Refactor

cluster/cluster.go

@@ -61,52 +61,26 @@ const (
 	DefaultBatchLinger = 500 * time.Millisecond
 )
 
-// Config configures the cluster. It is assembled by the server from its own config, which keeps
-// this package free of server types.
-type Config struct {
-	Enabled           bool          // Master switch; when false, New returns the nop cluster
-	NodeID            NodeID        // Stable per-node identifier; required
-	AdvertiseURL      string        // Base URL peers use to reach this node's fan-out endpoint
-	Secret            string        // Shared secret authenticating node-to-node fan-out requests
-	HeartbeatInterval time.Duration // How often the node registry heartbeat is refreshed
-	NodeTTL           time.Duration // Registry rows older than this do not count as live peers
-	BatchLinger       time.Duration // How long messages wait in a peer queue to form a batch; 0 = send immediately
-	StateInterval     time.Duration // How often the full subscription state is pushed to peers
-	MaxMessageBytes   int64         // Upper bound for a single message on the wire (batch limits derive from this)
-}
-
-// DeliverFunc hands a message received from a peer node to this node's local subscribers. The
-// server supplies it, which inverts the dependency: this package never imports the server.
-type DeliverFunc func(m *model.Message)
-
-// TopicsFunc returns the topics that currently have at least one live subscriber, computed
-// fresh on every call: membership is never tracked as a list, so topics "leave" simply by not
-// appearing in the next snapshot. The server supplies it (same inversion as DeliverFunc).
-type TopicsFunc func() []string
-
 // Cluster fans published messages out to peer cluster nodes and receives their fan-out requests.
 // Local delivery to a node's own subscribers still happens inline in the server; the cluster
 // only covers the cross-node hop.
 type Cluster interface {
-	// The internal peer API (/v1/internal/*), served on the dedicated cluster listener. The
-	// nop cluster answers 404.
 	http.Handler
-	// Relay sends a locally published message on to the peer nodes that may have subscribers
+	// ForwardMessage sends a locally published message on to the peer nodes that may have subscribers
 	// for its topic (all of them, when subscription knowledge is missing or stale). It is
 	// fire-and-forget and must not block the caller's request path.
-	Relay(m *model.Message) error
-	// AnnounceTopics tells peers that these topics just gained their first local subscriber,
-	// closing the routing-knowledge window to ~one round trip. Nop in single-node mode.
-	AnnounceTopics(topics []string)
+	ForwardMessage(m *model.Message) error
+	// BroadcastState pushes a subscription-state delta to ALL peers (unlike ForwardMessage,
+	// which routes), closing the routing-knowledge window to ~one round trip. Nop single-node.
+	BroadcastState(state *State)
 	// IsLeader reports whether this node holds the cluster leader lock. Singleton background
 	// jobs (e.g. the Firebase keepaliver) are gated on the leader.
 	IsLeader() bool
-	// TODO(T8): add Healthy() bool reflecting registration health (unhealthy when the last
-	// successful Register is older than NodeTTL, i.e. when peers stop relaying to this node),
-	// so /v1/health can pull a DB-partitioned node out of DNS rotation. Needs a fail-open
-	// answer at the health checker first: during a full DB outage ALL nodes fail to register
-	// while the mesh keeps delivering on stale peer caches, and naively pulling every node
-	// would turn a control-plane blip into a total outage.
+	// Healthy reports whether this node is fit to serve: its registry heartbeat is fresh
+	// enough (within NodeTTL) that peers still forward messages to it. Health checkers must
+	// fail open (never pull ALL nodes): during a full database outage every node reports
+	// unhealthy while the mesh keeps delivering on stale peer caches.
+	Healthy() bool
 	// Close stops the cluster and releases its resources.
 	Close() error
 }