2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
9 package org.opendaylight.controller.cluster.raft.behaviors;
11 import akka.actor.ActorRef;
12 import akka.actor.ActorSelection;
13 import akka.actor.Address;
14 import akka.cluster.Cluster;
15 import akka.cluster.ClusterEvent.CurrentClusterState;
16 import akka.cluster.Member;
17 import akka.cluster.MemberStatus;
18 import akka.japi.Procedure;
19 import com.google.common.annotations.VisibleForTesting;
20 import com.google.common.base.Stopwatch;
21 import java.io.IOException;
22 import java.util.ArrayList;
23 import java.util.List;
24 import java.util.Optional;
26 import java.util.concurrent.TimeUnit;
27 import java.util.concurrent.atomic.AtomicBoolean;
28 import javax.annotation.Nullable;
29 import org.opendaylight.controller.cluster.raft.RaftActorContext;
30 import org.opendaylight.controller.cluster.raft.RaftState;
31 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
32 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
33 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
34 import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
35 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
36 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
37 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
38 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
39 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
40 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
41 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
42 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
43 import org.opendaylight.controller.cluster.raft.persisted.Snapshot;
46 * The behavior of a RaftActor in the Follower raft state.
48 * <li> Respond to RPCs from candidates and leaders
49 * <li> If election timeout elapses without receiving AppendEntries
50 * RPC from current leader or granting vote to candidate:
51 * convert to candidate
54 public class Follower extends AbstractRaftActorBehavior {
55 private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18;
57 private final SyncStatusTracker initialSyncStatusTracker;
59 private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted();
60 private SnapshotTracker snapshotTracker = null;
61 private String leaderId;
62 private short leaderPayloadVersion;
64 public Follower(final RaftActorContext context) {
65 this(context, null, (short)-1);
68 public Follower(final RaftActorContext context, final String initialLeaderId,
69 final short initialLeaderPayloadVersion) {
70 super(context, RaftState.Follower);
71 this.leaderId = initialLeaderId;
72 this.leaderPayloadVersion = initialLeaderPayloadVersion;
74 initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), context.getConfigParams()
75 .getSyncIndexThreshold());
77 if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
78 actor().tell(TimeoutNow.INSTANCE, actor());
80 scheduleElection(electionDuration());
85 public final String getLeaderId() {
90 protected final void setLeaderId(@Nullable final String leaderId) {
91 this.leaderId = leaderId;
95 public short getLeaderPayloadVersion() {
96 return leaderPayloadVersion;
100 protected final void setLeaderPayloadVersion(final short leaderPayloadVersion) {
101 this.leaderPayloadVersion = leaderPayloadVersion;
104 private void restartLastLeaderMessageTimer() {
105 if (lastLeaderMessageTimer.isRunning()) {
106 lastLeaderMessageTimer.reset();
109 lastLeaderMessageTimer.start();
112 private boolean isLogEntryPresent(final long index) {
113 if (context.getReplicatedLog().isInSnapshot(index)) {
117 ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
118 return entry != null;
122 private void updateInitialSyncStatus(final long currentLeaderCommit, final String newLeaderId) {
123 initialSyncStatusTracker.update(newLeaderId, currentLeaderCommit, context.getCommitIndex());
127 protected RaftActorBehavior handleAppendEntries(final ActorRef sender, final AppendEntries appendEntries) {
129 int numLogEntries = appendEntries.getEntries().size();
130 if (log.isTraceEnabled()) {
131 log.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
132 } else if (log.isDebugEnabled() && numLogEntries > 0) {
133 log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
136 // TODO : Refactor this method into a bunch of smaller methods
137 // to make it easier to read. Before refactoring ensure tests
138 // cover the code properly
140 if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) {
141 log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the "
142 + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
143 closeSnapshotTracker();
146 if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
147 // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
148 AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
149 lastIndex(), lastTerm(), context.getPayloadVersion());
151 log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
152 sender.tell(reply, actor());
157 // If we got here then we do appear to be talking to the leader
158 leaderId = appendEntries.getLeaderId();
159 leaderPayloadVersion = appendEntries.getPayloadVersion();
161 // First check if the logs are in sync or not
162 long lastIndex = lastIndex();
164 if (isOutOfSync(appendEntries)) {
165 // We found that the log was out of sync so just send a negative
168 final AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
169 lastTerm(), context.getPayloadVersion());
171 log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply);
172 updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
173 sender.tell(reply, actor());
177 if (numLogEntries > 0) {
178 log.debug("{}: Number of entries to be appended = {}", logName(), numLogEntries);
180 // 3. If an existing entry conflicts with a new one (same index
181 // but different terms), delete the existing entry and all that
183 int addEntriesFrom = 0;
184 if (context.getReplicatedLog().size() > 0) {
186 // Find the entry up until the one that is not in the follower's log
187 for (int i = 0;i < numLogEntries; i++, addEntriesFrom++) {
188 ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
190 if (!isLogEntryPresent(matchEntry.getIndex())) {
191 // newEntry not found in the log
195 long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
197 log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
200 // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know
201 // what the term was so we'll assume it matches.
