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.Cancellable;
14 import com.google.common.annotations.VisibleForTesting;
15 import com.google.common.base.Optional;
16 import com.google.common.base.Preconditions;
17 import com.google.common.io.ByteSource;
18 import java.io.IOException;
19 import java.io.ObjectOutputStream;
20 import java.util.Collection;
21 import java.util.Collections;
22 import java.util.HashMap;
23 import java.util.Iterator;
24 import java.util.LinkedList;
25 import java.util.List;
27 import java.util.Map.Entry;
28 import java.util.Queue;
29 import java.util.concurrent.TimeUnit;
30 import javax.annotation.Nullable;
31 import org.opendaylight.controller.cluster.io.SharedFileBackedOutputStream;
32 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
33 import org.opendaylight.controller.cluster.messaging.SliceOptions;
34 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
35 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
36 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
37 import org.opendaylight.controller.cluster.raft.PeerInfo;
38 import org.opendaylight.controller.cluster.raft.RaftActorContext;
39 import org.opendaylight.controller.cluster.raft.RaftState;
40 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
41 import org.opendaylight.controller.cluster.raft.VotingState;
42 import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
43 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
44 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
45 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
46 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
47 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
48 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
49 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
50 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
51 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
52 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
53 import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
54 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
55 import org.opendaylight.controller.cluster.raft.persisted.Snapshot;
56 import scala.concurrent.duration.FiniteDuration;
59 * The behavior of a RaftActor when it is in the Leader state.
64 * <li> Upon election: send initial empty AppendEntries RPCs
65 * (heartbeat) to each server; repeat during idle periods to
66 * prevent election timeouts (§5.2)
67 * <li> If command received from client: append entry to local log,
68 * respond after entry applied to state machine (§5.3)
69 * <li> If last log index ≥ nextIndex for a follower: send
70 * AppendEntries RPC with log entries starting at nextIndex
71 * <li> If successful: update nextIndex and matchIndex for
73 * <li> If AppendEntries fails because of log inconsistency:
74 * decrement nextIndex and retry (§5.3)
75 * <li> If there exists an N such that N > commitIndex, a majority
76 * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
77 * set commitIndex = N (§5.3, §5.4).
80 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
81 private final Map<String, FollowerLogInformation> followerToLog = new HashMap<>();
84 * Lookup table for request contexts based on journal index. We could use a {@link Map} here, but we really
85 * expect the entries to be modified in sequence, hence we open-code the lookup.
86 * TODO: Evaluate the use of ArrayDeque(), as that has lower memory overhead. Non-head removals are more costly,
87 * but we already expect those to be far from frequent.
89 private final Queue<ClientRequestTracker> trackers = new LinkedList<>();
92 * Map of serialized AppendEntries output streams keyed by log index. This is used in conjunction with the
93 * appendEntriesMessageSlicer for slicing single ReplicatedLogEntry payloads that exceed the message size threshold.
94 * This Map allows the SharedFileBackedOutputStreams to be reused for multiple followers.
96 private final Map<Long, SharedFileBackedOutputStream> sharedSerializedAppendEntriesStreams = new HashMap<>();
97 private final MessageSlicer appendEntriesMessageSlicer;
99 private Cancellable heartbeatSchedule = null;
100 private Optional<SnapshotHolder> snapshotHolder = Optional.absent();
101 private int minReplicationCount;
103 protected AbstractLeader(final RaftActorContext context, final RaftState state,
104 @Nullable final AbstractLeader initializeFromLeader) {
105 super(context, state);
107 appendEntriesMessageSlicer = MessageSlicer.builder().logContext(logName())
108 .messageSliceSize(context.getConfigParams().getSnapshotChunkSize())
109 .expireStateAfterInactivity(context.getConfigParams().getElectionTimeOutInterval().toMillis() * 3,
110 TimeUnit.MILLISECONDS).build();
112 if (initializeFromLeader != null) {
113 followerToLog.putAll(initializeFromLeader.followerToLog);
114 snapshotHolder = initializeFromLeader.snapshotHolder;
115 trackers.addAll(initializeFromLeader.trackers);
117 for (PeerInfo peerInfo: context.getPeers()) {
118 FollowerLogInformation followerLogInformation = new FollowerLogInformation(peerInfo, context);
119 followerToLog.put(peerInfo.getId(), followerLogInformation);
123 log.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
125 updateMinReplicaCount();
127 // Immediately schedule a heartbeat
128 // Upon election: send initial empty AppendEntries RPCs
129 // (heartbeat) to each server; repeat during idle periods to
130 // prevent election timeouts (§5.2)
131 sendAppendEntries(0, false);
133 // It is important to schedule this heartbeat here
134 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
137 protected AbstractLeader(final RaftActorContext context, final RaftState state) {
138 this(context, state, null);
142 * Return an immutable collection of follower identifiers.
