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
8 package org.opendaylight.controller.cluster.raft.behaviors;
10 import static java.util.Objects.requireNonNull;
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.actor.Cancellable;
15 import com.google.common.annotations.VisibleForTesting;
16 import com.google.common.io.ByteSource;
17 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
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.Optional;
29 import java.util.OptionalInt;
30 import java.util.Queue;
31 import java.util.concurrent.TimeUnit;
32 import org.eclipse.jdt.annotation.Nullable;
33 import org.opendaylight.controller.cluster.io.SharedFileBackedOutputStream;
34 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
35 import org.opendaylight.controller.cluster.messaging.SliceOptions;
36 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
37 import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
38 import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
39 import org.opendaylight.controller.cluster.raft.PeerInfo;
40 import org.opendaylight.controller.cluster.raft.RaftActorContext;
41 import org.opendaylight.controller.cluster.raft.RaftState;
42 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
43 import org.opendaylight.controller.cluster.raft.VotingState;
44 import org.opendaylight.controller.cluster.raft.base.messages.CheckConsensusReached;
45 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
46 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
47 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
48 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
49 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
50 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
51 import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply;
52 import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
53 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
54 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
55 import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply;
56 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
57 import org.opendaylight.controller.cluster.raft.persisted.Snapshot;
58 import scala.concurrent.duration.FiniteDuration;
61 * The behavior of a RaftActor when it is in the Leader state.
66 * <li> Upon election: send initial empty AppendEntries RPCs
67 * (heartbeat) to each server; repeat during idle periods to
68 * prevent election timeouts (§5.2)
69 * <li> If command received from client: append entry to local log,
70 * respond after entry applied to state machine (§5.3)
71 * <li> If last log index ≥ nextIndex for a follower: send
72 * AppendEntries RPC with log entries starting at nextIndex
73 * <li> If successful: update nextIndex and matchIndex for
75 * <li> If AppendEntries fails because of log inconsistency:
76 * decrement nextIndex and retry (§5.3)
77 * <li> If there exists an N such that N > commitIndex, a majority
78 * of matchIndex[i] ≥ N, and log[N].term == currentTerm:
79 * set commitIndex = N (§5.3, §5.4).
82 public abstract class AbstractLeader extends AbstractRaftActorBehavior {
83 private final Map<String, FollowerLogInformation> followerToLog = new HashMap<>();
86 * Lookup table for request contexts based on journal index. We could use a {@link Map} here, but we really
87 * expect the entries to be modified in sequence, hence we open-code the lookup.
88 * TODO: Evaluate the use of ArrayDeque(), as that has lower memory overhead. Non-head removals are more costly,
89 * but we already expect those to be far from frequent.
91 private final Queue<ClientRequestTracker> trackers = new LinkedList<>();
94 * Map of serialized AppendEntries output streams keyed by log index. This is used in conjunction with the
95 * appendEntriesMessageSlicer for slicing single ReplicatedLogEntry payloads that exceed the message size threshold.
96 * This Map allows the SharedFileBackedOutputStreams to be reused for multiple followers.
98 private final Map<Long, SharedFileBackedOutputStream> sharedSerializedAppendEntriesStreams = new HashMap<>();
99 private final MessageSlicer appendEntriesMessageSlicer;
101 private Cancellable heartbeatSchedule = null;
102 private Optional<SnapshotHolder> snapshotHolder = Optional.empty();
103 private int minReplicationCount;
105 protected AbstractLeader(final RaftActorContext context, final RaftState state,
106 final @Nullable AbstractLeader initializeFromLeader) {
107 super(context, state);
109 appendEntriesMessageSlicer = MessageSlicer.builder().logContext(logName())
110 .messageSliceSize(context.getConfigParams().getSnapshotChunkSize())
111 .expireStateAfterInactivity(context.getConfigParams().getElectionTimeOutInterval().toMillis() * 3,
112 TimeUnit.MILLISECONDS).build();
114 if (initializeFromLeader != null) {
115 followerToLog.putAll(initializeFromLeader.followerToLog);
116 snapshotHolder = initializeFromLeader.snapshotHolder;
117 trackers.addAll(initializeFromLeader.trackers);
119 for (PeerInfo peerInfo: context.getPeers()) {
120 FollowerLogInformation followerLogInformation = new FollowerLogInformation(peerInfo, context);
121 followerToLog.put(peerInfo.getId(), followerLogInformation);
125 log.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
127 updateMinReplicaCount();
129 // Immediately schedule a heartbeat
130 // Upon election: send initial empty AppendEntries RPCs
131 // (heartbeat) to each server; repeat during idle periods to
132 // prevent election timeouts (§5.2)
133 sendAppendEntries(0, false);
135 // It is important to schedule this heartbeat here
136 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
139 protected AbstractLeader(final RaftActorContext context, final RaftState state) {
140 this(context, state, null);
144 * Return an immutable collection of follower identifiers.
