Adding a recovery mechanism for a split gossip cluster

Signed-off-by: Dani Louca <dani.louca@docker.com>

libnetwork/networkdb/cluster.go

@@ -2,6 +2,7 @@ package networkdb
 
 import (
 	"bytes"
+	"context"
 	"crypto/rand"
 	"encoding/hex"
 	"fmt"
@@ -17,10 +18,12 @@ import (
 )
 
 const (
-	reapPeriod       = 5 * time.Second
-	retryInterval    = 1 * time.Second
-	nodeReapInterval = 24 * time.Hour
-	nodeReapPeriod   = 2 * time.Hour
+	reapPeriod            = 5 * time.Second
+	rejoinClusterDuration = 10 * time.Second
+	rejoinInterval        = 60 * time.Second
+	retryInterval         = 1 * time.Second
+	nodeReapInterval      = 24 * time.Hour
+	nodeReapPeriod        = 2 * time.Hour
 )
 
 type logWriter struct{}
@@ -154,7 +157,7 @@ func (nDB *NetworkDB) clusterInit() error {
 		return fmt.Errorf("failed to create memberlist: %v", err)
 	}
 
-	nDB.stopCh = make(chan struct{})
+	nDB.ctx, nDB.cancelCtx = context.WithCancel(context.Background())
 	nDB.memberlist = mlist
 
 	for _, trigger := range []struct {
@@ -166,16 +169,17 @@ func (nDB *NetworkDB) clusterInit() error {
 		{config.PushPullInterval, nDB.bulkSyncTables},
 		{retryInterval, nDB.reconnectNode},
 		{nodeReapPeriod, nDB.reapDeadNode},
+		{rejoinInterval, nDB.rejoinClusterBootStrap},
 	} {
 		t := time.NewTicker(trigger.interval)
-		go nDB.triggerFunc(trigger.interval, t.C, nDB.stopCh, trigger.fn)
+		go nDB.triggerFunc(trigger.interval, t.C, trigger.fn)
 		nDB.tickers = append(nDB.tickers, t)
 	}
 
 	return nil
 }
 
-func (nDB *NetworkDB) retryJoin(members []string, stop <-chan struct{}) {
+func (nDB *NetworkDB) retryJoin(ctx context.Context, members []string) {
 	t := time.NewTicker(retryInterval)
 	defer t.Stop()
 
@@ -191,7 +195,7 @@ func (nDB *NetworkDB) retryJoin(members []string, stop <-chan struct{}) {
 				continue
 			}
 			return
-		case <-stop:
+		case <-ctx.Done():
 			return
 		}
 	}
@@ -202,8 +206,8 @@ func (nDB *NetworkDB) clusterJoin(members []string) error {
 	mlist := nDB.memberlist
 
 	if _, err := mlist.Join(members); err != nil {
-		// In case of failure, keep retrying join until it succeeds or the cluster is shutdown.
-		go nDB.retryJoin(members, nDB.stopCh)
+		// In case of failure, we no longer need to explicitly call retryJoin.
+		// rejoinClusterBootStrap, which runs every minute, will retryJoin for 10sec
 		return fmt.Errorf("could not join node to memberlist: %v", err)
 	}
 
@@ -225,7 +229,8 @@ func (nDB *NetworkDB) clusterLeave() error {
 		return err
 	}
 
-	close(nDB.stopCh)
+	// cancel the context
+	nDB.cancelCtx()
 
 	for _, t := range nDB.tickers {
 		t.Stop()
@@ -234,19 +239,19 @@ func (nDB *NetworkDB) clusterLeave() error {
 	return mlist.Shutdown()
 }
 