144 * @return Collection of follower IDs
146 public final Collection<String> getFollowerIds() {
147 return followerToLog.keySet();
150 public void addFollower(final String followerId) {
151 FollowerLogInformation followerLogInformation = new FollowerLogInformation(context.getPeerInfo(followerId),
153 followerToLog.put(followerId, followerLogInformation);
155 if (heartbeatSchedule == null) {
156 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
160 public void removeFollower(final String followerId) {
161 followerToLog.remove(followerId);
164 public void updateMinReplicaCount() {
166 for (PeerInfo peer: context.getPeers()) {
167 if (peer.isVoting()) {
172 minReplicationCount = getMajorityVoteCount(numVoting);
175 protected int getMinIsolatedLeaderPeerCount() {
176 //the isolated Leader peer count will be 1 less than the majority vote count.
177 //this is because the vote count has the self vote counted in it
179 //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
180 //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
181 //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
183 return minReplicationCount > 0 ? minReplicationCount - 1 : 0;
187 void setSnapshotHolder(@Nullable final SnapshotHolder snapshotHolder) {
188 this.snapshotHolder = Optional.fromNullable(snapshotHolder);
192 boolean hasSnapshot() {
193 return snapshotHolder.isPresent();
197 protected RaftActorBehavior handleAppendEntries(final ActorRef sender,
198 final AppendEntries appendEntries) {
200 log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
206 protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
207 final AppendEntriesReply appendEntriesReply) {
208 log.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
210 // Update the FollowerLogInformation
211 String followerId = appendEntriesReply.getFollowerId();
212 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
214 if (followerLogInformation == null) {
215 log.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
219 final long lastActivityNanos = followerLogInformation.nanosSinceLastActivity();
220 if (lastActivityNanos > context.getConfigParams().getElectionTimeOutInterval().toNanos()) {
221 log.warn("{} : handleAppendEntriesReply delayed beyond election timeout, "
222 + "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
223 logName(), appendEntriesReply, TimeUnit.NANOSECONDS.toMillis(lastActivityNanos),
224 context.getLastApplied(), context.getCommitIndex());
227 followerLogInformation.markFollowerActive();
228 followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
229 followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
230 followerLogInformation.setNeedsLeaderAddress(appendEntriesReply.isNeedsLeaderAddress());
232 long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
233 boolean updated = false;
234 if (appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) {
235 // The follower's log is actually ahead of the leader's log. Normally this doesn't happen
236 // in raft as a node cannot become leader if it's log is behind another's. However, the
237 // non-voting semantics deviate a bit from raft. Only voting members participate in
238 // elections and can become leader so it's possible for a non-voting follower to be ahead
239 // of the leader. This can happen if persistence is disabled and all voting members are
240 // restarted. In this case, the voting leader will start out with an empty log however
241 // the non-voting followers still retain the previous data in memory. On the first
242 // AppendEntries, the non-voting follower returns a successful reply b/c the prevLogIndex
243 // sent by the leader is -1 and thus the integrity checks pass. However the follower's returned
244 // lastLogIndex may be higher in which case we want to reset the follower by installing a
245 // snapshot. It's also possible that the follower's last log index is behind the leader's.
246 // However in this case the log terms won't match and the logs will conflict - this is handled
248 log.info("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} "
249 + "(snapshotIndex {}, snapshotTerm {}) - forcing install snaphot", logName(),
250 followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(),
251 context.getReplicatedLog().lastIndex(), context.getReplicatedLog().getSnapshotIndex(),
252 context.getReplicatedLog().getSnapshotTerm());
254 followerLogInformation.setMatchIndex(-1);
255 followerLogInformation.setNextIndex(-1);
257 initiateCaptureSnapshot(followerId);
260 } else if (appendEntriesReply.isSuccess()) {
261 long followersLastLogTermInLeadersLog = getLogEntryTerm(followerLastLogIndex);
262 if (followerLastLogIndex >= 0 && followersLastLogTermInLeadersLog >= 0
263 && followersLastLogTermInLeadersLog != appendEntriesReply.getLogLastTerm()) {
264 // The follower's last entry is present in the leader's journal but the terms don't match so the
265 // follower has a conflicting entry. Since the follower didn't report that it's out of sync, this means
266 // either the previous leader entry sent didn't conflict or the previous leader entry is in the snapshot
267 // and no longer in the journal. Either way, we set the follower's next index to 1 less than the last
268 // index reported by the follower. For the former case, the leader will send all entries starting with
269 // the previous follower's index and the follower will remove and replace the conflicting entries as
270 // needed. For the latter, the leader will initiate an install snapshot.