146 * @return Collection of follower IDs
148 public final Collection<String> getFollowerIds() {
149 return followerToLog.keySet();
152 public void addFollower(final String followerId) {
153 FollowerLogInformation followerLogInformation = new FollowerLogInformation(context.getPeerInfo(followerId),
155 followerToLog.put(followerId, followerLogInformation);
157 if (heartbeatSchedule == null) {
158 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
162 public void removeFollower(final String followerId) {
163 followerToLog.remove(followerId);
166 public void updateMinReplicaCount() {
168 for (PeerInfo peer: context.getPeers()) {
169 if (peer.isVoting()) {
174 minReplicationCount = getMajorityVoteCount(numVoting);
177 protected int getMinIsolatedLeaderPeerCount() {
178 //the isolated Leader peer count will be 1 less than the majority vote count.
179 //this is because the vote count has the self vote counted in it
181 //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0
182 //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
183 //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
185 return minReplicationCount > 0 ? minReplicationCount - 1 : 0;
189 void setSnapshotHolder(final @Nullable SnapshotHolder snapshotHolder) {
190 this.snapshotHolder = Optional.ofNullable(snapshotHolder);
194 boolean hasSnapshot() {
195 return snapshotHolder.isPresent();
199 protected RaftActorBehavior handleAppendEntries(final ActorRef sender,
200 final AppendEntries appendEntries) {
202 log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
208 protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
209 final AppendEntriesReply appendEntriesReply) {
210 log.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
212 // Update the FollowerLogInformation
213 String followerId = appendEntriesReply.getFollowerId();
214 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
216 if (followerLogInformation == null) {
217 log.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
221 final long lastActivityNanos = followerLogInformation.nanosSinceLastActivity();
222 if (lastActivityNanos > context.getConfigParams().getElectionTimeOutInterval().toNanos()) {
223 log.warn("{} : handleAppendEntriesReply delayed beyond election timeout, "
224 + "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
225 logName(), appendEntriesReply, TimeUnit.NANOSECONDS.toMillis(lastActivityNanos),
226 context.getLastApplied(), context.getCommitIndex());
229 followerLogInformation.markFollowerActive();
230 followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
231 followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
232 followerLogInformation.setNeedsLeaderAddress(appendEntriesReply.isNeedsLeaderAddress());
234 long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
235 boolean updated = false;
236 if (appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) {
237 // The follower's log is actually ahead of the leader's log. Normally this doesn't happen
238 // in raft as a node cannot become leader if it's log is behind another's. However, the
239 // non-voting semantics deviate a bit from raft. Only voting members participate in
240 // elections and can become leader so it's possible for a non-voting follower to be ahead
241 // of the leader. This can happen if persistence is disabled and all voting members are
242 // restarted. In this case, the voting leader will start out with an empty log however
243 // the non-voting followers still retain the previous data in memory. On the first
244 // AppendEntries, the non-voting follower returns a successful reply b/c the prevLogIndex
245 // sent by the leader is -1 and thus the integrity checks pass. However the follower's returned
246 // lastLogIndex may be higher in which case we want to reset the follower by installing a
247 // snapshot. It's also possible that the follower's last log index is behind the leader's.
248 // However in this case the log terms won't match and the logs will conflict - this is handled
250 log.info("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} "
251 + "(snapshotIndex {}, snapshotTerm {}) - forcing install snaphot", logName(),
252 followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(),
253 context.getReplicatedLog().lastIndex(), context.getReplicatedLog().getSnapshotIndex(),
254 context.getReplicatedLog().getSnapshotTerm());
256 followerLogInformation.setMatchIndex(-1);
257 followerLogInformation.setNextIndex(-1);
259 initiateCaptureSnapshot(followerId);
262 } else if (appendEntriesReply.isSuccess()) {
263 long followersLastLogTermInLeadersLog = getLogEntryTerm(followerLastLogIndex);
264 if (followerLastLogIndex >= 0 && followersLastLogTermInLeadersLog >= 0
265 && followersLastLogTermInLeadersLog != appendEntriesReply.getLogLastTerm()) {
266 // The follower's last entry is present in the leader's journal but the terms don't match so the
267 // follower has a conflicting entry. Since the follower didn't report that it's out of sync, this means
268 // either the previous leader entry sent didn't conflict or the previous leader entry is in the snapshot
269 // and no longer in the journal. Either way, we set the follower's next index to 1 less than the last
270 // index reported by the follower. For the former case, the leader will send all entries starting with
271 // the previous follower's index and the follower will remove and replace the conflicting entries as
272 // needed. For the latter, the leader will initiate an install snapshot.
