DbSet declaration — while the live data reaches
the AdminUI over the `driver-resilience-status` channel above. Keep-or-delete is Gitea **#524**.

## The three non-migrated channels (NOT migrated in Phase 5 — do not read this as "seven done")

The program sketch originally named seven observability topics for Phase 5. Three were scoped out,
each for a distinct, settled reason — and one of those three has since been deleted outright:

- **`redundancy-state`** — bidirectional, built directly from `Cluster.State`, and **pair-local**:
  it drives ServiceLevel and the Primary data-plane gate, consumed in-process by
  `OpcUaPublishActor`, `ScriptedAlarmHostActor`, `DriverHostActor`, and `HistorianAdapterActor`.
  It stays on DPS in **both** `MeshTransport` modes today, and it is genuinely mesh-bound by
  design — under Phase 6 each pair keeps sharing its own small mesh, so DPS keeps working for it
  in-mesh. Central's *display* of each pair's redundancy state is a Phase 6 cross-mesh concern
  (a future observability channel, once central no longer shares gossip with any site pair), not a
  Phase 5 telemetry-panel migration.
- **`fleet-status`** — **DELETED, 2026-08-06, Gitea #542.** There is no `fleet-status` channel any
  more, on DPS or anywhere else. It was central-internal rather than a node→central stream: the
  admin singleton `FleetStatusBroadcaster` built it from the admin node's own cluster membership/
  reachability/leader events and published to a SignalR hub that **no client anywhere ever
  connected to** — `Fleet.razor` polls `NodeDeploymentState` in the Config DB and always ignored
  the feed. Its second input, `DriverHostStatusHeartbeat`, had no production sender either. Phase 6
  then made the central-only view wrong as well (central's gossip ring is its own pair). The whole
  chain — broadcaster, singleton registration, DPS topic, SignalR bridge, `FleetStatusHub`, and the
  `FleetStatusChanged`/`FleetNodeStatus` messages — was removed rather than rewritten cross-mesh.
  Fleet visibility that does work today: `/fleet` over `NodeDeploymentState`, and the four Phase 5
  telemetry channels above, which are per-node dials and therefore already mesh-independent.
- **`deployment-acks`** — already migrated, but onto a different transport: it rides the Phase 2
  gRPC command transport (`MeshTransport:Mode=Grpc`, `CentralCommandService.ReportApply`) as a
  command-plane reply, not an observability broadcast. It was never a Phase 5 candidate.

See [`docs/Redundancy.md`](Redundancy.md#command-transport-centralnode) for how `redundancy-state`
and the command transports fit together.

## Authentication — fail-closed from day one

A shared node bearer key gates the stream: `Telemetry:ApiKey` (node/serve side) must equal
`TelemetryDial:ApiKey` (central/dial side). `TelemetryStreamAuthInterceptor` enforces it — path-scoped
to `/telemetry.v1.TelemetryStreamService/`, comparing the `Authorization: Bearer` token with
`CryptographicOperations.FixedTimeEquals`, and rejecting with `PermissionDenied`. An **unset** key
rejects every call rather than allowing an open stream — the same fail-closed posture as
`ConfigServeAuthInterceptor` and `LocalDbSyncAuthInterceptor`.

This **supersedes** [the design doc's §6.3](plans/2026-07-21-per-cluster-mesh-design.md), which had
provisionally decided to "match ScadaBridge's unauthenticated posture for now" for inter-cluster
transports. ScadaBridge itself has since closed that gap with this identical
interceptor pattern, so Phase 5 ships authenticated from the start rather than deferring auth to a
later hardening pass. See the design doc's superseded note for the full history.

**On a fused `admin,driver` node, `TelemetryStreamService` and `NodeCommandService` are `AllowAnonymous`
by design** — a *different* gate from the interceptor above. `TelemetryStreamAuthInterceptor` is the PSK
check; on an admin-role node the AdminUI cookie `FallbackPolicy` sits in front of it and would 401/redirect
a correctly-keyed call before the interceptor's verdict matters, because it runs in the authorization
middleware, strictly *before* the interceptor executes (archreview 03/S14). A driver-only node never
registers that middleware at all, so it is structurally unaffected either way. See
[`security.md` § Mesh gRPC ingress is `AllowAnonymous` by design](security.md#mesh-grpc-ingress-is-allowanonymous-by-design).

## Upgrading an existing deployment: populate `ClusterNode.GrpcPort` first (#493)

`GrpcPort` is a **nullable column added by Phase 5**, and nothing backfills it onto rows that
already existed. On an upgraded deployment every `ClusterNode` row therefore carries `NULL`,
central finds **no dial targets**, and flipping `TelemetryDial:Mode` to `Grpc` connects to
nothing. Fresh installs are unaffected. `AkkaPort` did not have this problem only because it is
`NOT NULL` with a default of 4053.

Nothing fails loudly, which is what makes this worth stating up front — the code degrades
gracefully, once per node, throttled:

```
ClusterNode <id> has no GrpcPort; it exposes no telemetry stream and is skipped in this dial refresh
(further skips of this node are silent)
```

Other symptoms: no `ESTABLISHED` sockets on the telemetry port (a `LISTEN` but nothing else in
`docker exec <node> cat /proc/net/tcp6`), and the `/alerts` / `/script-log` pill never turning
live in `Grpc` mode.