272 followerLogInformation.setNextIndex(followerLastLogIndex - 1);
275 log.info("{}: handleAppendEntriesReply: follower {} last log term {} for index {} conflicts with the "
276 + "leader's {} - set the follower's next index to {}", logName(),
277 followerId, appendEntriesReply.getLogLastTerm(), appendEntriesReply.getLogLastIndex(),
278 followersLastLogTermInLeadersLog, followerLogInformation.getNextIndex());
280 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
283 log.info("{}: handleAppendEntriesReply - received unsuccessful reply: {}, leader snapshotIndex: {}, "
284 + "snapshotTerm: {}, replicatedToAllIndex: {}", logName(), appendEntriesReply,
285 context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm(),
286 getReplicatedToAllIndex());
288 long followersLastLogTermInLeadersLogOrSnapshot = getLogEntryOrSnapshotTerm(followerLastLogIndex);
289 if (appendEntriesReply.isForceInstallSnapshot()) {
290 // Reset the followers match and next index. This is to signal that this follower has nothing
291 // in common with this Leader and so would require a snapshot to be installed
292 followerLogInformation.setMatchIndex(-1);
293 followerLogInformation.setNextIndex(-1);
295 // Force initiate a snapshot capture
296 initiateCaptureSnapshot(followerId);
297 } else if (followerLastLogIndex < 0 || followersLastLogTermInLeadersLogOrSnapshot >= 0
298 && followersLastLogTermInLeadersLogOrSnapshot == appendEntriesReply.getLogLastTerm()) {
299 // The follower's log is empty or the follower's last entry is present in the leader's journal or
300 // snapshot and the terms match so the follower is just behind the leader's journal from the last
301 // snapshot, if any. We'll catch up the follower quickly by starting at the follower's last log index.
303 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
305 log.info("{}: follower {} appears to be behind the leader from the last snapshot - "
306 + "updated: matchIndex: {}, nextIndex: {}", logName(), followerId,
307 followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
309 // The follower's log conflicts with leader's log so decrement follower's next index
310 // in an attempt to find where the logs match.
311 if (followerLogInformation.decrNextIndex(appendEntriesReply.getLogLastIndex())) {
314 log.info("{}: follower {} last log term {} conflicts with the leader's {} - dec next index to {}",
315 logName(), followerId, appendEntriesReply.getLogLastTerm(),
316 followersLastLogTermInLeadersLogOrSnapshot, followerLogInformation.getNextIndex());
321 if (log.isTraceEnabled()) {
322 log.trace("{}: handleAppendEntriesReply from {}: commitIndex: {}, lastAppliedIndex: {}, currentTerm: {}",
323 logName(), followerId, context.getCommitIndex(), context.getLastApplied(), currentTerm());
326 possiblyUpdateCommitIndex();
328 //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
329 sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
334 private void possiblyUpdateCommitIndex() {
335 // Figure out if we can update the the commitIndex as follows:
336 // If there exists an index N such that N > commitIndex, a majority of matchIndex[i] ≥ N,
337 // and log[N].term == currentTerm:
338 // set commitIndex = N (§5.3, §5.4).
339 for (long index = context.getCommitIndex() + 1; ; index++) {
340 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(index);
341 if (replicatedLogEntry == null) {
342 log.trace("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}",
343 logName(), index, context.getReplicatedLog().getSnapshotIndex(),
344 context.getReplicatedLog().size());
348 // Count our entry if it has been persisted.
349 int replicatedCount = replicatedLogEntry.isPersistencePending() ? 0 : 1;
351 if (replicatedCount == 0) {
352 // We don't commit and apply a log entry until we've gotten the ack from our local persistence,
353 // even though there *shouldn't* be any issue with updating the commit index if we get a consensus
354 // amongst the followers w/o the local persistence ack.
358 log.trace("{}: checking Nth index {}", logName(), index);
359 for (FollowerLogInformation info : followerToLog.values()) {
360 final PeerInfo peerInfo = context.getPeerInfo(info.getId());
361 if (info.getMatchIndex() >= index && peerInfo != null && peerInfo.isVoting()) {
363 } else if (log.isTraceEnabled()) {
364 log.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(),
365 info.getMatchIndex(), peerInfo);
369 if (log.isTraceEnabled()) {
370 log.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount,
371 minReplicationCount);
374 if (replicatedCount >= minReplicationCount) {
375 // Don't update the commit index if the log entry is from a previous term, as per §5.4.1:
376 // "Raft never commits log entries from previous terms by counting replicas".
377 // However we keep looping so we can make progress when new entries in the current term
378 // reach consensus, as per §5.4.1: "once an entry from the current term is committed by
379 // counting replicas, then all prior entries are committed indirectly".
380 if (replicatedLogEntry.getTerm() == currentTerm()) {
381 log.trace("{}: Setting commit index to {}", logName(), index);
382 context.setCommitIndex(index);
384 log.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, "
385 + "term {} does not match the current term {}", logName(), index,
386 replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm());
389 log.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
394 // Apply the change to the state machine
395 if (context.getCommitIndex() > context.getLastApplied()) {
396 log.debug("{}: Applying to log - commitIndex: {}, lastAppliedIndex: {}", logName(),
397 context.getCommitIndex(), context.getLastApplied());
399 applyLogToStateMachine(context.getCommitIndex());
402 if (!context.getSnapshotManager().isCapturing()) {
407 private boolean updateFollowerLogInformation(final FollowerLogInformation followerLogInformation,
408 final AppendEntriesReply appendEntriesReply) {
409 boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
410 updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
412 if (updated && log.isDebugEnabled()) {
414 "{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
415 logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
416 followerLogInformation.getNextIndex());
421 private void purgeInMemoryLog() {
422 //find the lowest index across followers which has been replicated to all.