274 followerLogInformation.setNextIndex(followerLastLogIndex - 1);
277 log.info("{}: handleAppendEntriesReply: follower {} last log term {} for index {} conflicts with the "
278 + "leader's {} - set the follower's next index to {}", logName(),
279 followerId, appendEntriesReply.getLogLastTerm(), appendEntriesReply.getLogLastIndex(),
280 followersLastLogTermInLeadersLog, followerLogInformation.getNextIndex());
282 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
285 log.info("{}: handleAppendEntriesReply - received unsuccessful reply: {}, leader snapshotIndex: {}, "
286 + "snapshotTerm: {}, replicatedToAllIndex: {}", logName(), appendEntriesReply,
287 context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm(),
288 getReplicatedToAllIndex());
290 long followersLastLogTermInLeadersLogOrSnapshot = getLogEntryOrSnapshotTerm(followerLastLogIndex);
291 if (appendEntriesReply.isForceInstallSnapshot()) {
292 // Reset the followers match and next index. This is to signal that this follower has nothing
293 // in common with this Leader and so would require a snapshot to be installed
294 followerLogInformation.setMatchIndex(-1);
295 followerLogInformation.setNextIndex(-1);
297 // Force initiate a snapshot capture
298 initiateCaptureSnapshot(followerId);
299 } else if (followerLastLogIndex < 0 || followersLastLogTermInLeadersLogOrSnapshot >= 0
300 && followersLastLogTermInLeadersLogOrSnapshot == appendEntriesReply.getLogLastTerm()) {
301 // The follower's log is empty or the follower's last entry is present in the leader's journal or
302 // snapshot and the terms match so the follower is just behind the leader's journal from the last
303 // snapshot, if any. We'll catch up the follower quickly by starting at the follower's last log index.
305 updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
307 log.info("{}: follower {} appears to be behind the leader from the last snapshot - "
308 + "updated: matchIndex: {}, nextIndex: {}", logName(), followerId,
309 followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
311 // The follower's log conflicts with leader's log so decrement follower's next index
312 // in an attempt to find where the logs match.
313 if (followerLogInformation.decrNextIndex(appendEntriesReply.getLogLastIndex())) {
316 log.info("{}: follower {} last log term {} conflicts with the leader's {} - dec next index to {}",
317 logName(), followerId, appendEntriesReply.getLogLastTerm(),
318 followersLastLogTermInLeadersLogOrSnapshot, followerLogInformation.getNextIndex());
323 if (log.isTraceEnabled()) {
324 log.trace("{}: handleAppendEntriesReply from {}: commitIndex: {}, lastAppliedIndex: {}, currentTerm: {}",
325 logName(), followerId, context.getCommitIndex(), context.getLastApplied(), currentTerm());
328 possiblyUpdateCommitIndex();
330 //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
331 sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
336 private void possiblyUpdateCommitIndex() {
337 // Figure out if we can update the the commitIndex as follows:
338 // If there exists an index N such that N > commitIndex, a majority of matchIndex[i] ≥ N,
339 // and log[N].term == currentTerm:
340 // set commitIndex = N (§5.3, §5.4).
341 for (long index = context.getCommitIndex() + 1; ; index++) {
342 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(index);
343 if (replicatedLogEntry == null) {
344 log.trace("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}",
345 logName(), index, context.getReplicatedLog().getSnapshotIndex(),
346 context.getReplicatedLog().size());
350 // Count our entry if it has been persisted.
351 int replicatedCount = replicatedLogEntry.isPersistencePending() ? 0 : 1;
353 if (replicatedCount == 0) {
354 // We don't commit and apply a log entry until we've gotten the ack from our local persistence,
355 // even though there *shouldn't* be any issue with updating the commit index if we get a consensus
356 // amongst the followers w/o the local persistence ack.
360 log.trace("{}: checking Nth index {}", logName(), index);
361 for (FollowerLogInformation info : followerToLog.values()) {
362 final PeerInfo peerInfo = context.getPeerInfo(info.getId());
363 if (info.getMatchIndex() >= index && peerInfo != null && peerInfo.isVoting()) {
365 } else if (log.isTraceEnabled()) {
366 log.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(),
367 info.getMatchIndex(), peerInfo);
371 if (log.isTraceEnabled()) {
372 log.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount,
373 minReplicationCount);
376 if (replicatedCount >= minReplicationCount) {
377 // Don't update the commit index if the log entry is from a previous term, as per §5.4.1:
378 // "Raft never commits log entries from previous terms by counting replicas".
379 // However we keep looping so we can make progress when new entries in the current term
380 // reach consensus, as per §5.4.1: "once an entry from the current term is committed by
381 // counting replicas, then all prior entries are committed indirectly".
382 if (replicatedLogEntry.getTerm() == currentTerm()) {
383 log.trace("{}: Setting commit index to {}", logName(), index);
384 context.setCommitIndex(index);
386 log.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, "
387 + "term {} does not match the current term {}", logName(), index,
388 replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm());
391 log.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
396 // Apply the change to the state machine
397 if (context.getCommitIndex() > context.getLastApplied()) {
398 log.debug("{}: Applying to log - commitIndex: {}, lastAppliedIndex: {}", logName(),
399 context.getCommitIndex(), context.getLastApplied());
401 applyLogToStateMachine(context.getCommitIndex());
404 if (!context.getSnapshotManager().isCapturing()) {
409 private boolean updateFollowerLogInformation(final FollowerLogInformation followerLogInformation,
410 final AppendEntriesReply appendEntriesReply) {
411 boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
412 updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
414 if (updated && log.isDebugEnabled()) {
416 "{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
417 logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
418 followerLogInformation.getNextIndex());
423 private void purgeInMemoryLog() {
424 //find the lowest index across followers which has been replicated to all.
