Run a separate in memory snapshot to reduce number of entries stored in raft memory storage

server/etcdserver/api/membership/cluster.go

@@ -899,7 +899,7 @@ func (c *RaftCluster) Store(store v2store.Store) {
 		if m.ClientURLs != nil {
 			mustUpdateMemberAttrInStore(c.lg, store, m)
 		}
-		c.lg.Info(
+		c.lg.Debug(
 			"snapshot storing member",
 			zap.String("id", m.ID.String()),
 			zap.Strings("peer-urls", m.PeerURLs),

server/etcdserver/server.go

@@ -109,6 +109,7 @@
 	readyPercentThreshold = 0.9
 
 	DowngradeEnabledPath = "/downgrade/enabled"
+	memorySnapshotCount  = 100
 )
 
 var (
@@ -291,9 +292,10 @@
 	clusterVersionChanged *notify.Notifier
 
 	*AccessController
-	// forceSnapshot can force snapshot be triggered after apply, independent of the snapshotCount.
+	// forceDiskSnapshot can force snapshot be triggered after apply, independent of the snapshotCount.
 	// Should only be set within apply code path. Used to force snapshot after cluster version downgrade.
-	forceSnapshot     bool
+	// TODO: Replace with flush db in v3.7 assuming v3.6 bootstraps from db file.
+	forceDiskSnapshot bool
 	corruptionChecker CorruptionChecker
 }
 
@@ -741,10 +743,11 @@
 }
 
 type etcdProgress struct {
-	confState raftpb.ConfState
-	snapi     uint64
-	appliedt  uint64
-	appliedi  uint64
+	confState           raftpb.ConfState
+	diskSnapshotIndex   uint64
+	memorySnapshotIndex uint64
+	appliedt            uint64
+	appliedi            uint64
 }
 
 // raftReadyHandler contains a set of EtcdServer operations to be called by raftNode,
@@ -809,10 +812,11 @@
 	s.r.start(rh)
 
 	ep := etcdProgress{
-		confState: sn.Metadata.ConfState,
-		snapi:     sn.Metadata.Index,
-		appliedt:  sn.Metadata.Term,
-		appliedi:  sn.Metadata.Index,
+		confState:           sn.Metadata.ConfState,
+		diskSnapshotIndex:   sn.Metadata.Index,
+		memorySnapshotIndex: sn.Metadata.Index,
+		appliedt:            sn.Metadata.Term,
+		appliedi:            sn.Metadata.Index,
 	}
 
 	defer func() {
@@ -979,7 +983,7 @@
 	// storage, since the raft routine might be slower than toApply routine.
 	<-apply.notifyc
 
-	s.triggerSnapshot(ep)
+	s.snapshotIfNeededAndCompactRaftLog(ep)
 	select {
 	// snapshot requested via send()
 	case m := <-s.r.msgSnapC:
@@ -998,15 +1002,15 @@
 	lg := s.Logger()
 	lg.Info(
 		"applying snapshot",
-		zap.Uint64("current-snapshot-index", ep.snapi),
+		zap.Uint64("current-snapshot-index", ep.diskSnapshotIndex),
 		zap.Uint64("current-applied-index", ep.appliedi),
 		zap.Uint64("incoming-leader-snapshot-index", toApply.snapshot.Metadata.Index),
 		zap.Uint64("incoming-leader-snapshot-term", toApply.snapshot.Metadata.Term),
 	)
 	defer func() {
 		lg.Info(
 			"applied snapshot",
-			zap.Uint64("current-snapshot-index", ep.snapi),
+			zap.Uint64("current-snapshot-index", ep.diskSnapshotIndex),
 			zap.Uint64("current-applied-index", ep.appliedi),
 			zap.Uint64("incoming-leader-snapshot-index", toApply.snapshot.Metadata.Index),
 			zap.Uint64("incoming-leader-snapshot-term", toApply.snapshot.Metadata.Term),
@@ -1017,7 +1021,7 @@
 	if toApply.snapshot.Metadata.Index <= ep.appliedi {
 		lg.Panic(
 			"unexpected leader snapshot from outdated index",
-			zap.Uint64("current-snapshot-index", ep.snapi),
+			zap.Uint64("current-snapshot-index", ep.diskSnapshotIndex),
 			zap.Uint64("current-applied-index", ep.appliedi),
 			zap.Uint64("incoming-leader-snapshot-index", toApply.snapshot.Metadata.Index),
 			zap.Uint64("incoming-leader-snapshot-term", toApply.snapshot.Metadata.Term),
@@ -1132,7 +1136,8 @@
 