423 // lastApplied if there are no followers, so that we keep clearing the log for single-node
424 // we would delete the in-mem log from that index on, in-order to minimize mem usage
425 // we would also share this info thru AE with the followers so that they can delete their log entries as well.
426 long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
427 for (FollowerLogInformation info : followerToLog.values()) {
428 minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
431 super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
435 protected ClientRequestTracker removeClientRequestTracker(final long logIndex) {
436 final Iterator<ClientRequestTracker> it = trackers.iterator();
437 while (it.hasNext()) {
438 final ClientRequestTracker t = it.next();
439 if (t.getIndex() == logIndex) {
449 protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
450 final RequestVoteReply requestVoteReply) {
454 protected void beforeSendHeartbeat(){}
457 public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
458 Preconditions.checkNotNull(sender, "sender should not be null");
460 if (appendEntriesMessageSlicer.handleMessage(message)) {
464 if (message instanceof RaftRPC) {
465 RaftRPC rpc = (RaftRPC) message;
466 // If RPC request or response contains term T > currentTerm:
467 // set currentTerm = T, convert to follower (§5.1)
468 // This applies to all RPC messages and responses
469 if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
470 log.info("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
471 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
473 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
475 // This is a special case. Normally when stepping down as leader we don't process and reply to the
476 // RaftRPC as per raft. But if we're in the process of transferring leadership and we get a
477 // RequestVote, process the RequestVote before switching to Follower. This enables the requesting
478 // candidate node to be elected the leader faster and avoids us possibly timing out in the Follower
479 // state and starting a new election and grabbing leadership back before the other candidate node can
480 // start a new election due to lack of responses. This case would only occur if there isn't a majority
481 // of other nodes available that can elect the requesting candidate. Since we're transferring
482 // leadership, we should make every effort to get the requesting node elected.
483 if (message instanceof RequestVote && context.getRaftActorLeadershipTransferCohort() != null) {
484 log.debug("{}: Leadership transfer in progress - processing RequestVote", logName());
485 super.handleMessage(sender, message);
488 return internalSwitchBehavior(RaftState.Follower);
492 if (message instanceof SendHeartBeat) {
493 beforeSendHeartbeat();
495 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
496 } else if (message instanceof SendInstallSnapshot) {
497 SendInstallSnapshot sendInstallSnapshot = (SendInstallSnapshot) message;
498 setSnapshotHolder(new SnapshotHolder(sendInstallSnapshot.getSnapshot(),
499 sendInstallSnapshot.getSnapshotBytes()));
500 sendInstallSnapshot();
501 } else if (message instanceof Replicate) {
502 replicate((Replicate) message);
503 } else if (message instanceof InstallSnapshotReply) {
504 handleInstallSnapshotReply((InstallSnapshotReply) message);
505 } else if (message instanceof CheckConsensusReached) {
506 possiblyUpdateCommitIndex();
508 return super.handleMessage(sender, message);
514 private void handleInstallSnapshotReply(final InstallSnapshotReply reply) {
515 log.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
517 String followerId = reply.getFollowerId();
518 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
519 if (followerLogInformation == null) {
520 // This can happen during AddServer if it times out.
521 log.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
522 logName(), followerId);
526 LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState();
527 if (installSnapshotState == null) {
528 log.error("{}: LeaderInstallSnapshotState not found for follower {} in InstallSnapshotReply",
529 logName(), followerId);
533 installSnapshotState.resetChunkTimer();
534 followerLogInformation.markFollowerActive();
536 if (installSnapshotState.getChunkIndex() == reply.getChunkIndex()) {
537 boolean wasLastChunk = false;
538 if (reply.isSuccess()) {
539 if (installSnapshotState.isLastChunk(reply.getChunkIndex())) {
540 //this was the last chunk reply
542 long followerMatchIndex = snapshotHolder.get().getLastIncludedIndex();
543 followerLogInformation.setMatchIndex(followerMatchIndex);
544 followerLogInformation.setNextIndex(followerMatchIndex + 1);
545 followerLogInformation.clearLeaderInstallSnapshotState();
547 log.info("{}: Snapshot successfully installed on follower {} (last chunk {}) - "
548 + "matchIndex set to {}, nextIndex set to {}", logName(), followerId, reply.getChunkIndex(),
549 followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
551 if (!anyFollowersInstallingSnapshot()) {
552 // once there are no pending followers receiving snapshots
553 // we can remove snapshot from the memory
554 setSnapshotHolder(null);
558 if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED) {
559 UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess =
560 new UnInitializedFollowerSnapshotReply(followerId);
561 context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor());
562 log.debug("Sent message UnInitializedFollowerSnapshotReply to self");
565 installSnapshotState.markSendStatus(true);
568 log.warn("{}: Received failed InstallSnapshotReply - will retry: {}", logName(), reply);
570 installSnapshotState.markSendStatus(false);
574 if (!context.getSnapshotManager().isCapturing()) {
575 // Since the follower is now caught up try to purge the log.