-func (nDB *NetworkDB) triggerFunc(stagger time.Duration, C <-chan time.Time, stop <-chan struct{}, f func()) {
+func (nDB *NetworkDB) triggerFunc(stagger time.Duration, C <-chan time.Time, f func()) {
 	// Use a random stagger to avoid syncronizing
 	randStagger := time.Duration(uint64(rnd.Int63()) % uint64(stagger))
 	select {
 	case <-time.After(randStagger):
-	case <-stop:
+	case <-nDB.ctx.Done():
 		return
 	}
 	for {
 		select {
 		case <-C:
 			f()
-		case <-stop:
+		case <-nDB.ctx.Done():
 			return
 		}
 	}
@@ -270,6 +275,35 @@ func (nDB *NetworkDB) reapDeadNode() {
 	}
 }
 
+// rejoinClusterBootStrap is called periodically to check if all bootStrap nodes are active in the cluster,
+// if not, call the cluster join to merge 2 separate clusters that are formed when all managers
+// stopped/started at the same time
+func (nDB *NetworkDB) rejoinClusterBootStrap() {
+	nDB.RLock()
+	if len(nDB.bootStrapIP) == 0 {
+		nDB.RUnlock()
+		return
+	}
+
+	bootStrapIPs := make([]string, 0, len(nDB.bootStrapIP))
+	for _, bootIP := range nDB.bootStrapIP {
+		for _, node := range nDB.nodes {
+			if node.Addr.Equal(bootIP) {
+				// One of the bootstrap nodes is part of the cluster, return
+				nDB.RUnlock()
+				return
+			}
+		}
+		bootStrapIPs = append(bootStrapIPs, bootIP.String())
+	}
+	nDB.RUnlock()
+	// None of the bootStrap nodes are in the cluster, call memberlist join
+	logrus.Debugf("rejoinClusterBootStrap, calling cluster join with bootStrap %v", bootStrapIPs)
+	ctx, cancel := context.WithTimeout(nDB.ctx, rejoinClusterDuration)
+	defer cancel()
+	nDB.retryJoin(ctx, bootStrapIPs)
+}
+
 func (nDB *NetworkDB) reconnectNode() {
 	nDB.RLock()
 	if len(nDB.failedNodes) == 0 {

libnetwork/networkdb/delegate.go

@@ -38,16 +38,11 @@ func (nDB *NetworkDB) handleNodeEvent(nEvent *NodeEvent) bool {
 	// If we are here means that the event is fresher and the node is known. Update the laport time
 	n.ltime = nEvent.LTime
 
-	// If it is a node leave event for a manager and this is the only manager we
-	// know of we want the reconnect logic to kick in. In a single manager
-	// cluster manager's gossip can't be bootstrapped unless some other node
-	// connects to it.
-	if len(nDB.bootStrapIP) == 1 && nEvent.Type == NodeEventTypeLeave {
-		for _, ip := range nDB.bootStrapIP {
-			if ip.Equal(n.Addr) {
-				return true
-			}
-		}
+	// If the node is not known from memberlist we cannot process save any state of it else if it actually
+	// dies we won't receive any notification and we will remain stuck with it
+	if _, ok := nDB.nodes[nEvent.NodeName]; !ok {
+		logrus.Error("node: %s is unknown to memberlist", nEvent.NodeName)
+		return false
 	}
 
 	switch nEvent.Type {

libnetwork/networkdb/networkdb.go

@@ -3,6 +3,7 @@ package networkdb
 //go:generate protoc -I.:../vendor/github.com/gogo/protobuf --gogo_out=import_path=github.com/docker/libnetwork/networkdb,Mgogoproto/gogo.proto=github.com/gogo/protobuf/gogoproto:. networkdb.proto
 
 import (
+	"context"
 	"fmt"
 	"net"
 	"os"
@@ -77,9 +78,10 @@ type NetworkDB struct {
 	// Broadcast queue for node event gossip.
 	nodeBroadcasts *memberlist.TransmitLimitedQueue
 
-	// A central stop channel to stop all go routines running on
+	// A central context to stop all go routines running on
 	// behalf of the NetworkDB instance.
-	stopCh chan struct{}
+	ctx       context.Context
+	cancelCtx context.CancelFunc
 
 	// A central broadcaster for all local watchers watching table
 	// events.