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.Cancellable;
13 import java.util.Random;
14 import java.util.concurrent.TimeUnit;
15 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
16 import org.opendaylight.controller.cluster.raft.RaftActorContext;
17 import org.opendaylight.controller.cluster.raft.RaftState;
18 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
19 import org.opendaylight.controller.cluster.raft.SerializationUtils;
20 import org.opendaylight.controller.cluster.raft.base.messages.ApplyJournalEntries;
21 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
22 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
23 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
24 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
25 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
26 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
27 import org.slf4j.Logger;
28 import scala.concurrent.duration.FiniteDuration;
31 * Abstract class that represents the behavior of a RaftActor
35 * <li> If commitIndex > lastApplied: increment lastApplied, apply
36 * log[lastApplied] to state machine (§5.3)
37 * <li> If RPC request or response contains term T > currentTerm:
38 * set currentTerm = T, convert to follower (§5.1)
40 public abstract class AbstractRaftActorBehavior implements RaftActorBehavior {
42 protected static final ElectionTimeout ELECTION_TIMEOUT = new ElectionTimeout();
45 * Information about the RaftActor whose behavior this class represents
47 protected final RaftActorContext context;
52 protected final Logger LOG;
57 private Cancellable electionCancel = null;
62 protected String leaderId = null;
64 private short leaderPayloadVersion = -1;
66 private long replicatedToAllIndex = -1;
68 private final String logName;
70 private final RaftState state;
72 protected AbstractRaftActorBehavior(RaftActorContext context, RaftState state) {
73 this.context = context;
75 this.LOG = context.getLogger();
77 logName = String.format("%s (%s)", context.getId(), state);
81 public RaftState state() {
85 public String logName() {
90 public void setReplicatedToAllIndex(long replicatedToAllIndex) {
91 this.replicatedToAllIndex = replicatedToAllIndex;
95 public long getReplicatedToAllIndex() {
96 return replicatedToAllIndex;
100 * Derived classes should not directly handle AppendEntries messages it
101 * should let the base class handle it first. Once the base class handles
102 * the AppendEntries message and does the common actions that are applicable
103 * in all RaftState's it will delegate the handling of the AppendEntries
104 * message to the derived class to do more state specific handling by calling
107 * @param sender The actor that sent this message
108 * @param appendEntries The AppendEntries message
109 * @return a new behavior if it was changed or the current behavior
111 protected abstract RaftActorBehavior handleAppendEntries(ActorRef sender,
112 AppendEntries appendEntries);
116 * appendEntries first processes the AppendEntries message and then
117 * delegates handling to a specific behavior
120 * @param appendEntries
121 * @return a new behavior if it was changed or the current behavior
123 protected RaftActorBehavior appendEntries(ActorRef sender,
124 AppendEntries appendEntries) {
126 // 1. Reply false if term < currentTerm (§5.1)
127 if (appendEntries.getTerm() < currentTerm()) {
128 if(LOG.isDebugEnabled()) {
129 LOG.debug("{}: Cannot append entries because sender term {} is less than {}",
130 logName(), appendEntries.getTerm(), currentTerm());
134 new AppendEntriesReply(context.getId(), currentTerm(), false,
135 lastIndex(), lastTerm(), context.getPayloadVersion()), actor()
141 return handleAppendEntries(sender, appendEntries);
145 * Derived classes should not directly handle AppendEntriesReply messages it
146 * should let the base class handle it first. Once the base class handles
147 * the AppendEntriesReply message and does the common actions that are
148 * applicable in all RaftState's it will delegate the handling of the
149 * AppendEntriesReply message to the derived class to do more state specific
150 * handling by calling this method
152 * @param sender The actor that sent this message
153 * @param appendEntriesReply The AppendEntriesReply message
154 * @return a new behavior if it was changed or the current behavior
156 protected abstract RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