579 ActorSelection followerActor = context.getPeerActorSelection(followerId);
580 if (followerActor != null) {
581 sendSnapshotChunk(followerActor, followerLogInformation);
586 log.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
587 logName(), reply.getChunkIndex(), followerId,
588 installSnapshotState.getChunkIndex());
590 if (reply.getChunkIndex() == LeaderInstallSnapshotState.INVALID_CHUNK_INDEX) {
591 // Since the Follower did not find this index to be valid we should reset the follower snapshot
592 // so that Installing the snapshot can resume from the beginning
593 installSnapshotState.reset();
598 private boolean anyFollowersInstallingSnapshot() {
599 for (FollowerLogInformation info: followerToLog.values()) {
600 if (info.getInstallSnapshotState() != null) {
609 private void replicate(final Replicate replicate) {
610 long logIndex = replicate.getReplicatedLogEntry().getIndex();
612 log.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}, isSendImmediate: {}", logName(),
613 replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass(),
614 replicate.isSendImmediate());
616 // Create a tracker entry we will use this later to notify the
618 if (replicate.getClientActor() != null) {
619 trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(),
623 boolean applyModificationToState = !context.anyVotingPeers()
624 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
626 if (applyModificationToState) {
627 context.setCommitIndex(logIndex);
628 applyLogToStateMachine(logIndex);
631 if (replicate.isSendImmediate() && !followerToLog.isEmpty()) {
632 sendAppendEntries(0, false);
636 protected void sendAppendEntries(final long timeSinceLastActivityIntervalNanos, final boolean isHeartbeat) {
637 // Send an AppendEntries to all followers
638 for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
639 final String followerId = e.getKey();
640 final FollowerLogInformation followerLogInformation = e.getValue();
641 // This checks helps not to send a repeat message to the follower
642 if (!followerLogInformation.isFollowerActive()
643 || followerLogInformation.nanosSinceLastActivity() >= timeSinceLastActivityIntervalNanos) {
644 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
650 * This method checks if any update needs to be sent to the given follower. This includes append log entries,
651 * sending next snapshot chunk, and initiating a snapshot.
653 private void sendUpdatesToFollower(final String followerId, final FollowerLogInformation followerLogInformation,
654 final boolean sendHeartbeat, final boolean isHeartbeat) {
656 ActorSelection followerActor = context.getPeerActorSelection(followerId);
657 if (followerActor != null) {
658 long followerNextIndex = followerLogInformation.getNextIndex();
659 boolean isFollowerActive = followerLogInformation.isFollowerActive();
660 boolean sendAppendEntries = false;
661 List<ReplicatedLogEntry> entries = Collections.emptyList();
663 LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState();
664 if (installSnapshotState != null) {
666 // if install snapshot is in process , then sent next chunk if possible
667 if (isFollowerActive) {
668 // 30 seconds with default settings, can be modified via heartbeat or election timeout factor
669 FiniteDuration snapshotReplyTimeout = context.getConfigParams().getHeartBeatInterval()
670 .$times(context.getConfigParams().getElectionTimeoutFactor() * 3);
672 if (installSnapshotState.isChunkTimedOut(snapshotReplyTimeout)) {
673 sendAppendEntries = !resendSnapshotChunk(followerActor, followerLogInformation);
674 } else if (installSnapshotState.canSendNextChunk()) {
675 sendSnapshotChunk(followerActor, followerLogInformation);
677 } else if (sendHeartbeat || followerLogInformation.hasStaleCommitIndex(context.getCommitIndex())) {
678 // we send a heartbeat even if we have not received a reply for the last chunk
679 sendAppendEntries = true;
681 } else if (followerLogInformation.isLogEntrySlicingInProgress()) {
682 sendAppendEntries = sendHeartbeat;
684 long leaderLastIndex = context.getReplicatedLog().lastIndex();
685 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
687 if (!isHeartbeat && log.isDebugEnabled() || log.isTraceEnabled()) {
688 log.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, "
689 + "leaderLastIndex: {}, leaderSnapShotIndex: {}", logName(), followerId, isFollowerActive,
690 followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
693 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
695 log.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
696 followerNextIndex, followerId);
698 if (followerLogInformation.okToReplicate(context.getCommitIndex())) {
699 entries = getEntriesToSend(followerLogInformation, followerActor);
700 sendAppendEntries = true;
702 } else if (isFollowerActive && followerNextIndex >= 0
703 && leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
704 // if the followers next index is not present in the leaders log, and
705 // if the follower is just not starting and if leader's index is more than followers index
706 // then snapshot should be sent
708 // Send heartbeat to follower whenever install snapshot is initiated.
