Make stream snapshots asynchronous

The monitorStream function can block for a long time when creating and
installing a snapshot of a stream's state. This can lead to increased
tail latency.
This commit extends the RaftNode interface with a new InstallSnapshotAsync
method. This method performs snapshot writing and WAL compaction in a
in a separate goroutine, making the process non-blocking. The existing
InstallSnapshot method is now a synchronous wrapper around the new
asynchronous implementation.

Signed-off-by: Daniele Sciascia <[email protected]>

Author: sciascid

server/jetstream_cluster.go

@@ -2519,9 +2519,21 @@ func (js *jetStream) monitorStream(mset *stream, sa *streamAssignment, sendSnaps
 	// fully recovered from disk.
 	isRecovering := true
 
-	doSnapshot := func() {
+	snapState := struct {
+		inProgress bool
+		curState   SimpleState
+		ch         chan InstalledSnapshot
+	}{
+		ch: make(chan InstalledSnapshot, 1),
+	}
+
+	wantSnapshot := func() bool {
 		if mset == nil || isRecovering || isRestore {
-			return
+			return false
+		}
+
+		if snapState.inProgress {
+			return false
 		}
 
 		// Before we actually calculate the detailed state and encode it, let's check the
@@ -2534,18 +2546,56 @@ func (js *jetStream) monitorStream(mset *stream, sa *streamAssignment, sendSnaps
 		// consumers on idle streams but better to be safe than sorry!
 		ne, nb := n.Size()
 		if curState == lastState && ne < compactNumMin && nb < compactSizeMin {
-			return
+			return false
+		}
+
+		snapState.curState = curState
+		return true
+	}
+
+	handleSnapshotErr := func(err error) {
+		switch err {
+		case nil:
+			lastState = snapState.curState
+		case errNoSnapAvailable, errNodeClosed, errCatchupsRunning:
+			// ignore the error
+		default:
+			s.RateLimitWarnf("Failed to install snapshot for '%s > %s' [%s]: %v",
+				mset.acc.Name, mset.name(), n.Group(), err)
+		}
+
+	}
+
+	doSnapshot := func() {
+		if wantSnapshot() {
+			// Make sure all pending data is flushed before allowing snapshots.
+			mset.flushAllPending()
+			err := n.InstallSnapshot(mset.stateSnapshot())
+			handleSnapshotErr(err)
 		}
+	}
 
-		// Make sure all pending data is flushed before allowing snapshots.
-		mset.flushAllPending()
-		if err := n.InstallSnapshot(mset.stateSnapshot()); err == nil {
-			lastState = curState
-		} else if err != errNoSnapAvailable && err != errNodeClosed && err != errCatchupsRunning {
-			s.RateLimitWarnf("Failed to install snapshot for '%s > %s' [%s]: %v", mset.acc.Name, mset.name(), n.Group(), err)
+	doSnapshotAsync := func() {
+		if wantSnapshot() {
+			// Make sure all pending data is flushed before allowing snapshots.
+			mset.flushAllPending()
+			n.InstallSnapshotAsync(mset.stateSnapshot(), snapState.ch)
+			snapState.inProgress = true
 		}
 	}
 
+	snapshotDone := func(snap InstalledSnapshot) {
+		handleSnapshotErr(snap.Err)
+		snapState.inProgress = false
+	}
+
+	defer func() {
+		if snapState.inProgress {
+			s := <-snapState.ch
+			snapshotDone(s)
+		}
+	}()
+
 	// We will establish a restoreDoneCh no matter what. Will never be triggered unless
 	// we replace with the restore chan.
 	restoreDoneCh := make(<-chan error)
@@ -2617,6 +2667,8 @@ func (js *jetStream) monitorStream(mset *stream, sa *streamAssignment, sendSnaps
 
 	for {
 		select {
+		case s := <-snapState.ch:
+			snapshotDone(s)
 		case <-s.quitCh:
 			// Server shutting down, but we might receive this before qch, so try to snapshot.
 			doSnapshot()
@@ -2726,7 +2778,7 @@ func (js *jetStream) monitorStream(mset *stream, sa *streamAssignment, sendSnaps
 			// Check about snapshotting
 			// If we have at least min entries to compact, go ahead and try to snapshot/compact.
 			if ne >= compactNumMin || nb > compactSizeMin || mset.getCLFS() > pclfs {
-				doSnapshot()
+				doSnapshotAsync()
 			}
 
 		case isLeader = <-lch:
@@ -2822,7 +2874,7 @@ func (js *jetStream) monitorStream(mset *stream, sa *streamAssignment, sendSnaps
 			stopDirectMonitoring()
 
 		case <-t.C:
-			doSnapshot()
+			doSnapshotAsync()
 
 		case <-uch:
 			// keep stream assignment current

server/raft.go

@@ -42,6 +42,7 @@ type RaftNode interface {
 	ProposeMulti(entries []*Entry) error
 	ForwardProposal(entry []byte) error
 	InstallSnapshot(snap []byte) error
+	InstallSnapshotAsync(snap []byte, ch chan<- InstalledSnapshot)
 	SendSnapshot(snap []byte) error
 	NeedSnapshot() bool
 	Applied(index uint64) (entries uint64, bytes uint64)
@@ -230,6 +231,7 @@ type raft struct {
 	observer     bool // The node is observing, i.e. not able to become leader
 	initializing bool // The node is new, and "empty log" checks can be temporarily relaxed.
 	scaleUp      bool // The node is part of a scale up, puts us in observer mode until the log contains data.
+	snapshotting bool // Snapshot is in progress
 }
 