**Before flipping any node to `Grpc`**, set each driver node's `GrpcPort` to that node's own
`Telemetry:GrpcListenPort` — via the AdminUI node editor, or directly:

```sql
UPDATE dbo.ClusterNode SET GrpcPort = <that node's Telemetry:GrpcListenPort> WHERE NodeId = '<node>';
```

There is deliberately **no EF data migration** backfilling this. The correct value is that node's
own listener port, which is per-deployment configuration the database cannot derive; a migration
could only invent one, and a *wrong* dial target is worse than the `NULL` the dialer already skips
explicitly. The docker-dev seed does backfill (`GrpcPort IS NULL AND CreatedBy = 'docker-dev-seed'`
→ 4056) because there the port is fixed by `docker-compose.yml` and therefore actually knowable.

## Reconnect and reliability

Central's `TelemetryDialSupervisor` (an admin-role actor) runs **one reconnecting dialer per
driver node**:

- **Discovery** — enabled, non-maintenance `ClusterNode` rows are read on start and refreshed every
  `TelemetryDial:ContactRefreshSeconds` (default 60), plus on an admin-side topology change. A row
  with a null `GrpcPort` is skipped (that node exposes no telemetry endpoint — logged, not an
  error). Dial targets are added/removed as the row set changes.
- **Reconnect backoff** — immediate first retry, then a fixed ~5 second backoff, indefinitely. The
  dialer never gives up permanently: this is an observability channel, not the data plane, so a
  persistently-unreachable node just means a persistently-stale panel, not a fleet fault.
- **Generation stamping** — each dialer carries a monotonically-increasing generation counter, so
  a late error or late event from a superseded stream (e.g. one that raced a reconnect) is
  recognized and ignored rather than corrupting the current stream's state.
- **Snapshot replay on reconnect** — the node-local hub keeps a last-value cache for
  `driver-health` and `driver-resilience-status` (there is nothing meaningful to replay for
  `alerts`/`script-logs`, which are append logs). On every new `Subscribe` — including a
  reconnect — the hub drains the cached snapshots to the new stream before live deltas, so
  `IDriverStatusSnapshotStore` / `IDriverResilienceStatusStore` re-prime immediately rather than
  sitting stale until the next natural event. `alerts`/`script-logs` simply tolerate the gap — a
  reconnect loses whatever transitions happened while disconnected, same as a DPS resubscribe would.
  The cache is evicted per instance (`ITelemetryLocalHub.Evict`) when `DriverHostActor.StopChild` removes a
  driver from the plan (#609), so a decommissioned instance's last state is not replayed to new subscribers.
  `StopChild` first releases the instance's resilience state (`IDriverCapabilityInvokerFactory.Release` →
  pipeline-builder + `DriverResilienceStatusTracker.Invalidate`): `DriverResilienceStatusPublisherService`
  re-emits the whole tracker every 5 s, so evicting only the hub was undone within one tick (review R1 M1).
  Residual: a Health emit already in flight from the stopping child can land after the eviction and
  re-cache that one snapshot — a narrow race, not a recurring replay.
- **Connection indicator** — the `/alerts` and `/script-log` live/disconnected pill is driven by
  **aggregate** stream health across all dialers: connected when **at least one** node stream is
  up, disconnected when all are down. This matches today's DPS `SubscribeAck`/`PostStop` pill
  semantics — the pill has never meant "every node is live," only "the panel has a source." The
  two store-backed panels (`driver-health`, `driver-resilience-status`) carry no fleet-wide
  connection flag today, on gRPC or on DPS; per-row staleness is unchanged.

## Configuration reference

See [`docs/Configuration.md` § `Telemetry` / `TelemetryDial`](Configuration.md#telemetry--telemetrydial-live-telemetry-transport)
for the full keys table. In brief:

- **`Telemetry`** (node/serve side): `Mode` (`Dps`|`Grpc`, default `Dps`), `GrpcListenPort` (`0` =
  disabled — the driver node's dedicated telemetry h2c port), `ApiKey` (the shared node key; supply
  via `${secret:}`/env, never commit).
- **`TelemetryDial`** (central side): `Mode` (`Dps`|`Grpc`), `ApiKey` (must equal the nodes'
  `Telemetry:ApiKey`), `ContactRefreshSeconds` (default `60`), `CallTimeoutSeconds` (default `30`).

A driver-role node with `Telemetry:Mode=Grpc` must set `GrpcListenPort > 0` (nothing to serve
otherwise), and `Grpc` mode on either side requires a non-empty `ApiKey` — both are enforced by
`ValidateOnStart` validators, fail-fast at boot.

## Relationship to the rest of the mesh program

- **Same dark-switch discipline as Phases 2 and 3.** `MeshTransport:Mode`, `ConfigSource:Mode`, and
  `Telemetry:Mode`/`TelemetryDial:Mode` are all independently-flippable, restart-only config
  changes, and in every case the "new" side is always wired so the flip needs no coordinated
  redeploy.
- **Why the inversion matters for Phase 6.** Every other cross-boundary transport in this program
  (`MeshTransport`, `ConfigSource`/`ConfigServe`) already has central as the addressable, discoverable
  side and the node as the caller or the callee-by-address. Telemetry is the one channel that is
  naturally node-sourced and fleet-wide-fanned, so putting central in the client role — dialing out
  to each node by its `ClusterNode`-recorded address — is what lets telemetry survive the mesh split
  cleanly: central never needs cluster membership with a site's mesh to keep watching it.
- **Status:** merged to `master` (`35552a96`); the live gate (flip the docker-dev rig to `Grpc`,
  confirm all panels stay green with the DPS telemetry bridges unspawned, kill-and-reconnect a site
  node) PASSED 2026-07-23. See `docs/plans/2026-07-22-per-cluster-mesh-program.md` § Phase 5 and
  `docs/plans/2026-07-23-mesh-phase5-grpc-telemetry-stream.md` for the implementation plan.