425 // lastApplied if there are no followers, so that we keep clearing the log for single-node
426 // we would delete the in-mem log from that index on, in-order to minimize mem usage
427 // we would also share this info thru AE with the followers so that they can delete their log entries as well.
428 long minReplicatedToAllIndex = followerToLog.isEmpty() ? context.getLastApplied() : Long.MAX_VALUE;
429 for (FollowerLogInformation info : followerToLog.values()) {
430 minReplicatedToAllIndex = Math.min(minReplicatedToAllIndex, info.getMatchIndex());
433 super.performSnapshotWithoutCapture(minReplicatedToAllIndex);
437 protected ClientRequestTracker removeClientRequestTracker(final long logIndex) {
438 final Iterator<ClientRequestTracker> it = trackers.iterator();
439 while (it.hasNext()) {
440 final ClientRequestTracker t = it.next();
441 if (t.getIndex() == logIndex) {
451 protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
452 final RequestVoteReply requestVoteReply) {
456 protected void beforeSendHeartbeat(){}
459 public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
460 requireNonNull(sender, "sender should not be null");
462 if (appendEntriesMessageSlicer.handleMessage(message)) {
466 if (message instanceof RaftRPC) {
467 RaftRPC rpc = (RaftRPC) message;
468 // If RPC request or response contains term T > currentTerm:
469 // set currentTerm = T, convert to follower (§5.1)
470 // This applies to all RPC messages and responses
471 if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
472 log.info("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
473 logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
475 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
477 // This is a special case. Normally when stepping down as leader we don't process and reply to the
478 // RaftRPC as per raft. But if we're in the process of transferring leadership and we get a
479 // RequestVote, process the RequestVote before switching to Follower. This enables the requesting
480 // candidate node to be elected the leader faster and avoids us possibly timing out in the Follower
481 // state and starting a new election and grabbing leadership back before the other candidate node can
482 // start a new election due to lack of responses. This case would only occur if there isn't a majority
483 // of other nodes available that can elect the requesting candidate. Since we're transferring
484 // leadership, we should make every effort to get the requesting node elected.
485 if (message instanceof RequestVote && context.getRaftActorLeadershipTransferCohort() != null) {
486 log.debug("{}: Leadership transfer in progress - processing RequestVote", logName());
487 super.handleMessage(sender, message);
490 return internalSwitchBehavior(RaftState.Follower);
494 if (message instanceof SendHeartBeat) {
495 beforeSendHeartbeat();
497 scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
498 } else if (message instanceof SendInstallSnapshot) {
499 SendInstallSnapshot sendInstallSnapshot = (SendInstallSnapshot) message;
500 setSnapshotHolder(new SnapshotHolder(sendInstallSnapshot.getSnapshot(),
501 sendInstallSnapshot.getSnapshotBytes()));
502 sendInstallSnapshot();
503 } else if (message instanceof Replicate) {
504 replicate((Replicate) message);
505 } else if (message instanceof InstallSnapshotReply) {
506 handleInstallSnapshotReply((InstallSnapshotReply) message);
507 } else if (message instanceof CheckConsensusReached) {
508 possiblyUpdateCommitIndex();
510 return super.handleMessage(sender, message);
516 @SuppressFBWarnings(value = "NP_NULL_PARAM_DEREF_ALL_TARGETS_DANGEROUS",
517 justification = "JDT nullness with SpotBugs at setSnapshotHolder(null)")
518 private void handleInstallSnapshotReply(final InstallSnapshotReply reply) {
519 log.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
521 String followerId = reply.getFollowerId();
522 FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
523 if (followerLogInformation == null) {
524 // This can happen during AddServer if it times out.
525 log.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
526 logName(), followerId);
530 LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState();
531 if (installSnapshotState == null) {
532 log.error("{}: LeaderInstallSnapshotState not found for follower {} in InstallSnapshotReply",
533 logName(), followerId);
537 installSnapshotState.resetChunkTimer();
538 followerLogInformation.markFollowerActive();
540 if (installSnapshotState.getChunkIndex() == reply.getChunkIndex()) {
541 boolean wasLastChunk = false;
542 if (reply.isSuccess()) {
543 if (installSnapshotState.isLastChunk(reply.getChunkIndex())) {
544 //this was the last chunk reply
546 long followerMatchIndex = snapshotHolder.get().getLastIncludedIndex();
547 followerLogInformation.setMatchIndex(followerMatchIndex);
548 followerLogInformation.setNextIndex(followerMatchIndex + 1);
549 followerLogInformation.clearLeaderInstallSnapshotState();
551 log.info("{}: Snapshot successfully installed on follower {} (last chunk {}) - "
552 + "matchIndex set to {}, nextIndex set to {}", logName(), followerId, reply.getChunkIndex(),
553 followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
555 if (!anyFollowersInstallingSnapshot()) {
556 // once there are no pending followers receiving snapshots
557 // we can remove snapshot from the memory
558 setSnapshotHolder(null);
562 if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED) {
563 UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess =
564 new UnInitializedFollowerSnapshotReply(followerId);
565 context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor());
566 log.debug("Sent message UnInitializedFollowerSnapshotReply to self");
569 installSnapshotState.markSendStatus(true);
572 log.warn("{}: Received failed InstallSnapshotReply - will retry: {}", logName(), reply);
574 installSnapshotState.markSendStatus(false);
578 if (!context.getSnapshotManager().isCapturing()) {
579 // Since the follower is now caught up try to purge the log.
