ScadaBridge/docs/plans/2026-08-07-secrets-grpc-hub-live-gate.md

Secrets gRPC hub — live gate on the docker cluster

Result: 3 of 4 checks PASS. Check 4 FAILS one clause of its stated criterion. NOT MERGED — the gate's merge condition is 4/4. Run 2026-08-07 on the local 8-node docker cluster.

Purpose

Terminal gate (Task 7) for feat/secrets-grpc-hub, which wires ZB.MOM.WW.Secrets.Replicator.Grpc 0.4.0 into ScadaBridge: the central role hosts a pull-only secrets_hub.v1 gRPC hub on its existing h2c control-plane listener, and a site node runs a sweep that converges its own local SQLite store from that hub.

This is the ScadaBridge half of the production-topology decision in scadaproj#3 (design: scadaproj/docs/plans/2026-08-07-secrets-production-topology-and-grpc-hub.md). The SqlServer replicator is not the production path here: it requires every site node to hold a connection string to central's database, which breaks ScadaBridge's standing rule that sites talk to central, not to central's DB.

Scope: rig config only. No product code was changed by this gate — see Defects.

Rig state

Fact	Value
Repo / branch	ScadaBridge @ feat/secrets-grpc-hub
Git SHA gated	fc784b413713b1643a21d573e55c00bcc05bd8ee (fc784b41)
Image	scadabridge:latest = 2958aa2c2bbb, built 2026-08-07 11:39 UTC
ZB.MOM.WW.ScadaBridge.Host.dll	09cf76dc550d32fe… — byte-identical on all 8 nodes
Secrets libs	ZB.MOM.WW.Secrets* 0.4.0 (Directory.Packages.props)
Working tree at build	only docker/docker-compose.yml modified (rig config, not in the image) + untracked .claude/
Hub enabled on	central-a, central-b (host) and site-a-a, site-a-b (followers)
Hub NOT enabled on	site-b-a, site-b-b, site-c-a, site-c-b — no Secrets__* env, no store file, no hub/sweep log line (verified)
Sweep interval	00:00:30 — the product default, deliberately not shortened, so the timings below are real
KEK	one shared committed dev-only key, identical on all four (kek_id sha256 7451bcbc1f1f on every row observed)

site-b and site-c were left off on purpose: the default-OFF posture is proven side by side on one rig, the same way site-a is the rig's only LocalDb-replicated pair.

Rig-config change made by this gate

docker/docker-compose.yml gained two YAML anchors and applied them to four services:

• x-secrets-hub-env (central pair + site-a pair) — dev KEK, Secrets__SqlitePath, Secrets__Replication__Enabled=true, Secrets__Replication__Mode=Grpc, Secrets__GrpcHub__BearerToken.
• x-secrets-hub-site-env (site-a pair only) — Secrets__GrpcHub__Endpoint=http://scadabridge-central-a:8083.

All values are DEV-ONLY and committed under the same exception the mesh PSKs and InboundApi:ApiKeyPepper already use. Secrets__SqlitePath points at /app/data because the appsettings default is a relative path resolving to /app — inside the image's writable layer, destroyed by every container recreate. The central pair gained a ./central-node-*/data:/app/data volume for this; the site pairs already had one.

Method

Secrets were seeded, read and deleted with the ZB.MOM.WW.Secrets.Cli (secret) published from scadaproj/ZB.MOM.WW.Secrets (0.4.0) and run in a throwaway container mounting one node's data volume, with that node's KEK. This is the method the sibling OtOpcUa gate used the same day (OtOpcUa/docs/plans/2026-08-07-secrets-pair-local-live-gate.md).

The CentralUI /admin/secrets page was not used, and for this topology that costs nothing: the hub is a pure reader of central's store on every call, with no write-triggered path and no manifest cache. A row written by any process into central's store is on exactly the same path a row written through the page would be. (Contrast the Akka replicator, where the two differ — hence that gate's residual 1, which does not apply here.)

Ciphertext identity is evidenced as a SHA-256 over the encrypted columns only (ciphertext‖nonce‖tag‖wrapped_dek‖wrap_nonce‖wrap_tag), plus a hash of kek_id. No plaintext, no token and no key material appears in this document. The secret values used were throwaway smoke strings.

gRPC calls for check 4 were made with fullstorydev/grpcurl on the scadabridge-net network, against the committed secrets_hub.proto.