157 AppendEntriesReply appendEntriesReply);
160 * requestVote handles the RequestVote message. This logic is common
165 * @return a new behavior if it was changed or the current behavior
167 protected RaftActorBehavior requestVote(ActorRef sender, RequestVote requestVote) {
169 LOG.debug("{}: In requestVote: {}", logName(), requestVote);
171 boolean grantVote = false;
173 // Reply false if term < currentTerm (§5.1)
174 if (requestVote.getTerm() < currentTerm()) {
177 // If votedFor is null or candidateId, and candidate’s log is at
178 // least as up-to-date as receiver’s log, grant vote (§5.2, §5.4)
179 } else if (votedFor() == null || votedFor()
180 .equals(requestVote.getCandidateId())) {
182 boolean candidateLatest = false;
185 // Raft determines which of two logs is more up-to-date
186 // by comparing the index and term of the last entries in the
187 // logs. If the logs have last entries with different terms, then
188 // the log with the later term is more up-to-date. If the logs
189 // end with the same term, then whichever log is longer is
191 if (requestVote.getLastLogTerm() > lastTerm()) {
192 candidateLatest = true;
193 } else if ((requestVote.getLastLogTerm() == lastTerm())
194 && requestVote.getLastLogIndex() >= lastIndex()) {
195 candidateLatest = true;
198 if (candidateLatest) {
200 context.getTermInformation().updateAndPersist(requestVote.getTerm(),
201 requestVote.getCandidateId());
205 RequestVoteReply reply = new RequestVoteReply(currentTerm(), grantVote);
207 LOG.debug("{}: requestVote returning: {}", logName(), reply);
209 sender.tell(reply, actor());
215 * Derived classes should not directly handle RequestVoteReply messages it
216 * should let the base class handle it first. Once the base class handles
217 * the RequestVoteReply message and does the common actions that are
218 * applicable in all RaftState's it will delegate the handling of the
219 * RequestVoteReply message to the derived class to do more state specific
220 * handling by calling this method
222 * @param sender The actor that sent this message
223 * @param requestVoteReply The RequestVoteReply message
224 * @return a new behavior if it was changed or the current behavior
226 protected abstract RaftActorBehavior handleRequestVoteReply(ActorRef sender,
227 RequestVoteReply requestVoteReply);
231 * @return a random election duration
233 protected FiniteDuration electionDuration() {
234 long variance = new Random().nextInt(context.getConfigParams().getElectionTimeVariance());
235 return context.getConfigParams().getElectionTimeOutInterval().$plus(
236 new FiniteDuration(variance, TimeUnit.MILLISECONDS));
240 * stop the scheduled election
242 protected void stopElection() {
243 if (electionCancel != null && !electionCancel.isCancelled()) {
244 electionCancel.cancel();
249 * schedule a new election
251 * @param interval the duration after which we should trigger a new election
253 protected void scheduleElection(FiniteDuration interval) {
256 // Schedule an election. When the scheduler triggers an ElectionTimeout
257 // message is sent to itself
259 context.getActorSystem().scheduler().scheduleOnce(interval,
260 context.getActor(), ELECTION_TIMEOUT,
261 context.getActorSystem().dispatcher(), context.getActor());
265 * @return the current term
267 protected long currentTerm() {
268 return context.getTermInformation().getCurrentTerm();
272 * @return the candidate for whom we voted in the current term
274 protected String votedFor() {
275 return context.getTermInformation().getVotedFor();
279 * @return the actor associated with this behavior
281 protected ActorRef actor() {
282 return context.getActor();
287 * @return the term from the last entry in the log
289 protected long lastTerm() {
290 return context.getReplicatedLog().lastTerm();
294 * @return the index from the last entry in the log
296 protected long lastIndex() {
297 return context.getReplicatedLog().lastIndex();
302 * @return the client request tracker for the specified logIndex
304 protected ClientRequestTracker findClientRequestTracker(long logIndex) {
310 * @return the client request tracker for the specified logIndex
312 protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
319 * @return log index from the previous to last entry in the log
321 protected long prevLogIndex(long index){
322 ReplicatedLogEntry prevEntry =
323 context.getReplicatedLog().get(index - 1);
324 if (prevEntry != null) {
