perf(group): reduce websocket fanout serialization

server/session.go

@@ -195,6 +195,9 @@ type Subscription struct {
 	supd chan<- *sessionUpdate
 }
 
+// sharedServerMessage is an immutable JSON payload reused by multiple WebSocket sessions.
+type sharedServerMessage []byte
+
 func (s *Session) enqueueOutbound(payload any) bool {
 	if s.isMultiplex() {
 		return s.enqueueMultiplexOutbound(payload, multiplexEnqueueWait)
@@ -234,8 +237,16 @@ func (s *Session) enqueueMultiplexOutbound(payload any, wait time.Duration) bool
 			s.clusterQueueMu.Unlock()
 			return true
 		}
+		if msg, ok := payload.(*ServerComMessage); ok {
+			msg.clusterQueuedAt = time.Now()
+		}
 		select {
 		case s.send <- payload:
+			peer := "unknown"
+			if s.clnode != nil {
+				peer = s.clnode.name
+			}
+			immetrics.ObserveClusterM2PEnqueue(peer, len(s.send))
 			s.clusterQueueMu.Unlock()
 			return true
 		default:
@@ -525,6 +536,17 @@ func (s *Session) queueOutBytes(data []byte) bool {
 	return true
 }
 
+func (s *Session) queueOutSharedBytes(data []byte) bool {
+	if s == nil || atomic.LoadInt32(&s.terminating) > 0 {
+		return true
+	}
+	if !s.enqueueOutbound(sharedServerMessage(data)) {
+		logs.Err.Println("s.queueOutSharedBytes: session's send queue full", s.sid)
+		return false
+	}
+	return true
+}
+
 func (s *Session) maybeScheduleClusterWriteLoop() {
 	if s.multi != nil {
 		s.multi.scheduleClusterWriteLoop()