Fix followers not receiving channel messages

This commit fixes a bug where the leader was not propagating to
the followers.

See #468

internal/server/follow.go

@@ -169,15 +169,19 @@ func (s *Server) followHandleCommand(args []string, followc int, w io.Writer) (i
 		return s.aofsz, errNoLongerFollowing
 	}
 	msg := &Message{Args: args}
-
 	_, d, err := s.command(msg, nil)
 	if err != nil {
 		if commandErrIsFatal(err) {
 			return s.aofsz, err
 		}
 	}
-	if err := s.writeAOF(args, &d); err != nil {
-		return s.aofsz, err
+	switch msg.Command() {
+	case "publish":
+		// Avoid writing these commands to the AOF
+	default:
+		if err := s.writeAOF(args, &d); err != nil {
+			return s.aofsz, err
+		}
 	}
 	if len(s.aofbuf) > 10240 {
 		s.flushAOF(false)

internal/server/pubqueue.go

@@ -0,0 +1,81 @@
+package server
+
+import (
+	"net"
+	"sync"
+
+	"github.com/tidwall/redcon"
+)
+
+type pubQueue struct {
+	cond    *sync.Cond
+	entries []pubQueueEntry // follower publish queue
+	closed  bool
+}
+
+type pubQueueEntry struct {
+	channel  string
+	messages []string
+}
+
+func (s *Server) startPublishQueue(wg *sync.WaitGroup) {
+	defer wg.Done()
+	var buf []byte
+	var conns []net.Conn
+	s.pubq.cond = sync.NewCond(&sync.Mutex{})
+	s.pubq.cond.L.Lock()
+	for {
+		for len(s.pubq.entries) > 0 {
+			entries := s.pubq.entries
+			s.pubq.entries = nil
+			s.pubq.cond.L.Unlock()
+			// Get follower connections
+			s.mu.RLock()
+			for conn := range s.aofconnM {
+				conns = append(conns, conn)
+			}
+			s.mu.RUnlock()
+			// Buffer the PUBLISH command pipeline
+			buf = buf[:0]
+			for _, entry := range entries {
+				for _, message := range entry.messages {
+					buf = redcon.AppendArray(buf, 3)
+					buf = redcon.AppendBulkString(buf, "PUBLISH")
+					buf = redcon.AppendBulkString(buf, entry.channel)
+					buf = redcon.AppendBulkString(buf, message)
+				}
+			}
+			// Publish to followers
+			for i, conn := range conns {
+				conn.Write(buf)
+				conns[i] = nil
+			}
+			conns = conns[:0]
+			s.pubq.cond.L.Lock()
+		}
+		if s.pubq.closed {
+			break
+		}
+		s.pubq.cond.Wait()
+	}
+	s.pubq.cond.L.Unlock()
+}
+
+func (s *Server) stopPublishQueue() {
+	s.pubq.cond.L.Lock()
+	s.pubq.closed = true
+	s.pubq.cond.Broadcast()
+	s.pubq.cond.L.Unlock()
+}
+
+func (s *Server) sendPublishQueue(channel string, message ...string) {
+	s.pubq.cond.L.Lock()
+	if !s.pubq.closed {
+		s.pubq.entries = append(s.pubq.entries, pubQueueEntry{
+			channel:  channel,
+			messages: message,
+		})
+	}
+	s.pubq.cond.Broadcast()
+	s.pubq.cond.L.Unlock()
+}

internal/server/pubsub.go

@@ -66,6 +66,7 @@ func (s *Server) Publish(channel string, message ...string) int {
 	}
 	s.pubsub.mu.RUnlock()
 
+	// broadcast to clients
 	for _, msg := range msgs {
 		msg.target.cond.L.Lock()
 		msg.target.msgs = append(msg.target.msgs, msg)
@@ -73,6 +74,9 @@ func (s *Server) Publish(channel string, message ...string) int {
 		msg.target.cond.L.Unlock()
 	}
 
+	// Broadcast to followers
+	s.sendPublishQueue(channel, message...)
+
 	return len(msgs)
 }
 

internal/server/server.go

@@ -140,6 +140,7 @@ type Server struct {
 	fcup     bool       // follow caught up
 	fcuponce bool       // follow caught up once
 	aofconnM map[net.Conn]io.Closer
+	pubq     pubQueue
 
 	// lua scripts
 	luascripts *lScriptMap
@@ -422,9 +423,12 @@ func Serve(opts Options) error {
 	go s.backgroundExpiring(&bgwg)
 	bgwg.Add(1)
 	go s.backgroundSyncAOF(&bgwg)
+	bgwg.Add(1)
+	go s.startPublishQueue(&bgwg)
 	defer func() {
 		log.Debug("Stopping background routines")
 		// Stop background routines
+		s.stopPublishQueue()
 		s.followc.Add(1) // this will force any follow communication to die
 		s.stopServer.Store(true)
 		if mln != nil {