325 return prevEntry.getIndex();
331 * @return log term from the previous to last entry in the log
333 protected long prevLogTerm(long index){
334 ReplicatedLogEntry prevEntry =
335 context.getReplicatedLog().get(index - 1);
336 if (prevEntry != null) {
337 return prevEntry.getTerm();
343 * Apply the provided index to the state machine
345 * @param index a log index that is known to be committed
347 protected void applyLogToStateMachine(final long index) {
348 long newLastApplied = context.getLastApplied();
349 // Now maybe we apply to the state machine
350 for (long i = context.getLastApplied() + 1;
351 i < index + 1; i++) {
352 ActorRef clientActor = null;
353 String identifier = null;
354 ClientRequestTracker tracker = removeClientRequestTracker(i);
356 if (tracker != null) {
357 clientActor = tracker.getClientActor();
358 identifier = tracker.getIdentifier();
360 ReplicatedLogEntry replicatedLogEntry =
361 context.getReplicatedLog().get(i);
363 if (replicatedLogEntry != null) {
364 // Send a local message to the local RaftActor (it's derived class to be
365 // specific to apply the log to it's index)
366 actor().tell(new ApplyState(clientActor, identifier,
367 replicatedLogEntry), actor());
370 //if one index is not present in the log, no point in looping
371 // around as the rest wont be present either
373 "{}: Missing index {} from log. Cannot apply state. Ignoring {} to {}",
374 logName(), i, i, index);
378 if(LOG.isDebugEnabled()) {
379 LOG.debug("{}: Setting last applied to {}", logName(), newLastApplied);
381 context.setLastApplied(newLastApplied);
383 // send a message to persist a ApplyLogEntries marker message into akka's persistent journal
384 // will be used during recovery
385 //in case if the above code throws an error and this message is not sent, it would be fine
386 // as the append entries received later would initiate add this message to the journal
387 actor().tell(new ApplyJournalEntries(context.getLastApplied()), actor());
390 protected Object fromSerializableMessage(Object serializable){
391 return SerializationUtils.fromSerializable(serializable);
395 public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
396 if (message instanceof AppendEntries) {
397 return appendEntries(sender, (AppendEntries) message);
398 } else if (message instanceof AppendEntriesReply) {
399 return handleAppendEntriesReply(sender, (AppendEntriesReply) message);
400 } else if (message instanceof RequestVote) {
401 return requestVote(sender, (RequestVote) message);
402 } else if (message instanceof RequestVoteReply) {
403 return handleRequestVoteReply(sender, (RequestVoteReply) message);
408 @Override public String getLeaderId() {
413 public short getLeaderPayloadVersion() {
414 return leaderPayloadVersion;
417 public void setLeaderPayloadVersion(short leaderPayloadVersion) {
418 this.leaderPayloadVersion = leaderPayloadVersion;
421 protected RaftActorBehavior switchBehavior(RaftActorBehavior behavior) {
422 LOG.info("{} :- Switching from behavior {} to {}", logName(), this.state(), behavior.state());
425 } catch (Exception e) {
426 LOG.error("{}: Failed to close behavior : {}", logName(), this.state(), e);
432 protected int getMajorityVoteCount(int numPeers) {
433 // Votes are required from a majority of the peers including self.
434 // The numMajority field therefore stores a calculated value
435 // of the number of votes required for this candidate to win an
436 // election based on it's known peers.
437 // If a peer was added during normal operation and raft replicas
438 // came to know about them then the new peer would also need to be
439 // taken into consideration when calculating this value.
440 // Here are some examples for what the numMajority would be for n
442 // 0 peers = 1 numMajority -: (0 + 1) / 2 + 1 = 1
443 // 2 peers = 2 numMajority -: (2 + 1) / 2 + 1 = 2
444 // 4 peers = 3 numMajority -: (4 + 1) / 2 + 1 = 3
449 numMajority = (numPeers + self) / 2 + 1;
457 * Performs a snapshot with no capture on the replicated log.
458 * It clears the log from the supplied index or last-applied-1 which ever is minimum.
460 * @param snapshotCapturedIndex
462 protected void performSnapshotWithoutCapture(final long snapshotCapturedIndex) {
463 long actualIndex = context.getSnapshotManager().trimLog(snapshotCapturedIndex, this);
465 if(actualIndex != -1){
466 setReplicatedToAllIndex(actualIndex);
470 protected String getId(){
471 return context.getId();