202 if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) {
206 if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
208 log.info("{}: Removing entries from log starting at {}", logName(), matchEntry.getIndex());
210 // Entries do not match so remove all subsequent entries
211 if (!context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) {
212 // Could not remove the entries - this means the matchEntry index must be in the
213 // snapshot and not the log. In this case the prior entries are part of the state
214 // so we must send back a reply to force a snapshot to completely re-sync the
215 // follower's log and state.
217 log.info("{}: Could not remove entries - sending reply to force snapshot", logName());
218 updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
219 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
220 lastTerm(), context.getPayloadVersion(), true), actor());
226 updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
227 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
228 lastTerm(), context.getPayloadVersion(), true), actor());
234 lastIndex = lastIndex();
235 log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(),
236 lastIndex, addEntriesFrom);
238 // When persistence successfully completes for each new log entry appended, we need to determine if we
239 // should capture a snapshot to compact the persisted log. shouldCaptureSnapshot tracks whether or not
240 // one of the log entries has exceeded the log size threshold whereby a snapshot should be taken. However
241 // we don't initiate the snapshot at that log entry but rather after the last log entry has been persisted.
242 // This is done because subsequent log entries after the one that tripped the threshold may have been
243 // applied to the state already, as the persistence callback occurs async, and we want those entries
244 // purged from the persisted log as well.
245 final AtomicBoolean shouldCaptureSnapshot = new AtomicBoolean(false);
246 final Procedure<ReplicatedLogEntry> appendAndPersistCallback = logEntry -> {
247 final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
248 final ReplicatedLogEntry lastEntryToAppend = entries.get(entries.size() - 1);
249 if (shouldCaptureSnapshot.get() && logEntry == lastEntryToAppend) {
250 context.getSnapshotManager().capture(context.getReplicatedLog().last(), getReplicatedToAllIndex());
254 // 4. Append any new entries not already in the log
255 for (int i = addEntriesFrom; i < numLogEntries; i++) {
256 ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
258 log.debug("{}: Append entry to log {}", logName(), entry.getData());
260 context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false);
262 shouldCaptureSnapshot.compareAndSet(false,
263 context.getReplicatedLog().shouldCaptureSnapshot(entry.getIndex()));
265 if (entry.getData() instanceof ServerConfigurationPayload) {
266 context.updatePeerIds((ServerConfigurationPayload)entry.getData());
270 log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
273 // 5. If leaderCommit > commitIndex, set commitIndex =
274 // min(leaderCommit, index of last new entry)
276 lastIndex = lastIndex();
277 long prevCommitIndex = context.getCommitIndex();
279 if (appendEntries.getLeaderCommit() > prevCommitIndex) {
280 context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
283 if (prevCommitIndex != context.getCommitIndex()) {
284 log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
287 AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
288 lastIndex, lastTerm(), context.getPayloadVersion());
290 if (log.isTraceEnabled()) {
291 log.trace("{}: handleAppendEntries returning : {}", logName(), reply);
292 } else if (log.isDebugEnabled() && numLogEntries > 0) {
293 log.debug("{}: handleAppendEntries returning : {}", logName(), reply);
296 // Reply to the leader before applying any previous state so as not to hold up leader consensus.
