feat(cluster): auto-down downing strategy — either-node crash now fails over (owner decision 2026-07-21: availability over partition-safety)
Two-node keep-oldest could NEVER survive a crash of the oldest/active node:
Akka.NET 1.5.62 KeepOldest.OldestDecision only lets down-if-alone rescue a
side with >= 2 members, so the 1-vs-1 survivor takes DownReachable and downs
ITSELF — proven live on the rig ('SBR took decision ... including myself')
before this change. static-quorum(1) is worse (IsTooManyMembers -> DownAll);
keep-majority just re-keys the fatal crash to the lowest address.
SplitBrainResolverStrategy gains 'auto-down' (new default): BuildHocon emits
Akka's AutoDowning provider with auto-down-unreachable-after = StableAfter.
The leader among the REACHABLE members downs the unreachable peer, so the
survivor takes over singletons and /health/active in ~25s regardless of which
node died. Accepted trade (explicit owner decision): a real network partition
runs dual-active until an operator restarts one side. keep-oldest remains
supported; DownIfAlone validation is now scoped to it.
Live drill on the rebuilt rig: active-crash TAKEOVER in 28s (victim still
down; all 7 singletons Younger->Oldest), standby-crash removal 27s with 0
routing blips; victims rejoin as standby in 2s. New real-cluster tests pin
both directions (SbrFailoverTests.AutoDown_*); TwoNodeClusterFixture gains a
strategy knob. All 16 appsettings flipped (src, docker, docker-env2, and the
gitignored wonder-app-vd03 overlay on disk — owner must sync to the host).
Docs: decision record docs/plans/2026-07-21-auto-down-availability-decision.md,
Component-ClusterInfrastructure downing section rewritten, drill + README
reworked (active mode now asserts takeover), deferred-work SBR row resolved.
This commit is contained in:
Joseph Doherty
+14
-12
@@ -273,29 +273,31 @@ All test passwords are `password`. See `infra/glauth/config.toml` for the full l
|
||||
### Automated failover drill (`failover-drill.sh`)
|
||||
|
||||
```bash
|
||||
DRILL_MODE=standby bash docker/failover-drill.sh # default — survivable younger-node crash
|
||||
DRILL_MODE=active bash docker/failover-drill.sh # oldest-node crash — measures the registered outage gap
|
||||
DRILL_MODE=standby bash docker/failover-drill.sh # default — younger-node crash, active untouched
|
||||
DRILL_MODE=active bash docker/failover-drill.sh # oldest-node crash — survivor must TAKE OVER
|
||||
```
|
||||
|
||||
The scripted drill (`docker kill` = SIGKILL, the hard-crash path — a `docker stop` would take the graceful `CoordinatedShutdown` path and would not prove crash recovery) has **two modes**, because under the unified oldest-member semantics the *active* node IS the oldest, i.e. the one crash two-node keep-oldest cannot survive:
|
||||
The scripted drill (`docker kill` = SIGKILL, the hard-crash path — a `docker stop` would take the graceful `CoordinatedShutdown` path and would not prove crash recovery) has **two modes**, and since the **auto-down decision (2026-07-21)** both expect recovery — the cluster runs Akka's `AutoDowning` provider (`auto-down-unreachable-after` = 15s), under which the leader among the *reachable* members downs the unreachable peer, so a crash of either node fails over:
|
||||
|
||||
- **`DRILL_MODE=standby` (default) — kills the STANDBY (younger) central node.** The survivable direction: SBR downs the crashed member and the active node keeps its singletons. Expected result: **no routing outage at all** (the active node is never touched, so `/health/active` blips = 0) and member removal on the survivor within **~25s** (10s failure-detection threshold + 15s stable-after; the 2s heartbeat interval is not additive). PASS = the survivor logs the member removal within `TIMEOUT_S` (default 90s) while routing stays up.