709 sendAppendEntries = true;
710 if (canInstallSnapshot(followerNextIndex)) {
711 log.info("{}: Initiating install snapshot to follower {}: follower nextIndex: {}, leader "
712 + "snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(), followerId,
713 followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
714 context.getReplicatedLog().size());
716 initiateCaptureSnapshot(followerId);
718 // It doesn't seem like we should ever reach here - most likely indicates sonething is
720 log.info("{}: Follower {} is behind but cannot install snapshot: follower nextIndex: {}, "
721 + "leader snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(),
722 followerId, followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
723 context.getReplicatedLog().size());
726 } else if (sendHeartbeat || followerLogInformation.hasStaleCommitIndex(context.getCommitIndex())) {
727 // we send an AppendEntries, even if the follower is inactive
728 // in-order to update the followers timestamp, in case it becomes active again
729 sendAppendEntries = true;
734 if (sendAppendEntries) {
735 sendAppendEntriesToFollower(followerActor, entries, followerLogInformation);
740 private List<ReplicatedLogEntry> getEntriesToSend(final FollowerLogInformation followerLogInfo,
741 final ActorSelection followerActor) {
742 // Try to get all the entries in the journal but not exceeding the max data size for a single AppendEntries
744 int maxEntries = (int) context.getReplicatedLog().size();
745 final int maxDataSize = context.getConfigParams().getSnapshotChunkSize();
746 final long followerNextIndex = followerLogInfo.getNextIndex();
747 List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex,
748 maxEntries, maxDataSize);
750 // If the first entry's size exceeds the max data size threshold, it will be returned from the call above. If
751 // that is the case, then we need to slice it into smaller chunks.
752 if (!(entries.size() == 1 && entries.get(0).getData().size() > maxDataSize)) {
753 // Don't need to slice.
757 log.debug("{}: Log entry size {} exceeds max payload size {}", logName(), entries.get(0).getData().size(),
760 // If an AppendEntries has already been serialized for the log index then reuse the
761 // SharedFileBackedOutputStream.
762 final Long logIndex = entries.get(0).getIndex();
763 SharedFileBackedOutputStream fileBackedStream = sharedSerializedAppendEntriesStreams.get(logIndex);
764 if (fileBackedStream == null) {
765 fileBackedStream = context.getFileBackedOutputStreamFactory().newSharedInstance();
767 final AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
768 getLogEntryIndex(followerNextIndex - 1), getLogEntryTerm(followerNextIndex - 1), entries,
769 context.getCommitIndex(), getReplicatedToAllIndex(), context.getPayloadVersion());
771 log.debug("{}: Serializing {} for slicing for follower {}", logName(), appendEntries,
772 followerLogInfo.getId());
774 try (ObjectOutputStream out = new ObjectOutputStream(fileBackedStream)) {
775 out.writeObject(appendEntries);
776 } catch (IOException e) {
777 log.error("{}: Error serializing {}", logName(), appendEntries, e);
778 fileBackedStream.cleanup();
779 return Collections.emptyList();
782 sharedSerializedAppendEntriesStreams.put(logIndex, fileBackedStream);
784 fileBackedStream.setOnCleanupCallback(index -> {
785 log.debug("{}: On SharedFileBackedOutputStream cleanup for index {}", logName(), index);
786 sharedSerializedAppendEntriesStreams.remove(index);
789 log.debug("{}: Reusing SharedFileBackedOutputStream for follower {}", logName(), followerLogInfo.getId());
790 fileBackedStream.incrementUsageCount();
793 log.debug("{}: Slicing stream for index {}, follower {}", logName(), logIndex, followerLogInfo.getId());
795 // Record that slicing is in progress for the follower.
796 followerLogInfo.setSlicedLogEntryIndex(logIndex);
798 final FollowerIdentifier identifier = new FollowerIdentifier(followerLogInfo.getId());
799 appendEntriesMessageSlicer.slice(SliceOptions.builder().identifier(identifier)
800 .fileBackedOutputStream(fileBackedStream).sendTo(followerActor).replyTo(actor())
801 .onFailureCallback(failure -> {
802 log.error("{}: Error slicing AppendEntries for follower {}", logName(),
803 followerLogInfo.getId(), failure);
804 followerLogInfo.setSlicedLogEntryIndex(FollowerLogInformation.NO_INDEX);
807 return Collections.emptyList();
810 private void sendAppendEntriesToFollower(final ActorSelection followerActor, final List<ReplicatedLogEntry> entries,
811 final FollowerLogInformation followerLogInformation) {
812 // In certain cases outlined below we don't want to send the actual commit index to prevent the follower from
813 // possibly committing and applying conflicting entries (those with same index, different term) from a prior
814 // term that weren't replicated to a majority, which would be a violation of raft.
