CentralCommunicationActor.HandleHeartbeat was forwarding each incoming
HeartbeatMessage to Context.Parent, which resolves to the /user
guardian — a non-actor. Every site heartbeat went straight to dead
letters (~1026 per central node per 30 minutes at the default ~2s
interval across three sites).
The aggregator now exposes MarkHeartbeat(siteId, receivedAt) which
bumps LastReportReceivedAt on already-known sites (and clears IsOnline
if it had flipped) without touching LatestReport. Heartbeats from
unregistered sites are dropped — first registration still happens on
the first full report. CentralCommunicationActor calls this in place
of the no-op Tell.
The result: heartbeats now serve their stated health-monitoring
purpose (per CLAUDE.md) by keeping a site marked online between the
30s full reports if a single report is briefly delayed, and the dead
letter noise disappears entirely.
CentralHealthAggregator is a per-node hosted singleton, but site health
reports flow through ClusterClient which round-robins each report to one
central node only. The other node's aggregator never saw those reports
and marked sites offline at the 60s threshold — sites constantly flapped
between online and offline on the monitoring page.
On receive, the active CentralCommunicationActor now republishes a
SiteHealthReportReplica wrapper on a DistributedPubSub topic. Both
central nodes subscribe to the topic and process replicas through a
dedicated path that updates the local aggregator without re-broadcasting
(avoids fan-out loops). The aggregator's existing sequence-number
idempotency makes self-delivery a cheap no-op.
DistributedPubSubExtensionProvider is now listed in the HOCON
`akka.extensions` block so the mediator is initialised at cluster
start, eliminating a race where the first Subscribe arrived before the
extension was loaded.
ClusterClient Sender refs are temporary proxies — valid for immediate reply
but not durable for future Tells. Events now flow as DebugStreamEvent through
SiteCommunicationActor → ClusterClient → CentralCommunicationActor → bridge
actor (same pattern as health reports). Also fix DebugStreamHub to use
IHubContext for long-lived callbacks instead of transient hub instance.
Central and site clusters now communicate via ClusterClient/
ClusterClientReceptionist instead of direct ActorSelection. Both
CentralCommunicationActor and SiteCommunicationActor are registered
with their cluster's receptionist. Central creates one ClusterClient
per site using NodeA/NodeB contact points from the DB. Sites configure
multiple CentralContactPoints for automatic failover between central
nodes. ISiteClientFactory enables test injection.
Sites now send SiteHealthReport via AkkaHealthReportTransport →
SiteCommunicationActor → CentralCommunicationActor → CentralHealthAggregator.
Added IHealthReportTransport impl, ISiteIdentityProvider impl, registered
HealthReportSender on site nodes, and added SiteHealthReport handler in
CentralCommunicationActor. Health Dashboard now shows all 3 sites online.
Central now resolves site Akka remoting addresses from the Sites DB table
(NodeAAddress/NodeBAddress) instead of relying on runtime RegisterSite
messages. Eliminates the race condition where sites starting before central
had their registration dead-lettered. Addresses are cached in
CentralCommunicationActor with 60s periodic refresh and on-demand refresh
when sites are added/edited/deleted via UI or CLI.
Joseph Doherty