|
||||
- **`DRILL_MODE=active` — kills the ACTIVE (oldest) central node.** Expected result: a **total central outage** until the victim container is restarted — this is the registered deferred keep-oldest decision (master tracker 2026-07-08): keep-oldest downs the partition *without* the oldest, so the younger survivor downs itself, and it cannot re-form a cluster alone (see the seed-node constraint below). The drill confirms the dark window, then recovery within ~2 min of restarting the victim. The mode exists to make the registered gap *observable*, not to pretend it is covered.
|
||||
- **`DRILL_MODE=standby` (default) — kills the STANDBY (younger) central node.** The active node is untouched: expected result is **no routing outage at all** (`/health/active` blips = 0) and member removal on the survivor within **~25s** (10s failure-detection threshold + 15s auto-down window; the 2s heartbeat interval is not additive). PASS = the survivor logs the downing/removal within `TIMEOUT_S` (default 90s) while routing stays up.
|
||||
- **`DRILL_MODE=active` — kills the ACTIVE (oldest) central node.** The survivor must **take over while the victim is still down**: it auto-downs the dead oldest, becomes the oldest member itself, re-hosts all singletons, and its `/health/active` goes 200. PASS = survivor active within `TIMEOUT_S`, then Traefik routing to it. (Under the pre-2026-07-21 `keep-oldest` strategy this direction was a proven total outage — the younger survivor took `DownReachable` and downed itself, because Akka's `down-if-alone` only rescues a side with ≥ 2 members.)
|
||||
|
||||
The drill exercises S1 (SBR downing on hard crash), S3 (single active node routed through Traefik), and the Task 20 restart/rejoin contract. Requires a running cluster (`bash docker/deploy.sh`) and `curl` + `docker` on the host.
|
||||
Both modes finish by restarting the victim and confirming it rejoins as a ready standby. The drill exercises downing-on-hard-crash, S3 (single active node routed through Traefik), and the Task 20 restart/rejoin contract. Requires a running cluster (`bash docker/deploy.sh`) and `curl` + `docker` on the host.
|
||||
|
||||
**Seed-node bootstrap constraint.** Only the FIRST seed in `Cluster:SeedNodes` may self-join to form a *new* cluster. Both central nodes list `scadabridge-central-a` first (`docker/central-node-a/appsettings.Central.json`, `docker/central-node-b/appsettings.Central.json`), so a lone restarted `central-b` (with `central-a` still down) loops on `InitJoin` forever — it never reaches `Up`, and `/health/active` never returns 200. Operator recovery actions: **(1)** restart the dead first-seed node (`central-a`) — preferred; or **(2)** restart the survivor with a self-first seed override (env `ScadaBridge__Cluster__SeedNodes__0=akka.tcp://scadabridge@<self-host>:8081`, `ScadaBridge__Cluster__SeedNodes__1=<peer>`). The repo deliberately does NOT ship self-first ordering per node: with *both* nodes self-first, a simultaneous cold start can let each self-join independently → two one-node clusters that never merge (the cold-start split-brain the identical-seed-order convention exists to prevent). The real remedy is the pending keep-oldest topology/strategy decision (deferred, owner: user).
|
||||
**Partition trade (accepted).** Auto-down is availability-first: in a *real network partition* (both nodes alive, link cut) each side downs the other and both run active — dual-active until an operator restarts one side after the partition heals. This was an explicit owner decision (2026-07-21): site pairs have no shared lease infrastructure to arbitrate, and a stalled system is a bigger risk than a rare partition. See `docs/plans/2026-07-21-auto-down-availability-decision.md`.