297 updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
298 sender.tell(reply, actor());
300 // If commitIndex > lastApplied: increment lastApplied, apply
301 // log[lastApplied] to state machine (§5.3)
302 // check if there are any entries to be applied. last-applied can be equal to last-index
303 if (appendEntries.getLeaderCommit() > context.getLastApplied()
304 && context.getLastApplied() < lastIndex) {
305 if (log.isDebugEnabled()) {
306 log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {}, "
307 + "context.getLastApplied(): {}, lastIndex(): {}", logName(),
308 appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
311 applyLogToStateMachine(appendEntries.getLeaderCommit());
314 if (!context.getSnapshotManager().isCapturing()) {
315 super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
321 private boolean isOutOfSync(final AppendEntries appendEntries) {
323 final long lastIndex = lastIndex();
324 if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
326 // The follower's log is out of sync because the leader does have an entry at prevLogIndex and this
327 // follower has no entries in it's log.
329 log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(),
330 appendEntries.getPrevLogIndex());
334 if (lastIndex > -1) {
335 if (isLogEntryPresent(appendEntries.getPrevLogIndex())) {
336 final long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
337 if (prevLogTerm != appendEntries.getPrevLogTerm()) {
339 // The follower's log is out of sync because the Leader's prevLogIndex entry does exist
340 // in the follower's log but it has a different term in it
342 log.info("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append "
343 + "entries prevLogTerm {} - lastIndex: {}, snapshotIndex: {}", logName(),
344 appendEntries.getPrevLogIndex(), prevLogTerm, appendEntries.getPrevLogTerm(), lastIndex,
345 context.getReplicatedLog().getSnapshotIndex());
348 } else if (appendEntries.getPrevLogIndex() != -1) {
350 // The follower's log is out of sync because the Leader's prevLogIndex entry was not found in it's log
352 log.info("{}: The log is not empty but the prevLogIndex {} was not found in it - lastIndex: {}, "
353 + "snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
354 context.getReplicatedLog().getSnapshotIndex());
359 if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
360 && appendEntries.getReplicatedToAllIndex() != -1) {
361 if (!isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
362 // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
363 // the previous entry in it's in-memory journal
365 log.info("{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the "
366 + "in-memory journal", logName(), appendEntries.getReplicatedToAllIndex());
370 final List<ReplicatedLogEntry> entries = appendEntries.getEntries();
371 if (entries.size() > 0 && !isLogEntryPresent(entries.get(0).getIndex() - 1)) {
372 log.info("{}: Cannot append entries because the calculated previousIndex {} was not found in the "
373 + "in-memory journal", logName(), entries.get(0).getIndex() - 1);
382 protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
383 final AppendEntriesReply appendEntriesReply) {
388 protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
389 final RequestVoteReply requestVoteReply) {
394 public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
395 if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
396 return handleElectionTimeout(message);
399 if (!(message instanceof RaftRPC)) {
400 // The rest of the processing requires the message to be a RaftRPC
404 final RaftRPC rpc = (RaftRPC) message;
405 // If RPC request or response contains term T > currentTerm:
406 // set currentTerm = T, convert to follower (§5.1)
407 // This applies to all RPC messages and responses
408 if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
409 log.info("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
410 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
412 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
415 if (rpc instanceof InstallSnapshot) {
416 handleInstallSnapshot(sender, (InstallSnapshot) rpc);
417 restartLastLeaderMessageTimer();
418 scheduleElection(electionDuration());
422 if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
423 restartLastLeaderMessageTimer();
424 scheduleElection(electionDuration());
427 return super.handleMessage(sender, rpc);
430 private RaftActorBehavior handleElectionTimeout(final Object message) {
431 // If the message is ElectionTimeout, verify we haven't actually seen a message from the leader
432 // during the election timeout interval. It may that the election timer expired b/c this actor
433 // was busy and messages got delayed, in which case leader messages would be backed up in the
434 // queue but would be processed before the ElectionTimeout message and thus would restart the
435 // lastLeaderMessageTimer.