583 ActorSelection followerActor = context.getPeerActorSelection(followerId);
584 if (followerActor != null) {
585 sendSnapshotChunk(followerActor, followerLogInformation);
590 log.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
591 logName(), reply.getChunkIndex(), followerId,
592 installSnapshotState.getChunkIndex());
594 if (reply.getChunkIndex() == LeaderInstallSnapshotState.INVALID_CHUNK_INDEX) {
595 // Since the Follower did not find this index to be valid we should reset the follower snapshot
596 // so that Installing the snapshot can resume from the beginning
597 installSnapshotState.reset();
602 private boolean anyFollowersInstallingSnapshot() {
603 for (FollowerLogInformation info: followerToLog.values()) {
604 if (info.getInstallSnapshotState() != null) {
613 private void replicate(final Replicate replicate) {
614 long logIndex = replicate.getReplicatedLogEntry().getIndex();
616 log.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}, isSendImmediate: {}", logName(),
617 replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass(),
618 replicate.isSendImmediate());
620 // Create a tracker entry we will use this later to notify the
622 if (replicate.getClientActor() != null) {
623 trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(),
627 boolean applyModificationToState = !context.anyVotingPeers()
628 || context.getRaftPolicy().applyModificationToStateBeforeConsensus();
630 if (applyModificationToState) {
631 context.setCommitIndex(logIndex);
632 applyLogToStateMachine(logIndex);
635 if (replicate.isSendImmediate() && !followerToLog.isEmpty()) {
636 sendAppendEntries(0, false);
640 protected void sendAppendEntries(final long timeSinceLastActivityIntervalNanos, final boolean isHeartbeat) {
641 // Send an AppendEntries to all followers
642 for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
643 final String followerId = e.getKey();
644 final FollowerLogInformation followerLogInformation = e.getValue();
645 // This checks helps not to send a repeat message to the follower
646 if (!followerLogInformation.isFollowerActive()
647 || followerLogInformation.nanosSinceLastActivity() >= timeSinceLastActivityIntervalNanos) {
648 sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
654 * This method checks if any update needs to be sent to the given follower. This includes append log entries,
655 * sending next snapshot chunk, and initiating a snapshot.
657 private void sendUpdatesToFollower(final String followerId, final FollowerLogInformation followerLogInformation,
658 final boolean sendHeartbeat, final boolean isHeartbeat) {
660 ActorSelection followerActor = context.getPeerActorSelection(followerId);
661 if (followerActor != null) {
662 long followerNextIndex = followerLogInformation.getNextIndex();
663 boolean isFollowerActive = followerLogInformation.isFollowerActive();
664 boolean sendAppendEntries = false;
665 List<ReplicatedLogEntry> entries = Collections.emptyList();
667 LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState();
668 if (installSnapshotState != null) {
670 // if install snapshot is in process , then sent next chunk if possible
671 if (isFollowerActive) {
672 // 30 seconds with default settings, can be modified via heartbeat or election timeout factor
673 FiniteDuration snapshotReplyTimeout = context.getConfigParams().getHeartBeatInterval()
674 .$times(context.getConfigParams().getElectionTimeoutFactor() * 3);
676 if (installSnapshotState.isChunkTimedOut(snapshotReplyTimeout)) {
677 sendAppendEntries = !resendSnapshotChunk(followerActor, followerLogInformation);
678 } else if (installSnapshotState.canSendNextChunk()) {
679 sendSnapshotChunk(followerActor, followerLogInformation);
681 } else if (sendHeartbeat || followerLogInformation.hasStaleCommitIndex(context.getCommitIndex())) {
682 // we send a heartbeat even if we have not received a reply for the last chunk
683 sendAppendEntries = true;
685 } else if (followerLogInformation.isLogEntrySlicingInProgress()) {
686 sendAppendEntries = sendHeartbeat;
688 long leaderLastIndex = context.getReplicatedLog().lastIndex();
689 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
691 if (!isHeartbeat && log.isDebugEnabled() || log.isTraceEnabled()) {
692 log.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, "
693 + "leaderLastIndex: {}, leaderSnapShotIndex: {}", logName(), followerId, isFollowerActive,
694 followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
697 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
699 log.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
700 followerNextIndex, followerId);
702 if (followerLogInformation.okToReplicate(context.getCommitIndex())) {
703 entries = getEntriesToSend(followerLogInformation, followerActor);
704 sendAppendEntries = true;
706 } else if (isFollowerActive && followerNextIndex >= 0
707 && leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
708 // if the followers next index is not present in the leaders log, and
709 // if the follower is just not starting and if leader's index is more than followers index
710 // then snapshot should be sent
712 // Send heartbeat to follower whenever install snapshot is initiated.