815 // - if the follower isn't active. In this case we don't know the state of the follower and we send an
816 // empty AppendEntries as a heart beat to prevent election.
817 // - if we're in the process of installing a snapshot. In this case we don't send any new entries but still
818 // need to send AppendEntries to prevent election.
819 // - if we're in the process of slicing an AppendEntries with a large log entry payload. In this case we
820 // need to send an empty AppendEntries to prevent election.
821 boolean isInstallingSnaphot = followerLogInformation.getInstallSnapshotState() != null;
822 long leaderCommitIndex = isInstallingSnaphot || followerLogInformation.isLogEntrySlicingInProgress()
823 || !followerLogInformation.isFollowerActive() ? -1 : context.getCommitIndex();
825 long followerNextIndex = followerLogInformation.getNextIndex();
826 AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
827 getLogEntryIndex(followerNextIndex - 1),
828 getLogEntryTerm(followerNextIndex - 1), entries,
829 leaderCommitIndex, super.getReplicatedToAllIndex(), context.getPayloadVersion(),
830 followerLogInformation.getRaftVersion(), followerLogInformation.needsLeaderAddress(getId()));
832 if (!entries.isEmpty() || log.isTraceEnabled()) {
833 log.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerLogInformation.getId(),
837 followerLogInformation.setSentCommitIndex(leaderCommitIndex);
838 followerActor.tell(appendEntries, actor());
842 * Initiates a snapshot capture to install on a follower.
845 * Install Snapshot works as follows
846 * 1. Leader initiates the capture snapshot by calling createSnapshot on the RaftActor.
847 * 2. On receipt of the CaptureSnapshotReply message, the RaftActor persists the snapshot and makes a call to
848 * the Leader's handleMessage with a SendInstallSnapshot message.
849 * 3. The Leader obtains and stores the Snapshot from the SendInstallSnapshot message and sends it in chunks to
850 * the Follower via InstallSnapshot messages.
851 * 4. For each chunk, the Follower sends back an InstallSnapshotReply.
852 * 5. On receipt of the InstallSnapshotReply for the last chunk, the Leader marks the install complete for that
854 * 6. If another follower requires a snapshot and a snapshot has been collected (via SendInstallSnapshot)
855 * then send the existing snapshot in chunks to the follower.
857 * @param followerId the id of the follower.
858 * @return true if capture was initiated, false otherwise.
860 public boolean initiateCaptureSnapshot(final String followerId) {
861 FollowerLogInformation followerLogInfo = followerToLog.get(followerId);
862 if (snapshotHolder.isPresent()) {
863 // If a snapshot is present in the memory, most likely another install is in progress no need to capture
864 // snapshot. This could happen if another follower needs an install when one is going on.
865 final ActorSelection followerActor = context.getPeerActorSelection(followerId);
867 // Note: sendSnapshotChunk will set the LeaderInstallSnapshotState.
868 sendSnapshotChunk(followerActor, followerLogInfo);
872 boolean captureInitiated = context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
873 this.getReplicatedToAllIndex(), followerId);
874 if (captureInitiated) {
875 followerLogInfo.setLeaderInstallSnapshotState(new LeaderInstallSnapshotState(
876 context.getConfigParams().getSnapshotChunkSize(), logName()));
879 return captureInitiated;
882 private boolean canInstallSnapshot(final long nextIndex) {
883 // If the follower's nextIndex is -1 then we might as well send it a snapshot
884 // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present
886 return nextIndex == -1 || !context.getReplicatedLog().isPresent(nextIndex)
887 && context.getReplicatedLog().isInSnapshot(nextIndex);
892 private void sendInstallSnapshot() {
893 log.debug("{}: sendInstallSnapshot", logName());
894 for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
895 String followerId = e.getKey();
896 ActorSelection followerActor = context.getPeerActorSelection(followerId);
897 FollowerLogInformation followerLogInfo = e.getValue();
899 if (followerActor != null) {
900 long nextIndex = followerLogInfo.getNextIndex();
901 if (followerLogInfo.getInstallSnapshotState() != null
902 || context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED
903 || canInstallSnapshot(nextIndex)) {
904 sendSnapshotChunk(followerActor, followerLogInfo);
911 * Sends a snapshot chunk to a given follower
912 * InstallSnapshot should qualify as a heartbeat too.
914 private void sendSnapshotChunk(final ActorSelection followerActor, final FollowerLogInformation followerLogInfo) {
915 if (snapshotHolder.isPresent()) {
916 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
917 if (installSnapshotState == null) {
918 installSnapshotState = new LeaderInstallSnapshotState(context.getConfigParams().getSnapshotChunkSize(),
920 followerLogInfo.setLeaderInstallSnapshotState(installSnapshotState);
924 // Ensure the snapshot bytes are set - this is a no-op.