 type proposedEntry struct {
@@ -307,6 +309,7 @@ var (
 	errNodeClosed        = errors.New("raft: node is closed")
 	errBadSnapName       = errors.New("raft: snapshot name could not be parsed")
 	errNoSnapAvailable   = errors.New("raft: no snapshot available")
+	errSnapInProgress    = errors.New("raft: snapshot is already in progress")
 	errCatchupsRunning   = errors.New("raft: snapshot can not be installed while catchups running")
 	errSnapshotCorrupt   = errors.New("raft: snapshot corrupt")
 	errTooManyPrefs      = errors.New("raft: stepdown requires at most one preferred new leader")
@@ -1242,49 +1245,103 @@ func (n *raft) SendSnapshot(data []byte) error {
 	return nil
 }
 
-// Used to install a snapshot for the given term and applied index. This will release
-// all of the log entries up to and including index. This should not be called with
-// entries that have been applied to the FSM but have not been applied to the raft state.
-func (n *raft) InstallSnapshot(data []byte) error {
+type InstalledSnapshot struct {
+	Term  uint64
+	Index uint64
+	Path  string
+	Err   error
+}
+
+func (n *raft) installSnapshotAsync(encoded []byte, snap InstalledSnapshot,
+	ch chan<- InstalledSnapshot) {
+	go func() {
+		snap.Err = writeFileWithSync(snap.Path, encoded, defaultFilePerms)
+		n.Lock()
+		if snap.Err == nil {
+			// Delete our previous snapshot file if it exists.
+			if n.snapfile != _EMPTY_ && n.snapfile != snap.Path {
+				os.Remove(n.snapfile)
+			}
+			// Remember our latest snapshot file.
+			n.snapfile = snap.Path
+			_, snap.Err = n.wal.Compact(snap.Index + 1)
+			if snap.Err != nil {
+				n.setWriteErrLocked(snap.Err)
+			} else {
+				var state StreamState
+				n.wal.FastState(&state)
+				n.papplied = snap.Index
+				n.bytes = state.Bytes
+			}
+		}
+		n.snapshotting = false
+		n.Unlock()
+		ch <- snap
+	}()
+}
+
+// InstallSnapshotAsync installs a snapshot asynchronously. It writes the
+// snapshot to disk and compacts the WAL in a separate goroutine. The caller
+// is notified of the result on the provided channel.
+func (n *raft) InstallSnapshotAsync(data []byte, ch chan<- InstalledSnapshot) {
 	if n.State() == Closed {
-		return errNodeClosed
+		ch <- InstalledSnapshot{Err: errNodeClosed}
+		return
 	}
 
 	n.Lock()
 	defer n.Unlock()
 
+	if n.snapshotting {
+		ch <- InstalledSnapshot{Err: errSnapInProgress}
+		return
+	}
+
 	// If a write error has occurred already then stop here.
-	if werr := n.werr; werr != nil {
-		return werr
+	if n.werr != nil {
+		ch <- InstalledSnapshot{Err: n.werr}
+		return
 	}
 
-	// Check that a catchup isn't already taking place. If it is then we won't
-	// allow installing snapshots until it is done.
+	// Check that a catchup isn't already taking place. If it is then we
+	// won't allow installing snapshots until it is done.
 	if len(n.progress) > 0 || n.paused {
-		return errCatchupsRunning
+		ch <- InstalledSnapshot{Err: errCatchupsRunning}
+		return
 	}
 
 	if n.applied == 0 {
-		n.debug("Not snapshotting as there are no applied entries")
-		return errNoSnapAvailable
+		ch <- InstalledSnapshot{Err: errNoSnapAvailable}
+		return
 	}
 
-	var term uint64
-	if ae, _ := n.loadEntry(n.applied); ae != nil {
-		term = ae.term
-	} else {
-		n.debug("Not snapshotting as entry %d is not available", n.applied)
-		return errNoSnapAvailable
+	ae, _ := n.loadEntry(n.applied)
+	if ae == nil {
+		ch <- InstalledSnapshot{Err: errNoSnapAvailable}
+		return
 	}
 
-	n.debug("Installing snapshot of %d bytes [%d:%d]", len(data), term, n.applied)
-
-	return n.installSnapshot(&snapshot{
-		lastTerm:  term,
+	peers := encodePeerState(&peerState{n.peerNames(), n.csz, n.extSt})
+	encoded := n.encodeSnapshot(&snapshot{
+		lastTerm:  ae.term,
 		lastIndex: n.applied,
-		peerstate: encodePeerState(&peerState{n.peerNames(), n.csz, n.extSt}),
-		data:      data,
-	})
+		peerstate: peers,
+		data:      data})
+
+	snapDir := filepath.Join(n.sd, snapshotsDir)
+	snapFile := filepath.Join(snapDir, fmt.Sprintf(snapFileT, ae.term, n.applied))
+	snap := InstalledSnapshot{Term: ae.term, Index: n.applied, Path: snapFile}
+	n.installSnapshotAsync(encoded, snap, ch)
+}
+
+// InstallSnapshot installs a snapshot for the current applied index. This is a
+// synchronous call that will block until the snapshot is installed, and will
+// release all of the log entries up to the applied index.
+func (n *raft) InstallSnapshot(data []byte) error {
+	ch := make(chan InstalledSnapshot, 1)
+	n.InstallSnapshotAsync(data, ch)
+	snap := <-ch
+	return snap.Err
 }
 
 // Install the snapshot.