713 sendAppendEntries = true;
714 if (canInstallSnapshot(followerNextIndex)) {
715 log.info("{}: Initiating install snapshot to follower {}: follower nextIndex: {}, leader "
716 + "snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(), followerId,
717 followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
718 context.getReplicatedLog().size());
720 initiateCaptureSnapshot(followerId);
722 // It doesn't seem like we should ever reach here - most likely indicates sonething is
724 log.info("{}: Follower {} is behind but cannot install snapshot: follower nextIndex: {}, "
725 + "leader snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(),
726 followerId, followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
727 context.getReplicatedLog().size());
730 } else if (sendHeartbeat || followerLogInformation.hasStaleCommitIndex(context.getCommitIndex())) {
731 // we send an AppendEntries, even if the follower is inactive
732 // in-order to update the followers timestamp, in case it becomes active again
733 sendAppendEntries = true;
738 if (sendAppendEntries) {
739 sendAppendEntriesToFollower(followerActor, entries, followerLogInformation);
744 private List<ReplicatedLogEntry> getEntriesToSend(final FollowerLogInformation followerLogInfo,
745 final ActorSelection followerActor) {
746 // Try to get all the entries in the journal but not exceeding the max data size for a single AppendEntries
748 int maxEntries = (int) context.getReplicatedLog().size();
749 final int maxDataSize = context.getConfigParams().getSnapshotChunkSize();
750 final long followerNextIndex = followerLogInfo.getNextIndex();
751 List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex,
752 maxEntries, maxDataSize);
754 // If the first entry's size exceeds the max data size threshold, it will be returned from the call above. If
755 // that is the case, then we need to slice it into smaller chunks.
756 if (!(entries.size() == 1 && entries.get(0).getData().size() > maxDataSize)) {
757 // Don't need to slice.
761 log.debug("{}: Log entry size {} exceeds max payload size {}", logName(), entries.get(0).getData().size(),
764 // If an AppendEntries has already been serialized for the log index then reuse the
765 // SharedFileBackedOutputStream.
766 final Long logIndex = entries.get(0).getIndex();
767 SharedFileBackedOutputStream fileBackedStream = sharedSerializedAppendEntriesStreams.get(logIndex);
768 if (fileBackedStream == null) {
769 fileBackedStream = context.getFileBackedOutputStreamFactory().newSharedInstance();
771 final AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
772 getLogEntryIndex(followerNextIndex - 1), getLogEntryTerm(followerNextIndex - 1), entries,
773 context.getCommitIndex(), getReplicatedToAllIndex(), context.getPayloadVersion());
775 log.debug("{}: Serializing {} for slicing for follower {}", logName(), appendEntries,
776 followerLogInfo.getId());
778 try (ObjectOutputStream out = new ObjectOutputStream(fileBackedStream)) {
779 out.writeObject(appendEntries);
780 } catch (IOException e) {
781 log.error("{}: Error serializing {}", logName(), appendEntries, e);
782 fileBackedStream.cleanup();
783 return Collections.emptyList();
786 sharedSerializedAppendEntriesStreams.put(logIndex, fileBackedStream);
788 fileBackedStream.setOnCleanupCallback(index -> {
789 log.debug("{}: On SharedFileBackedOutputStream cleanup for index {}", logName(), index);
790 sharedSerializedAppendEntriesStreams.remove(index);
793 log.debug("{}: Reusing SharedFileBackedOutputStream for follower {}", logName(), followerLogInfo.getId());
794 fileBackedStream.incrementUsageCount();
797 log.debug("{}: Slicing stream for index {}, follower {}", logName(), logIndex, followerLogInfo.getId());
799 // Record that slicing is in progress for the follower.
800 followerLogInfo.setSlicedLogEntryIndex(logIndex);
802 final FollowerIdentifier identifier = new FollowerIdentifier(followerLogInfo.getId());
803 appendEntriesMessageSlicer.slice(SliceOptions.builder().identifier(identifier)
804 .fileBackedOutputStream(fileBackedStream).sendTo(followerActor).replyTo(actor())
805 .onFailureCallback(failure -> {
806 log.error("{}: Error slicing AppendEntries for follower {}", logName(),
807 followerLogInfo.getId(), failure);
808 followerLogInfo.setSlicedLogEntryIndex(FollowerLogInformation.NO_INDEX);
811 return Collections.emptyList();
814 private void sendAppendEntriesToFollower(final ActorSelection followerActor, final List<ReplicatedLogEntry> entries,
815 final FollowerLogInformation followerLogInformation) {
816 // In certain cases outlined below we don't want to send the actual commit index to prevent the follower from
817 // possibly committing and applying conflicting entries (those with same index, different term) from a prior
818 // term that weren't replicated to a majority, which would be a violation of raft.