 	ep.appliedt = toApply.snapshot.Metadata.Term
 	ep.appliedi = toApply.snapshot.Metadata.Index
-	ep.snapi = ep.appliedi
+	ep.diskSnapshotIndex = ep.appliedi
+	ep.memorySnapshotIndex = ep.appliedi
 	ep.confState = toApply.snapshot.Metadata.ConfState
 
 	// As backends and implementations like alarmsStore changed, we need
@@ -1188,31 +1193,26 @@
 }
 
 func (s *EtcdServer) ForceSnapshot() {
-	s.forceSnapshot = true
+	s.forceDiskSnapshot = true
 }
 
-func (s *EtcdServer) triggerSnapshot(ep *etcdProgress) {
-	if !s.shouldSnapshot(ep) {
+func (s *EtcdServer) snapshotIfNeededAndCompactRaftLog(ep *etcdProgress) {
+	//TODO: Remove disk snapshot in v3.7
+	shouldSnapshotToDisk := s.shouldSnapshotToDisk(ep)
+	shouldSnapshotToMemory := s.shouldSnapshotToMemory(ep)
+	if !shouldSnapshotToDisk && !shouldSnapshotToMemory {
 		return
 	}
-	lg := s.Logger()
-	lg.Info(
-		"triggering snapshot",
-		zap.String("local-member-id", s.MemberID().String()),
-		zap.Uint64("local-member-applied-index", ep.appliedi),
-		zap.Uint64("local-member-snapshot-index", ep.snapi),
-		zap.Uint64("local-member-snapshot-count", s.Cfg.SnapshotCount),
-		zap.Bool("snapshot-forced", s.forceSnapshot),
-	)
-	s.forceSnapshot = false
-
-	s.snapshot(ep.appliedi, ep.confState)
+	s.snapshot(ep, shouldSnapshotToDisk)
 	s.compactRaftLog(ep.appliedi)
-	ep.snapi = ep.appliedi
 }
 
-func (s *EtcdServer) shouldSnapshot(ep *etcdProgress) bool {
-	return (s.forceSnapshot && ep.appliedi != ep.snapi) || (ep.appliedi-ep.snapi > s.Cfg.SnapshotCount)
+func (s *EtcdServer) shouldSnapshotToDisk(ep *etcdProgress) bool {
+	return (s.forceDiskSnapshot && ep.appliedi != ep.diskSnapshotIndex) || (ep.appliedi-ep.diskSnapshotIndex > s.Cfg.SnapshotCount)
+}
+
+func (s *EtcdServer) shouldSnapshotToMemory(ep *etcdProgress) bool {
+	return ep.appliedi > ep.memorySnapshotIndex+memorySnapshotCount
 }
 
 func (s *EtcdServer) hasMultipleVotingMembers() bool {
@@ -2128,23 +2128,33 @@
 }
 
