Raft batching: Reduce lock contention on raft mutex

This is an attempt to reduce contention between Propose() and
sendAppendEntry(). Change Propose() to acquire a read lock on Raft, and
avoid locking Raft during storeToWAL() (which potentially does IO and
may take a long time). This works as long as sendAppendEntry() is called
from the Raft's goroutine only, unless the entry does not require to be
stored to the Raft log. So the rest of the changes are for enforcing the
above requirement:
  * Change EntryLeaderTransfer so that it is not store to the Raft log.
  * Push EntryPeerState and EntrySnapshot entries to the proposal queue.
  * Make sure EntrySnapshot entries skip the leader check, so make sure
    those are not batched together with other entries.

For performance testing only at this point.

Author: sciascid

server/filestore.go

@@ -8195,6 +8195,10 @@ func fileStoreMsgSizeEstimate(slen, maxPayload int) uint64 {
 	return uint64(emptyRecordLen + slen + 4 + maxPayload)
 }
 
+func (fs *fileStore) MsgSize(msg []byte) uint64 {
+	return fileStoreMsgSizeRaw(0, 0, len(msg))
+}
+
 // ResetState resets any state that's temporary. For example when changing leaders.
 func (fs *fileStore) ResetState() {
 	fs.mu.Lock()

server/jetstream_cluster_1_test.go

@@ -6973,7 +6973,7 @@ func TestJetStreamClusterCatchupMustStallWhenBehindOnApplies(t *testing.T) {
 	n := mset.node.(*raft)
 	n.Lock()
 	ae := n.buildAppendEntry(nil)
-	err = n.storeToWAL(ae)
+	_, _, err = n.storeToWAL(ae)
 	n.Unlock()
 	index, commit, applied := n.Progress()
 	require_NoError(t, err)

server/jetstream_cluster_4_test.go

@@ -4269,7 +4269,7 @@ func TestJetStreamClusterConsumerDontSendSnapshotOnLeaderChange(t *testing.T) {
 	rn.Lock()
 	entries := []*Entry{{EntryNormal, updateDeliveredBuffer()}, {EntryNormal, updateAcksBuffer()}}
 	ae := encode(t, rn.buildAppendEntry(entries))
-	err = rn.storeToWAL(ae)
+	_, _, err = rn.storeToWAL(ae)
 	minPindex := rn.pindex
 	rn.Unlock()
 	require_NoError(t, err)

server/memstore.go

@@ -2068,6 +2068,10 @@ func memStoreMsgSize(subj string, hdr, msg []byte) uint64 {
 	return memStoreMsgSizeRaw(len(subj), len(hdr), len(msg))
 }
 
+func (ms *memStore) MsgSize(msg []byte) uint64 {
+	return memStoreMsgSizeRaw(0, 0, len(msg))
+}
+
 // ResetState resets any state that's temporary. For example when changing leaders.
 func (ms *memStore) ResetState() {
 	ms.mu.Lock()

server/raft.go

@@ -91,6 +91,7 @@ type RaftNode interface {
 type WAL interface {
 	Type() StorageType
 	StoreMsg(subj string, hdr, msg []byte, ttl int64) (uint64, int64, error)
+	MsgSize(msg []byte) uint64
 	LoadMsg(index uint64, sm *StoreMsg) (*StoreMsg, error)
 	RemoveMsg(index uint64) (bool, error)
 	Compact(index uint64) (uint64, error)
@@ -867,19 +868,22 @@ func (s *Server) transferRaftLeaders() bool {
 // Propose will propose a new entry to the group.
 // This should only be called on the leader.
 func (n *raft) Propose(data []byte) error {
-	n.Lock()
-	defer n.Unlock()
-	// Check state under lock, we might not be leader anymore.
-	if state := n.State(); state != Leader {
+	n.RLock()
+	state := n.State()
+	writeError := n.werr
+	prop := n.prop
+	n.RUnlock()
+
+	if state != Leader {
 		n.debug("Proposal ignored, not leader (state: %v)", state)
 		return errNotLeader
 	}
 
-	// Error if we had a previous write error.
-	if werr := n.werr; werr != nil {
-		return werr
+	if writeError != nil {
+		return writeError
 	}
-	n.prop.push(newProposedEntry(newEntry(EntryNormal, data), _EMPTY_))
+
+	prop.push(newProposedEntry(newEntry(EntryNormal, data), _EMPTY_))
 	return nil
 }
 
@@ -1014,10 +1018,9 @@ func (n *raft) AdjustBootClusterSize(csz int) error {
 // Must be the leader.
 func (n *raft) AdjustClusterSize(csz int) error {
 	n.Lock()
-	defer n.Unlock()
-
 	// Check state under lock, we might not be leader anymore.
 	if n.State() != Leader {
+		n.Unlock()
 		return errNotLeader
 	}
 	// Same floor as bootstrap.
@@ -1029,8 +1032,10 @@ func (n *raft) AdjustClusterSize(csz int) error {
 	// a quorum.
 	n.csz = csz
 	n.qn = n.csz/2 + 1
+	n.Unlock()
 