819 // - if the follower isn't active. In this case we don't know the state of the follower and we send an
820 // empty AppendEntries as a heart beat to prevent election.
821 // - if we're in the process of installing a snapshot. In this case we don't send any new entries but still
822 // need to send AppendEntries to prevent election.
823 // - if we're in the process of slicing an AppendEntries with a large log entry payload. In this case we
824 // need to send an empty AppendEntries to prevent election.
825 boolean isInstallingSnaphot = followerLogInformation.getInstallSnapshotState() != null;
826 long leaderCommitIndex = isInstallingSnaphot || followerLogInformation.isLogEntrySlicingInProgress()
827 || !followerLogInformation.isFollowerActive() ? -1 : context.getCommitIndex();
829 long followerNextIndex = followerLogInformation.getNextIndex();
830 AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
831 getLogEntryIndex(followerNextIndex - 1),
832 getLogEntryTerm(followerNextIndex - 1), entries,
833 leaderCommitIndex, super.getReplicatedToAllIndex(), context.getPayloadVersion(),
834 followerLogInformation.getRaftVersion(), followerLogInformation.needsLeaderAddress(getId()));
836 if (!entries.isEmpty() || log.isTraceEnabled()) {
837 log.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerLogInformation.getId(),
841 followerLogInformation.setSentCommitIndex(leaderCommitIndex);
842 followerActor.tell(appendEntries, actor());
846 * Initiates a snapshot capture to install on a follower.
849 * Install Snapshot works as follows
850 * 1. Leader initiates the capture snapshot by calling createSnapshot on the RaftActor.
851 * 2. On receipt of the CaptureSnapshotReply message, the RaftActor persists the snapshot and makes a call to
852 * the Leader's handleMessage with a SendInstallSnapshot message.
853 * 3. The Leader obtains and stores the Snapshot from the SendInstallSnapshot message and sends it in chunks to
854 * the Follower via InstallSnapshot messages.
855 * 4. For each chunk, the Follower sends back an InstallSnapshotReply.
856 * 5. On receipt of the InstallSnapshotReply for the last chunk, the Leader marks the install complete for that
858 * 6. If another follower requires a snapshot and a snapshot has been collected (via SendInstallSnapshot)
859 * then send the existing snapshot in chunks to the follower.
861 * @param followerId the id of the follower.
862 * @return true if capture was initiated, false otherwise.
864 public boolean initiateCaptureSnapshot(final String followerId) {
865 FollowerLogInformation followerLogInfo = followerToLog.get(followerId);
866 if (snapshotHolder.isPresent()) {
867 // If a snapshot is present in the memory, most likely another install is in progress no need to capture
868 // snapshot. This could happen if another follower needs an install when one is going on.
869 final ActorSelection followerActor = context.getPeerActorSelection(followerId);
871 // Note: sendSnapshotChunk will set the LeaderInstallSnapshotState.
872 sendSnapshotChunk(followerActor, followerLogInfo);
876 boolean captureInitiated = context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
877 this.getReplicatedToAllIndex(), followerId);
878 if (captureInitiated) {
879 followerLogInfo.setLeaderInstallSnapshotState(new LeaderInstallSnapshotState(
880 context.getConfigParams().getSnapshotChunkSize(), logName()));
883 return captureInitiated;
886 private boolean canInstallSnapshot(final long nextIndex) {
887 // If the follower's nextIndex is -1 then we might as well send it a snapshot
888 // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present
890 return nextIndex == -1 || !context.getReplicatedLog().isPresent(nextIndex)
891 && context.getReplicatedLog().isInSnapshot(nextIndex);
896 private void sendInstallSnapshot() {
897 log.debug("{}: sendInstallSnapshot", logName());
898 for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
899 String followerId = e.getKey();
900 ActorSelection followerActor = context.getPeerActorSelection(followerId);
901 FollowerLogInformation followerLogInfo = e.getValue();
903 if (followerActor != null) {
904 long nextIndex = followerLogInfo.getNextIndex();
905 if (followerLogInfo.getInstallSnapshotState() != null
906 || context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED
907 || canInstallSnapshot(nextIndex)) {
908 sendSnapshotChunk(followerActor, followerLogInfo);
915 * Sends a snapshot chunk to a given follower
916 * InstallSnapshot should qualify as a heartbeat too.
918 private void sendSnapshotChunk(final ActorSelection followerActor, final FollowerLogInformation followerLogInfo) {
919 if (snapshotHolder.isPresent()) {
920 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
921 if (installSnapshotState == null) {
922 installSnapshotState = new LeaderInstallSnapshotState(context.getConfigParams().getSnapshotChunkSize(),
924 followerLogInfo.setLeaderInstallSnapshotState(installSnapshotState);
928 // Ensure the snapshot bytes are set - this is a no-op.