 // TODO: non-blocking snapshot
-func (s *EtcdServer) snapshot(snapi uint64, confState raftpb.ConfState) {
-	d := GetMembershipInfoInV2Format(s.Logger(), s.cluster)
-	// commit kv to write metadata (for example: consistent index) to disk.
-	//
-	// This guarantees that Backend's consistent_index is >= index of last snapshot.
-	//
-	// KV().commit() updates the consistent index in backend.
-	// All operations that update consistent index must be called sequentially
-	// from applyAll function.
-	// So KV().Commit() cannot run in parallel with toApply. It has to be called outside
-	// the go routine created below.
-	s.KV().Commit()
-
+func (s *EtcdServer) snapshot(ep *etcdProgress, toDisk bool) {
 	lg := s.Logger()
+	d := GetMembershipInfoInV2Format(lg, s.cluster)
+	if toDisk {
+		s.Logger().Info(
+			"triggering snapshot",
+			zap.String("local-member-id", s.MemberID().String()),
+			zap.Uint64("local-member-applied-index", ep.appliedi),
+			zap.Uint64("local-member-snapshot-index", ep.diskSnapshotIndex),
+			zap.Uint64("local-member-snapshot-count", s.Cfg.SnapshotCount),
+			zap.Bool("snapshot-forced", s.forceDiskSnapshot),
+		)
+		s.forceDiskSnapshot = false
+		// commit kv to write metadata (for example: consistent index) to disk.
+		//
+		// This guarantees that Backend's consistent_index is >= index of last snapshot.
+		//
+		// KV().commit() updates the consistent index in backend.
+		// All operations that update consistent index must be called sequentially
+		// from applyAll function.
+		// So KV().Commit() cannot run in parallel with toApply. It has to be called outside
+		// the go routine created below.
+		s.KV().Commit()
+	}
 
 	// For backward compatibility, generate v2 snapshot from v3 state.
-	snap, err := s.r.raftStorage.CreateSnapshot(snapi, &confState, d)
+	snap, err := s.r.raftStorage.CreateSnapshot(ep.appliedi, &ep.confState, d)
 	if err != nil {
 		// the snapshot was done asynchronously with the progress of raft.
 		// raft might have already got a newer snapshot.
@@ -2153,21 +2163,25 @@
 		}
 		lg.Panic("failed to create snapshot", zap.Error(err))
 	}
+	ep.memorySnapshotIndex = ep.appliedi
 
 	verifyConsistentIndexIsLatest(lg, snap, s.consistIndex.ConsistentIndex())
 
-	// SaveSnap saves the snapshot to file and appends the corresponding WAL entry.
-	if err = s.r.storage.SaveSnap(snap); err != nil {
-		lg.Panic("failed to save snapshot", zap.Error(err))
-	}
-	if err = s.r.storage.Release(snap); err != nil {
-		lg.Panic("failed to release wal", zap.Error(err))
-	}
+	if toDisk {
+		// SaveSnap saves the snapshot to file and appends the corresponding WAL entry.
+		if err = s.r.storage.SaveSnap(snap); err != nil {
+			lg.Panic("failed to save snapshot", zap.Error(err))
+		}
+		ep.diskSnapshotIndex = ep.appliedi
+		if err = s.r.storage.Release(snap); err != nil {
+			lg.Panic("failed to release wal", zap.Error(err))
+		}
 
-	lg.Info(
-		"saved snapshot",
-		zap.Uint64("snapshot-index", snap.Metadata.Index),
-	)
+		lg.Info(
+			"saved snapshot to disk",
+			zap.Uint64("snapshot-index", snap.Metadata.Index),
+		)
+	}
 }
 
 func (s *EtcdServer) compactRaftLog(snapi uint64) {
@@ -2188,7 +2202,6 @@
 	if snapi > s.Cfg.SnapshotCatchUpEntries {
 		compacti = snapi - s.Cfg.SnapshotCatchUpEntries
 	}
-
 	err := s.r.raftStorage.Compact(compacti)
 	if err != nil {
 		// the compaction was done asynchronously with the progress of raft.
@@ -2198,7 +2211,7 @@
 		}
 		lg.Panic("failed to compact", zap.Error(err))
 	}
-	lg.Info(
+	lg.Debug(
 		"compacted Raft logs",
 		zap.Uint64("compact-index", compacti),
 	)

server/etcdserver/server_test.go

@@ -627,8 +627,8 @@ func TestApplyMultiConfChangeShouldStop(t *testing.T) {
 	}
 }
 
-// TestSnapshot should snapshot the store and cut the persistent
-func TestSnapshot(t *testing.T) {
+// TestSnapshotDisk should save the snapshot to disk and release old snapshots
+func TestSnapshotDisk(t *testing.T) {
 	revertFunc := verify.DisableVerifications()
 	defer revertFunc()
 
@@ -667,24 +667,65 @@ func TestSnapshot(t *testing.T) {
 		gaction, _ := p.Wait(2)
 		defer func() { ch <- struct{}{} }()
 