-	n.sendPeerState()
+	n.prop.push(newProposedEntry(
+		newEntry(EntryPeerState, encodePeerState(n.currentPeerState())), _EMPTY_))
 	return nil
 }
 
@@ -1213,10 +1218,8 @@ func (n *raft) encodeSnapshot(snap *snapshot) []byte {
 // Should only be used when the upper layers know this is most recent.
 // Used when restoring streams, moving a stream from R1 to R>1, etc.
 func (n *raft) SendSnapshot(data []byte) error {
-	n.Lock()
-	defer n.Unlock()
-	// Don't check if we're leader before sending and storing, this is used on scaleup.
-	n.sendAppendEntryLocked([]*Entry{{EntrySnapshot, data}}, false)
+	// TODO Need to copy data?
+	n.prop.push(newProposedEntry(newEntry(EntrySnapshot, data), _EMPTY_))
 	return nil
 }
 
@@ -1714,6 +1717,8 @@ func (n *raft) StepDown(preferred ...string) error {
 	// Send the append entry directly rather than via the proposals queue,
 	// as we will switch to follower state immediately and will blow away
 	// the contents of the proposal queue in the process.
+	// Also, we won't store the entry in the Raft log, so it is OK ot call
+	// into sendAppendEntry() directly from here.
 	if maybeLeader != noLeader {
 		n.debug("Selected %q for new leader, stepping down due to leadership transfer", maybeLeader)
 		ae := newEntry(EntryLeaderTransfer, []byte(maybeLeader))
@@ -1829,13 +1834,16 @@ func (n *raft) Peers() []*Peer {
 // Update and propose our known set of peers.
 func (n *raft) ProposeKnownPeers(knownPeers []string) {
 	n.Lock()
-	defer n.Unlock()
 	// If we are the leader update and send this update out.
 	if n.State() != Leader {
+		n.Unlock()
 		return
 	}
 	n.updateKnownPeersLocked(knownPeers)
-	n.sendPeerState()
+	n.Unlock()
+
+	n.prop.push(newProposedEntry(
+		newEntry(EntryPeerState, encodePeerState(n.currentPeerState())), _EMPTY_))
 }
 
 // Update our known set of peers.
@@ -2630,12 +2638,11 @@ func (n *raft) runAsLeader() {
 		n.unsubscribe(rpsub)
 		n.Unlock()
 	}()
-
-	// To send out our initial peer state.
-	n.sendPeerState()
 	n.Unlock()
 
 	var propBatch []*proposedEntry
+	n.sendAppendEntry(
+		[]*Entry{{EntryPeerState, encodePeerState(n.currentPeerState())}})
 
 	hb := time.NewTicker(hbInterval)
 	defer hb.Stop()
@@ -2739,6 +2746,14 @@ func (n *raft) composeBatch(allProposals []*proposedEntry) ([]*Entry, []*propose
 	end := 0
 	for end < len(allProposals) {
 		p := allProposals[end]
+		// If we have a snapshot do not batch with anything else.
+		if p.Type == EntrySnapshot {
+			if end == 0 {
+				sz = len(p.Data) + 1
+				end = 1
+			}
+			break
+		}
 		sz += len(p.Data) + 1
 		end++
 		if sz < maxBatch && end < maxEntries {
@@ -3812,13 +3827,24 @@ CONTINUE:
 	if ae.shouldStore() {
 		// Only store if an original which will have sub != nil
 		if sub != nil {
-			if err := n.storeToWAL(ae); err != nil {
+			n.Unlock()
+			size, seq, err := n.storeToWAL(ae)
+			n.Lock()
+			if err != nil {
 				if err != ErrStoreClosed {
 					n.warn("Error storing entry to WAL: %v", err)
 				}
+				if err == errEntryStoreFailed {
+					n.resetWAL()
+					n.cancelCatchup()
+				}
 				n.Unlock()
 				return
 			}
+			n.bytes += size
+			n.pterm = ae.term
+			n.pindex = seq
+			n.active = time.Now()
 			n.cachePendingEntry(ae)
 			n.resetInitializing()
 		} else {
@@ -3979,50 +4005,54 @@ func (n *raft) buildAppendEntry(entries []*Entry) *appendEntry {
 	return newAppendEntry(n.id, n.term, n.commit, n.pterm, n.pindex, entries)
 }
 
-// Determine if we should store an entry. This stops us from storing
-// heartbeat messages.
+// Determine if we should store an entry.
+// This stops us from storing heartbeat and leader transfer messages.
 func (ae *appendEntry) shouldStore() bool {
-	return ae != nil && len(ae.entries) > 0
+	if ae == nil {
+		return false
+	}
+	l := len(ae.entries)
+	if l == 0 {
+		return false
+	}
+	if l == 1 {
+		return ae.entries[0].Type != EntryLeaderTransfer
+	}
+	return true
+}
+
+func (ae *appendEntry) shouldCheckLeader() bool {
+	if ae != nil && len(ae.entries) == 1 &&
+		ae.entries[0].Type == EntrySnapshot {
+		return true
+	}
+	return false
 }
 