929 installSnapshotState.setSnapshotBytes(snapshotHolder.get().getSnapshotBytes());
931 if (!installSnapshotState.canSendNextChunk()) {
935 byte[] nextSnapshotChunk = installSnapshotState.getNextChunk();
937 log.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerLogInfo.getId(),
938 nextSnapshotChunk.length);
940 int nextChunkIndex = installSnapshotState.incrementChunkIndex();
941 Optional<ServerConfigurationPayload> serverConfig = Optional.empty();
942 if (installSnapshotState.isLastChunk(nextChunkIndex)) {
943 serverConfig = Optional.ofNullable(context.getPeerServerInfo(true));
946 sendSnapshotChunk(followerActor, followerLogInfo, nextSnapshotChunk, nextChunkIndex, serverConfig);
948 log.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}", logName(), followerActor.path(),
949 installSnapshotState.getChunkIndex(), installSnapshotState.getTotalChunks());
951 } catch (IOException e) {
952 log.warn("{}: Unable to send chunk: {}/{}. Reseting snapshot progress. Snapshot state: {}", logName(),
953 installSnapshotState.getChunkIndex(), installSnapshotState.getTotalChunks(),
954 installSnapshotState);
955 installSnapshotState.reset();
960 private void sendSnapshotChunk(final ActorSelection followerActor, final FollowerLogInformation followerLogInfo,
961 final byte[] snapshotChunk, final int chunkIndex,
962 final Optional<ServerConfigurationPayload> serverConfig) {
963 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
965 installSnapshotState.startChunkTimer();
967 new InstallSnapshot(currentTerm(), context.getId(),
968 snapshotHolder.get().getLastIncludedIndex(),
969 snapshotHolder.get().getLastIncludedTerm(),
972 installSnapshotState.getTotalChunks(),
973 OptionalInt.of(installSnapshotState.getLastChunkHashCode()),
975 ).toSerializable(followerLogInfo.getRaftVersion()),
980 private boolean resendSnapshotChunk(final ActorSelection followerActor,
981 final FollowerLogInformation followerLogInfo) {
982 if (!snapshotHolder.isPresent()) {
983 // Seems like we should never hit this case, but just in case we do, reset the snapshot progress so that it
984 // can restart from the next AppendEntries.
985 log.warn("{}: Attempting to resend snapshot with no snapshot holder present.", logName());
986 followerLogInfo.clearLeaderInstallSnapshotState();
990 LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
991 // we are resending, timer needs to be reset
992 installSnapshotState.resetChunkTimer();
993 installSnapshotState.markSendStatus(false);
995 sendSnapshotChunk(followerActor, followerLogInfo);
1000 private void sendHeartBeat() {
1001 if (!followerToLog.isEmpty()) {
1002 log.trace("{}: Sending heartbeat", logName());
1003 sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toNanos(), true);
1005 appendEntriesMessageSlicer.checkExpiredSlicedMessageState();
1009 private void stopHeartBeat() {
1010 if (heartbeatSchedule != null && !heartbeatSchedule.isCancelled()) {
1011 heartbeatSchedule.cancel();
1015 private void scheduleHeartBeat(final FiniteDuration interval) {
1016 if (followerToLog.isEmpty()) {
1017 // Optimization - do not bother scheduling a heartbeat as there are
1024 // Schedule a heartbeat. When the scheduler triggers a SendHeartbeat
1025 // message is sent to itself.
1026 // Scheduling the heartbeat only once here because heartbeats do not
1027 // need to be sent if there are other messages being sent to the remote
1029 heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce(
1030 interval, context.getActor(), SendHeartBeat.INSTANCE,
1031 context.getActorSystem().dispatcher(), context.getActor());
1035 public void close() {
1037 appendEntriesMessageSlicer.close();
1041 public final String getLeaderId() {
1042 return context.getId();
1046 public final short getLeaderPayloadVersion() {
1047 return context.getPayloadVersion();
1050 protected boolean isLeaderIsolated() {
1051 int minPresent = getMinIsolatedLeaderPeerCount();
1052 for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
1053 final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId());
1054 if (peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
1056 if (minPresent == 0) {
1061 return minPresent != 0;
1064 // called from example-actor for printing the follower-states
1065 public String printFollowerStates() {
1066 final StringBuilder sb = new StringBuilder();
1069 for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
1071 sb.append(followerLogInformation.getId());
1072 sb.append(" state:");
1073 sb.append(followerLogInformation.isFollowerActive());
1078 return sb.toString();
1082 public FollowerLogInformation getFollower(final String followerId) {
1083 return followerToLog.get(followerId);
1087 public int followerLogSize() {
1088 return followerToLog.size();
1091 static class SnapshotHolder {
1092 private final long lastIncludedTerm;
1093 private final long lastIncludedIndex;
1094 private final ByteSource snapshotBytes;
1096 SnapshotHolder(final Snapshot snapshot, final ByteSource snapshotBytes) {
1097 this.lastIncludedTerm = snapshot.getLastAppliedTerm();
1098 this.lastIncludedIndex = snapshot.getLastAppliedIndex();
1099 this.snapshotBytes = snapshotBytes;
1102 long getLastIncludedTerm() {
1103 return lastIncludedTerm;
1106 long getLastIncludedIndex() {
1107 return lastIncludedIndex;
1110 ByteSource getSnapshotBytes() {
1111 return snapshotBytes;