Results

#	Check	Result	Timing
1	Convergence — central write reaches both site-a nodes, byte-identical, decrypt-verified	PASS	5 s (budget 30 s + margin)
2	Cold-boot offline — site pair boots and serves last-known-good with central down; convergence resumes	PASS	resumed in 14 s / 43 s
3	Tombstone — delete propagates, survives a pair restart	PASS	≤ 9 s
4	Auth negatives + log hygiene	FAIL (1 clause of 3)	—

Baseline before seeding: all four stores empty.

Check 1 — convergence · PASS

Seeded hub-gate-smoke-1 on central-a only at 11:46:57Z.

central-node-a   hub-gate-smoke-1  rev=0 del=0 cipher_sha256=00a14a2eb8de987fb75a4515 kek=7451bcbc1f1f updated=2026-08-07T11:46:57.1403922+00:00
site-a-node-a    hub-gate-smoke-1  rev=0 del=0 cipher_sha256=00a14a2eb8de987fb75a4515 kek=7451bcbc1f1f updated=2026-08-07T11:46:57.1403922+00:00
site-a-node-b    hub-gate-smoke-1  rev=0 del=0 cipher_sha256=00a14a2eb8de987fb75a4515 kek=7451bcbc1f1f updated=2026-08-07T11:46:57.1403922+00:00

Ciphertext hash, kek_id, revision and updated_utc are identical on all three — the row is relayed verbatim, not re-encrypted, so the last-writer-wins ordering key survives the hop. First poll after the seed (11:47:05Z) already showed both followers converged.

Both halves logged it, and central served the fetch:

site-a-a  [11:47:02 INF] Secret hub sync converged: pulled 1 row(s) from the central hub.
site-a-b  [11:47:02 INF] Secret hub sync converged: pulled 1 row(s) from the central hub.
central-a [11:47:02 INF] Request finished HTTP/2 POST .../SecretsHub/GetSecrets - 200 application/grpc 0.9899ms

Convergence = 11:46:57Z → 11:47:02Z = 5 s, one sweep tick.

Decrypt-verified on both followers: secret get hub-gate-smoke-1 returned the exact plaintext seeded on central-a, on site-a-a and site-a-b.

Check 2 — cold-boot offline · PASS

Central pair stopped 11:48:27Z → 11:48:37Z.

Sweep degrades, does not crash. One warning per interval on each follower, carrying the transport fault and nothing else:

site-a-a [11:49:02 WRN] Secret hub sync failed; the node continues serving its local store and will retry on the next interval.
         Grpc.Core.RpcException: Status(StatusCode="Unavailable", Detail="Error connecting to subchannel." …)

Site pair restarted together (the supported pattern) 11:49:31Z → 11:50:12Z, with central still down. site-a-a booted at 11:49:42Z; its SyncOnStartup sweep warned at 11:49:42Z and retried at 11:50:12 / 11:50:42 / 11:51:12 / 11:51:42 / 11:52:12 / 11:52:42. Both containers running, RestartCount=0, zero FATAL / Unhandled exception / Application startup exception lines.

Last-known-good served with central down — at 11:51:08Z, secret get hub-gate-smoke-1 returned the correct plaintext on both site-a-a and site-a-b.

Central pair started 11:51:37Z. hub-gate-smoke-2 seeded on central-a at 11:52:29Z — after the site nodes were already up, so this proves convergence resumed rather than re-bootstrapped:

site-a-b [11:52:43 INF] Secret hub sync converged: pulled 1 row(s) from the central hub.   →  14 s
site-a-a [11:53:12 INF] Secret hub sync converged: pulled 1 row(s) from the central hub.   →  43 s

Both rows byte-identical to central's (cipher_sha256=4f022e222cfc2f8f77dcfb0b, updated=2026-08-07T11:52:29.6983529+00:00).

Honest note on the 43 s. site-a-a's first post-outage sweep (11:52:42Z) still failed, with SocketException: Name or service not known — a stale DNS/subchannel view of a container that had been docker stopped and docker started. It recovered unaided on the very next tick. So after a central outage, worst-case convergence on this rig is two sweep intervals, not one. Bounded, self-healing, no intervention — but recorded so the 5 s figure from check 1 is not read as the post-outage number.