|
||||
|
||||
> **Observed results** (plan R2-01 T3):
|
||||
**Seed-node bootstrap constraint (still applies to boot-alone).** Only the FIRST seed in `Cluster:SeedNodes` may self-join to form a *new* cluster. Both central nodes list `scadabridge-central-a` first (`docker/central-node-a/appsettings.Central.json`, `docker/central-node-b/appsettings.Central.json`), so a lone *restarted* `central-b` (with `central-a` still down) loops on `InitJoin` forever. Under auto-down this no longer causes the active-crash outage (the survivor keeps running — it never restarts), but it still bites when a node must boot alone (cold start of only the non-first-seed VM, or the survivor crashing while its peer is still dead). Operator recovery: **(1)** restart the first-seed node (`central-a`) — preferred; or **(2)** restart the survivor with a self-first seed override (env `ScadaBridge__Cluster__SeedNodes__0=akka.tcp://scadabridge@<self-host>:8081`, `ScadaBridge__Cluster__SeedNodes__1=<peer>`). The repo deliberately does NOT ship self-first ordering per node: with *both* nodes self-first, a simultaneous cold start can let each self-join independently → two one-node clusters that never merge.
|
||||
|
||||
> **Observed results** (auto-down decision verification):
|
||||
>
|
||||
> **Run 2026-07-13** against a freshly-deployed cluster on `main` @ `99544985` (round-2 merged image; `active=central-a`). Both directions behaved exactly as the design predicts.
|
||||
> **Run 2026-07-21** against a freshly-deployed cluster with `SplitBrainResolverStrategy: auto-down` (first drill: `active=central-a`). Both directions recovered.
|
||||
>
|
||||
> | Direction (`DRILL_MODE`) | Outcome | Measured |
|
||||
> |--------------------------|---------|----------|
|
||||
> | `standby` (younger-node crash) | **PASS** — SBR downed+removed the crashed `central-b`; active `central-a` kept all 7 singletons; recovered on restart. | Member removed in **27s** (budget ~25s: 10s detection + 15s stable-after); **0** `/health/active` routing blips (active node never touched); routable **0s** after victim restart. |
|
||||
> | `active` (oldest-node crash) | **Outage as designed** — killing the oldest/active `central-a` made the younger `central-b` self-down (total central outage — the registered keep-oldest gap); recovered after restarting the victim, `central-b` then assuming Oldest and re-hosting all singletons. | Outage confirmed at **9s**; central routable again **4s** after restarting `central-a`. |
|
||||
> | `active` (oldest-node crash) | **PASS — TAKEOVER** — `central-b` auto-downed the dead oldest, went `Younger -> Oldest` on all 7 singletons, and served `/health/active` **while the victim was still down**; restarted victim rejoined as standby. | Survivor active + Traefik routing in **28s** (budget ~25s: 10s detection + 15s auto-down + hand-over); victim ready **2s** after restart. |
|
||||
> | `standby` (younger-node crash) | **PASS** — active node untouched; survivor downed+removed the crashed member; restarted victim rejoined as standby. | Member removed in **27s**; **0** `/health/active` routing blips; victim ready **2s** after restart. |
|
||||
>
|
||||
> Notes: the `standby` PASS shows the survivable direction is clean end-to-end (SBR `DownUnreachable` decision + per-singleton "Member removed" in the survivor log, zero routing interruption). The `active` result **empirically confirms the deferred keep-oldest topology gap** (master tracker 2026-07-08 / `docs/plans/2026-07-08-deferred-work-register.md`): a hard crash of the active/oldest central node is a total outage until that node (the first seed) is restarted — the remedy remains the pending topology/strategy decision. In-process envelope (`FailoverTimingTests`, plan R2-01 T4) independently measured full failover at **33.7s**.
|
||||
> Historical baseline (keep-oldest, run 2026-07-13 on `99544985`): `standby` PASS with member removal in 27s / 0 routing blips; `active` was a **total outage** — `central-b` self-downed ~20s after the kill (live SBR log 2026-07-21: `SBR took decision Akka.Cluster.SBR.DownReachable … including myself`) and could not re-bootstrap until `central-a` returned. That result is what motivated the auto-down decision. In-process envelope (`FailoverTimingTests`) measured full failover at **33.7s**.
|
||||
|
||||
### Central Failover
|
||||
|
||||
|
||||
Reference in New Issue
Block a user