-		if len(gaction) != 2 {
-			t.Errorf("len(action) = %d, want 2", len(gaction))
-			return
-		}
-		if !reflect.DeepEqual(gaction[0], testutil.Action{Name: "SaveSnap"}) {
-			t.Errorf("action = %s, want SaveSnap", gaction[0])
-		}
+		assert.Len(t, gaction, 2)
+		assert.Equal(t, testutil.Action{Name: "SaveSnap"}, gaction[0])
+		assert.Equal(t, testutil.Action{Name: "Release"}, gaction[1])
+	}()
+	ep := etcdProgress{appliedi: 1, confState: raftpb.ConfState{Voters: []uint64{1}}}
+	srv.snapshot(&ep, true)
+	<-ch
+	assert.Empty(t, st.Action())
+	assert.Equal(t, uint64(1), ep.diskSnapshotIndex)
+	assert.Equal(t, uint64(1), ep.memorySnapshotIndex)
+}
 
-		if !reflect.DeepEqual(gaction[1], testutil.Action{Name: "Release"}) {
-			t.Errorf("action = %s, want Release", gaction[1])
-		}
+func TestSnapshotMemory(t *testing.T) {
+	revertFunc := verify.DisableVerifications()
+	defer revertFunc()
+
+	be, _ := betesting.NewDefaultTmpBackend(t)
+	defer betesting.Close(t, be)
+
+	s := raft.NewMemoryStorage()
+	s.Append([]raftpb.Entry{{Index: 1}})
+	st := mockstore.NewRecorderStream()
+	p := mockstorage.NewStorageRecorderStream("")
+	r := newRaftNode(raftNodeConfig{
+		lg:          zaptest.NewLogger(t),
+		Node:        newNodeNop(),
+		raftStorage: s,
+		storage:     p,
+	})
+	srv := &EtcdServer{
+		lgMu:         new(sync.RWMutex),
+		lg:           zaptest.NewLogger(t),
+		r:            *r,
+		v2store:      st,
+		consistIndex: cindex.NewConsistentIndex(be),
+	}
+	srv.kv = mvcc.New(zaptest.NewLogger(t), be, &lease.FakeLessor{}, mvcc.StoreConfig{})
+	defer func() {
+		assert.NoError(t, srv.kv.Close())
 	}()
+	srv.be = be
 
-	srv.snapshot(1, raftpb.ConfState{Voters: []uint64{1}})
+	cl := membership.NewCluster(zaptest.NewLogger(t))
+	srv.cluster = cl
+
+	ch := make(chan struct{}, 1)
+
+	go func() {
+		gaction, _ := p.Wait(1)
+		defer func() { ch <- struct{}{} }()
+
+		assert.Empty(t, gaction)
+	}()
+	ep := etcdProgress{appliedi: 1, confState: raftpb.ConfState{Voters: []uint64{1}}}
+	srv.snapshot(&ep, false)
 	<-ch
-	if len(st.Action()) != 0 {
-		t.Errorf("no action expected on v2store. Got %d actions", len(st.Action()))
-	}
+	assert.Empty(t, st.Action())
+	assert.Equal(t, uint64(0), ep.diskSnapshotIndex)
+	assert.Equal(t, uint64(1), ep.memorySnapshotIndex)
 }
 
 // TestSnapshotOrdering ensures raft persists snapshot onto disk before

tests/integration/v3_watch_restore_test.go

@@ -111,7 +111,9 @@ func TestV3WatchRestoreSnapshotUnsync(t *testing.T) {
 	//
 	// Since there is no way to confirm server has compacted the log, we
 	// use log monitor to watch and expect "compacted Raft logs" content.
-	expectMemberLog(t, clus.Members[initialLead], 5*time.Second, "compacted Raft logs", 2)
+	// In v3.6 we no longer generates "compacted Raft logs" log as raft compaction happens independently to snapshot.
+	// For now let's use snapshot log which should be equivalent to compaction.
+	expectMemberLog(t, clus.Members[initialLead], 5*time.Second, "saved snapshot to disk", 2)
 
 	// After RecoverPartition, leader L will send snapshot to slow F_m0
 	// follower, because F_m0(index:8) is 'out of date' compared to