Check 3 — tombstone · PASS

hub-gate-smoke-1 deleted on central-a at 11:54:05Z. hub-gate-smoke-2 left live as a control.

central-node-a  hub-gate-smoke-1  rev=1 del=1 updated=2026-08-07T11:54:05.9812744+00:00 deleted=2026-08-07T11:54:05.9812744+00:00
site-a-node-a   hub-gate-smoke-1  rev=1 del=1 updated=2026-08-07T11:54:05.9812744+00:00 deleted=2026-08-07T11:54:05.9812744+00:00
site-a-node-b   hub-gate-smoke-1  rev=1 del=1 updated=2026-08-07T11:54:05.9812744+00:00 deleted=2026-08-07T11:54:05.9812744+00:00

Both followers held the tombstone at the 11:54:14Z poll — ≤ 9 s. Delete is a soft tombstone (is_deleted=1, revision bumped, updated_utc restamped), which is what lets it win under last-writer-wins. secret get returned {"error":"not-found"} and secret list returned only hub-gate-smoke-2 on both followers.

No resurrection across a pair restart. Site pair restarted together 11:55:09Z → 11:55:50Z, with central up and sweeping the whole time — so the sweep had every opportunity to re-pull the row. Re-checked at 11:57:05Z (75 s / 2+ sweeps later):

• hub-gate-smoke-1 still rev=1 del=1 on both, updated_utc unchanged; secret get → {"error":"not-found"} on both.
• hub-gate-smoke-2 survived on both and decrypted to the correct plaintext on both.

Check 4 — auth negatives + log hygiene · FAIL (one clause)

Three calls to SecretsHub/GetManifest on scadabridge-central-a:8083 at 11:57:31–32Z:

Call	Result
no authorization header	Code: Unauthenticated / Message: Unauthenticated.
wrong bearer (Bearer not-the-hub-token)	Code: Unauthenticated / Message: Unauthenticated.
correct bearer (positive control)	200, manifest returned (2 entries)

PASS — identical denial. Both negatives are byte-identical in code and detail: a caller learns it was refused and nothing else. The positive control proves the endpoint is live and the refusals are the auth gate, not a dead route.

PASS — log hygiene, zero hits. Every one of the 8 nodes' docker logs and every on-disk Serilog file under docker/*/logs/ was grepped for the dev bearer token value, the dev KEK value, and both secret plaintexts:

central-a  token=0 kek=0 plain1=0 plain2=0 authz-header-mentions=0
central-b  token=0 kek=0 plain1=0 plain2=0 authz-header-mentions=0
site-a-a   token=0 kek=0 plain1=0 plain2=0 authz-header-mentions=0
site-a-b   token=0 kek=0 plain1=0 plain2=0 authz-header-mentions=0
site-b-a / site-b-b / site-c-a / site-c-b   all zero
on-disk log files: zero matching files for all four search strings

The token appears only where it legitimately may: the committed rig config (docker/docker-compose.yml) and the resulting container environment.

FAIL — no server-side warning. The criterion asks for a server-side warning on denial. What the hub actually emits is one Information-level line per denial, from Grpc.AspNetCore.Server, not from the hub's own gate:

[11:57:31 INF] Error status code 'Unauthenticated' with detail 'Unauthenticated.' raised.
[11:57:32 INF] Error status code 'Unauthenticated' with detail 'Unauthenticated.' raised.

Nothing at WRN or above was emitted by either denial — verified by dumping every log line in the 11:57:31–32Z window, not by a filtered grep.

This is not a wiring mistake and not fixable in this repo: SecretsHubAuthInterceptor (in ZB.MOM.WW.Secrets.Replicator.Grpc 0.4.0) deliberately logs nothing on a denial. Its only warning is WarnUnconfiguredOnce, emitted once when the hub has no token configured at all — which is not the case here. So the shipped design and this gate's criterion disagree, and the disagreement lives in the library.

The consequence is real but narrow: a follower whose token was mis-rotated is refused on every sweep and stops converging, while central records that only at Information. An operator watching for warnings sees the follower's own Secret hub sync failed WRN (which does fire), but central shows nothing above INF.

Not patched in ScadaBridge. A host-side interceptor could log the denial, but that would (a) contradict a decision the library documents and reasons about explicitly, in the consumer rather than at the layer that owns it, and (b) put an unbounded log write on an unauthenticated endpoint. Both are worse than the gap. The correct fix is a rate-limited warning in SecretsHubAuthInterceptor, or an explicit decision that Information is right and this criterion should be relaxed. That call is not this gate's to make — see Follow-ups.

Defects found

None in product code. One defect in the gate method, caught and corrected before any result was recorded:

Host-side seeding of a bind-mounted SQLite store is not coherent with the running container (Docker Desktop for macOS). The first attempt at check 1 seeded central-a's store with the CLI running natively on the host. The row was visible to the host and to a fresh helper container, but the running node never saw it: four consecutive sweeps returned an empty manifest, no GetSecrets was ever issued, and a wal_checkpoint(TRUNCATE) did not help. Restarting central-a then lost the row entirely — it was gone from both the host and container views. Cause is the virtiofs/page-cache boundary between the macOS host and the Linux guest: SQLite's WAL and -shm mappings are not shared across it.