925 installSnapshotState.setSnapshotBytes(snapshotHolder.get().getSnapshotBytes());
927 if (!installSnapshotState.canSendNextChunk()) {
931 byte[] nextSnapshotChunk = installSnapshotState.getNextChunk();
933 log.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerLogInfo.getId(),
934 nextSnapshotChunk.length);
936 int nextChunkIndex = installSnapshotState.incrementChunkIndex();
937 Optional<ServerConfigurationPayload> serverConfig = Optional.absent();
938 if (installSnapshotState.isLastChunk(nextChunkIndex)) {
939 serverConfig = Optional.fromNullable(context.getPeerServerInfo(true));
942 sendSnapshotChunk(followerActor, followerLogInfo, nextSnapshotChunk, nextChunkIndex, serverConfig);
944 } catch (IOException e) {
945 throw new RuntimeException(e);
948 log.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}", logName(), followerActor.path(),
949 installSnapshotState.getChunkIndex(), installSnapshotState.getTotalChunks());
953 private void sendSnapshotChunk(final ActorSelection followerActor, final FollowerLogInformation followerLogInfo,
954 final byte[] snapshotChunk, final int chunkIndex,
955 final Optional<ServerConfigurationPayload> serverConfig) {
956 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
958 installSnapshotState.startChunkTimer();
960 new InstallSnapshot(currentTerm(), context.getId(),
961 snapshotHolder.get().getLastIncludedIndex(),
962 snapshotHolder.get().getLastIncludedTerm(),
965 installSnapshotState.getTotalChunks(),
966 Optional.of(installSnapshotState.getLastChunkHashCode()),
968 ).toSerializable(followerLogInfo.getRaftVersion()),
973 private boolean resendSnapshotChunk(final ActorSelection followerActor,
974 final FollowerLogInformation followerLogInfo) {
975 if (!snapshotHolder.isPresent()) {
976 // Seems like we should never hit this case, but just in case we do, reset the snapshot progress so that it
977 // can restart from the next AppendEntries.
978 log.warn("{}: Attempting to resend snapshot with no snapshot holder present.", logName());
979 followerLogInfo.clearLeaderInstallSnapshotState();
983 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
984 // we are resending, timer needs to be reset
985 installSnapshotState.resetChunkTimer();
986 installSnapshotState.markSendStatus(false);
988 sendSnapshotChunk(followerActor, followerLogInfo);
993 private void sendHeartBeat() {
994 if (!followerToLog.isEmpty()) {
995 log.trace("{}: Sending heartbeat", logName());
996 sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toNanos(), true);
998 appendEntriesMessageSlicer.checkExpiredSlicedMessageState();
1002 private void stopHeartBeat() {
1003 if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
1004 heartbeatSchedule.cancel();
1008 private void scheduleHeartBeat(final FiniteDuration interval) {
1009 if (followerToLog.isEmpty()) {
1010 // Optimization - do not bother scheduling a heartbeat as there are
1017 // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
1018 // message is sent to itself.
1019 // Scheduling the heartbeat only once here because heartbeats do not
1020 // need to be sent if there are other messages being sent to the remote
1022 heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
1023 interval, context.getActor(), SendHeartBeat.INSTANCE,
1024 context.getActorSystem().dispatcher(), context.getActor());
1028 public void close() {
1030 appendEntriesMessageSlicer.close();
1034 public final String getLeaderId() {
1035 return context.getId();
1039 public final short getLeaderPayloadVersion() {
1040 return context.getPayloadVersion();
1043 protected boolean isLeaderIsolated() {
1044 int minPresent = getMinIsolatedLeaderPeerCount();
1045 for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
1046 final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId());
1047 if (peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
1049 if (minPresent == 0) {
1054 return minPresent != 0;
1057 // called from example-actor for printing the follower-states
1058 public String printFollowerStates() {
1059 final StringBuilder sb = new StringBuilder();
1062 for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
1064 sb.append(followerLogInformation.getId());
1065 sb.append(" state:");
1066 sb.append(followerLogInformation.isFollowerActive());
1071 return sb.toString();
1075 public FollowerLogInformation getFollower(final String followerId) {
1076 return followerToLog.get(followerId);
1080 public int followerLogSize() {
1081 return followerToLog.size();
1084 static class SnapshotHolder {
1085 private final long lastIncludedTerm;
1086 private final long lastIncludedIndex;
1087 private final ByteSource snapshotBytes;
1089 SnapshotHolder(final Snapshot snapshot, final ByteSource snapshotBytes) {
1090 this.lastIncludedTerm = snapshot.getLastAppliedTerm();
1091 this.lastIncludedIndex = snapshot.getLastAppliedIndex();
1092 this.snapshotBytes = snapshotBytes;
1095 long getLastIncludedTerm() {
1096 return lastIncludedTerm;
1099 long getLastIncludedIndex() {
1100 return lastIncludedIndex;
1103 ByteSource getSnapshotBytes() {
1104 return snapshotBytes;