 // Store our append entry to our WAL.
-// lock should be held.
-func (n *raft) storeToWAL(ae *appendEntry) error {
+// Returns the number of bytes written and the sequence number
+// assigned to the message.
+func (n *raft) storeToWAL(ae *appendEntry) (uint64, uint64, error) {
 	if ae == nil {
-		return fmt.Errorf("raft: Missing append entry for storage")
+		return 0, 0, fmt.Errorf("raft: Missing append entry for storage")
 	}
+
 	if n.werr != nil {
-		return n.werr
+		return 0, 0, n.werr
 	}
 
 	seq, _, err := n.wal.StoreMsg(_EMPTY_, nil, ae.buf, 0)
 	if err != nil {
-		n.setWriteErrLocked(err)
-		return err
+		return 0, 0, err
 	}
-
 	// Sanity checking for now.
 	if index := ae.pindex + 1; index != seq {
 		n.warn("Wrong index, ae is %+v, index stored was %d, n.pindex is %d, will reset", ae, seq, n.pindex)
-		if n.State() == Leader {
-			n.stepdownLocked(n.selectNextLeader())
-		}
-		// Reset and cancel any catchup.
-		n.resetWAL()
-		n.cancelCatchup()
-		return errEntryStoreFailed
+		return 0, 0, errEntryStoreFailed
 	}
 
-	var sz uint64
-	if n.wtype == FileStorage {
-		sz = fileStoreMsgSize(_EMPTY_, nil, ae.buf)
-	} else {
-		sz = memStoreMsgSize(_EMPTY_, nil, ae.buf)
-	}
-	n.bytes += sz
-	n.pterm = ae.term
-	n.pindex = seq
-	return nil
+	sz := n.wal.MsgSize(ae.buf)
+	return sz, seq, nil
 }
 
 const (
@@ -4031,19 +4061,25 @@ const (
 	paeWarnModulo    = 5_000
 )
 
+// sendAppendEntry builds a appendEntry and stores it to the WAL,
+// before sending it to the followers.
+// It is expected for this method to be called from Raft's main
+// goroutine, unless the appendEntry does not need to be stored
+// to the WAL (heartbeat or EntryLeaderTransfer)
 func (n *raft) sendAppendEntry(entries []*Entry) {
-	n.Lock()
-	defer n.Unlock()
-	n.sendAppendEntryLocked(entries, true)
-}
-func (n *raft) sendAppendEntryLocked(entries []*Entry, checkLeader bool) {
-	// Safeguard against sending an append entry right after a stepdown from a different goroutine.
-	// Specifically done while holding the lock to not race.
-	if checkLeader && n.State() != Leader {
+	// Safeguard against sending an append entry right after a stepdown
+	// from a different goroutine. Specifically done while holding the
+	// lock to not race.
+	n.RLock()
+	state := n.State()
+	ae := n.buildAppendEntry(entries)
+	n.RUnlock()
+
+	if ae.shouldCheckLeader() && state != Leader {
 		n.debug("Not sending append entry, not leader")
+		ae.returnToPool()
 		return
 	}
-	ae := n.buildAppendEntry(entries)
 
 	var err error
 	var scratch [1024]byte
@@ -4055,12 +4091,30 @@ func (n *raft) sendAppendEntryLocked(entries []*Entry, checkLeader bool) {
 	// If we have entries store this in our wal.
 	shouldStore := ae.shouldStore()
 	if shouldStore {
-		if err := n.storeToWAL(ae); err != nil {
+		size, seq, err := n.storeToWAL(ae)
+		n.Lock()
+		if err != nil {
+			n.setWriteErrLocked(err)
+			if err == errEntryStoreFailed {
+				if n.State() == Leader {
+					n.stepdownLocked(n.selectNextLeader())
+				}
+				// are we sure we want this?
+				n.resetWAL()
+				n.cancelCatchup()
+			}
+			n.Unlock()
 			return
 		}
+
+		n.bytes += size
+		n.pterm = ae.term
+		n.pindex = seq
 		n.active = time.Now()
 		n.cachePendingEntry(ae)
+		n.Unlock()
 	}
+
 	n.sendRPC(n.asubj, n.areply, ae.buf)
 	if !shouldStore {
 		ae.returnToPool()
@@ -4149,23 +4203,15 @@ func (n *raft) peerNames() []string {
 
 func (n *raft) currentPeerState() *peerState {
 	n.RLock()
-	ps := n.currentPeerStateLocked()
-	n.RUnlock()
-	return ps
-}
-
-func (n *raft) currentPeerStateLocked() *peerState {
+	defer n.RUnlock()
 	return &peerState{n.peerNames(), n.csz, n.extSt}
 }
 
-// sendPeerState will send our current peer state to the cluster.
-// Lock should be held.
-func (n *raft) sendPeerState() {
-	n.sendAppendEntryLocked([]*Entry{{EntryPeerState, encodePeerState(n.currentPeerStateLocked())}}, true)
-}
-
 // Send a heartbeat.
 func (n *raft) sendHeartbeat() {
+	// OK to call sendAppendEntry() directly here.
+	// No need to push heardbeats into prop queue
+	// because we don't store those into the log.
 	n.sendAppendEntry(nil)
 }