Corrected by doing every store access — seed, read, delete — from a throwaway container on the same guest kernel, which is what the sibling OtOpcUa gate already did. The rig was returned to an empty-store baseline before check 1 was re-run. Recorded because the failure mode is a convincing false negative: it looks exactly like "the hub is broken".

Residuals

1. The hub client dials a SINGLE endpoint — no failover. Known and documented on the branch. Observed live and directly contrasted on the same node in the same minute:

   site-a-a [11:51:13 WRN] Central control-plane endpoint http://scadabridge-central-a:8083 is
            unavailable; site site-a failed over to http://scadabridge-central-b:8083.
   site-a-a [11:51:12 WRN] Secret hub sync failed; the node continues serving its local store …
   

   ScadaBridge:Communication:CentralGrpcEndpoints is a list and failed over; the hub client stalled. Survivable — the sweep is best-effort and the node keeps its full local last-known-good store — but secrets stop converging until that one central node returns.

2. The central pair does NOT converge with itself, and this makes residual 1 sharper than it reads. Both central nodes host the hub (verified: central-b answers, and refuses an unauthenticated call identically). But central-b's store stayed empty for the whole run while central-a held both secrets:

   central-b, correct bearer, 11:59:32Z:  {}          ← empty manifest, authenticated
   central-node-b (store):                (empty)
   

   Nothing replicates central-a ↔ central-b in Mode=Grpc — the topology is pull-only central → sites. So a hypothetical failover to central-b would not merely stall, it would succeed against an empty hub. By the reconciler's pull-only algorithm that is not data loss (a name absent from the manifest is simply not pulled, never deleted), but it is a silent stop to convergence, and a deployment that authors a secret on the central node the sites do not dial will see it reach nobody. Not live-tested here — asserted from the algorithm plus the observed empty manifest. Any production enablement needs an answer for how the central pair itself converges.

3. Both site-a nodes dial the same central node. Not a rig shortcut — it is the only shape the client supports (residual 1). Worth restating so the run is not read as having exercised two independent hub sources.

4. The /admin/secrets write path was not exercised. Seeding was CLI-into-central's-store. For this topology that is the same path (see Method), so it is a coverage note, not a gap in the result — but the page's own Blazor/authz behaviour is unproven by this gate. ScadaBridge#22 covered it separately.

5. Delete-while-a-follower-is-offline was not tested. Check 3 restarts the pair after both already held the tombstone. The harder resurrection case — deleting at central while a follower is down, then bringing it back holding a live copy — is covered by the library's last-writer-wins tests but not live here. Under a pull-only hub it is structurally safer than in a bidirectional topology (the follower can never push its stale live row back), which is why it was not prioritised.

6. Replication stays default-OFF in the product. Secrets:Replication:Enabled is false and Mode is SqlServer in the shipped appsettings.json. It is enabled on this docker rig only, on four of eight nodes, with a dev KEK and a dev token. Production enablement additionally needs a real KEK supplied out of band and a real token — never committed, and never a ${secret:} reference.

Follow-ups

• File against ZB.MOM.WW.Secrets.Replicator.Grpc: decide whether a denied hub call should emit a rate-limited Warning on the server. Today it emits nothing from the gate itself, so a mis-rotated follower token is visible at central only at Information level. This is the sole reason this gate is not 4/4.
• File against ScadaBridge: endpoint failover for the hub client, and/or a documented answer for how the central pair's own stores converge (residual 2). Both are the same underlying question — "which central node is authoritative for secrets".

Conclusion

The pull-only gRPC secrets hub works, and works the way the branch says it does: convergence is one sweep tick, ciphertext is relayed verbatim so last-writer-wins ordering survives the hop, a site node boots and serves its full last-known-good store with the entire central pair stopped, convergence resumes unaided when central returns, tombstones propagate and do not resurrect across a pair restart, the auth gate refuses uniformly, and no token, KEK or plaintext reaches any log on any of the eight nodes.

One clause of check 4 is not met — denials are recorded at Information, not Warning — and that clause is a property of the library, not of this branch. The gate's merge condition is 4/4, so this branch is not merged by this gate. Resolving the logging question (fix the library, or relax the criterion) is the only thing standing between this result and a merge.