436 long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
437 long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
438 boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning()
439 || lastLeaderMessageInterval >= electionTimeoutInMillis;
441 if (canStartElection()) {
442 if (message instanceof TimeoutNow) {
443 log.debug("{}: Received TimeoutNow - switching to Candidate", logName());
444 return internalSwitchBehavior(RaftState.Candidate);
445 } else if (noLeaderMessageReceived) {
446 // Check the cluster state to see if the leader is known to be up before we go to Candidate.
447 // However if we haven't heard from the leader in a long time even though the cluster state
448 // indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
450 long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
451 if (isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
452 log.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
453 scheduleElection(electionDuration());
455 log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
456 return internalSwitchBehavior(RaftState.Candidate);
459 log.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
460 logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
461 scheduleElection(electionDuration());
463 } else if (message instanceof ElectionTimeout) {
464 if (noLeaderMessageReceived) {
468 scheduleElection(electionDuration());
474 private boolean isLeaderAvailabilityKnown() {
475 if (leaderId == null) {
479 Optional<Cluster> cluster = context.getCluster();
480 if (!cluster.isPresent()) {
484 ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
485 if (leaderActor == null) {
489 Address leaderAddress = leaderActor.anchorPath().address();
491 CurrentClusterState state = cluster.get().state();
492 Set<Member> unreachable = state.getUnreachable();
494 log.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
497 for (Member m: unreachable) {
498 if (leaderAddress.equals(m.address())) {
499 log.info("{}: Leader {} is unreachable", logName(), leaderAddress);
504 for (Member m: state.getMembers()) {
505 if (leaderAddress.equals(m.address())) {
506 if (m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
507 log.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
508 leaderAddress, m.status());
511 log.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
512 leaderAddress, m.status());
518 log.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
523 private void handleInstallSnapshot(final ActorRef sender, final InstallSnapshot installSnapshot) {
525 log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
527 leaderId = installSnapshot.getLeaderId();
529 if (snapshotTracker == null) {
530 snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId(),
534 updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
537 final InstallSnapshotReply reply = new InstallSnapshotReply(
538 currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
540 if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
541 installSnapshot.getLastChunkHashCode())) {
543 log.info("{}: Snapshot installed from leader: {}", logName(), installSnapshot.getLeaderId());
545 Snapshot snapshot = Snapshot.create(
546 context.getSnapshotManager().convertSnapshot(snapshotTracker.getSnapshotBytes()),
548 installSnapshot.getLastIncludedIndex(),
549 installSnapshot.getLastIncludedTerm(),
550 installSnapshot.getLastIncludedIndex(),
551 installSnapshot.getLastIncludedTerm(),
552 context.getTermInformation().getCurrentTerm(),
553 context.getTermInformation().getVotedFor(),
554 installSnapshot.getServerConfig().orNull());
556 ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
558 public void onSuccess() {
559 log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
561 sender.tell(reply, actor());
565 public void onFailure() {
566 sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
570 actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
572 closeSnapshotTracker();
574 log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
576 sender.tell(reply, actor());
578 } catch (IOException e) {
579 log.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
581 sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
582 -1, false), actor());
584 closeSnapshotTracker();
588 private void closeSnapshotTracker() {
589 if (snapshotTracker != null) {
590 snapshotTracker.close();
591 snapshotTracker = null;
596 public void close() {
597 closeSnapshotTracker();
602 SnapshotTracker getSnapshotTracker() {
603 return snapshotTracker;