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 com.google.common.base.Preconditions;
14 import java.util.Random;
15 import java.util.concurrent.TimeUnit;
16 import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
17 import org.opendaylight.controller.cluster.raft.RaftActorContext;
18 import org.opendaylight.controller.cluster.raft.RaftState;
19 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
20 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
21 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
22 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
23 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
24 import org.opendaylight.controller.cluster.raft.messages.RequestVote;
25 import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
26 import org.opendaylight.controller.cluster.raft.persisted.ApplyJournalEntries;
27 import org.slf4j.Logger;
28 import scala.concurrent.duration.FiniteDuration;
31 * Abstract class that provides common code for a RaftActor behavior.
33 public abstract class AbstractRaftActorBehavior implements RaftActorBehavior {
35 * Information about the RaftActor whose behavior this class represents.
37 protected final RaftActorContext context;
40 * Used for message logging.
42 protected final Logger log;
45 * Prepended to log messages to provide appropriate context.
47 private final String logName;
50 * The RaftState corresponding to his behavior.
52 private final RaftState state;
55 * Used to cancel a scheduled election.
57 private Cancellable electionCancel = null;
60 * The index of the last log entry that has been replicated to all raft peers.
62 private long replicatedToAllIndex = -1;
64 AbstractRaftActorBehavior(final RaftActorContext context, final RaftState state) {
65 this.context = Preconditions.checkNotNull(context);
66 this.state = Preconditions.checkNotNull(state);
67 this.log = context.getLogger();
69 logName = String.format("%s (%s)", context.getId(), state);
72 public static RaftActorBehavior createBehavior(final RaftActorContext context, final RaftState state) {
75 return new Candidate(context);
77 return new Follower(context);
79 return new IsolatedLeader(context);
81 return new Leader(context);
83 return new PreLeader(context);
85 throw new IllegalArgumentException("Unhandled state " + state);
90 public final RaftState state() {
94 protected final String logName() {
99 public void setReplicatedToAllIndex(long replicatedToAllIndex) {
100 this.replicatedToAllIndex = replicatedToAllIndex;
104 public long getReplicatedToAllIndex() {
105 return replicatedToAllIndex;
109 * Derived classes should not directly handle AppendEntries messages it
110 * should let the base class handle it first. Once the base class handles
111 * the AppendEntries message and does the common actions that are applicable
112 * in all RaftState's it will delegate the handling of the AppendEntries
113 * message to the derived class to do more state specific handling by calling
116 * @param sender The actor that sent this message
117 * @param appendEntries The AppendEntries message
118 * @return a new behavior if it was changed or the current behavior
120 protected abstract RaftActorBehavior handleAppendEntries(ActorRef sender,
121 AppendEntries appendEntries);
124 * Handles the common logic for the AppendEntries message and delegates handling to the derived class.
126 * @param sender the ActorRef that sent the message
127 * @param appendEntries the message
128 * @return a new behavior if it was changed or the current behavior
130 protected RaftActorBehavior appendEntries(ActorRef sender, AppendEntries appendEntries) {
132 // 1. Reply false if term < currentTerm (ยง5.1)
133 if (appendEntries.getTerm() < currentTerm()) {
134 log.debug("{}: Cannot append entries because sender term {} is less than {}", logName(),
135 appendEntries.getTerm(), currentTerm());
137 sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex(), lastTerm(),
138 context.getPayloadVersion()), actor());
143 return handleAppendEntries(sender, appendEntries);
147 * Derived classes should not directly handle AppendEntriesReply messages it
148 * should let the base class handle it first. Once the base class handles
149 * the AppendEntriesReply message and does the common actions that are
150 * applicable in all RaftState's it will delegate the handling of the
151 * AppendEntriesReply message to the derived class to do more state specific
152 * handling by calling this method
154 * @param sender The actor that sent this message
155 * @param appendEntriesReply The AppendEntriesReply message
156 * @return a new behavior if it was changed or the current behavior
158 protected abstract RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
159 AppendEntriesReply appendEntriesReply);
162 * Handles the logic for the RequestVote message that is common for all behaviors.
164 * @param sender the ActorRef that sent the message
165 * @param requestVote the message
166 * @return a new behavior if it was changed or the current behavior
168 protected RaftActorBehavior requestVote(ActorRef sender, RequestVote requestVote) {
170 log.debug("{}: In requestVote: {}", logName(), requestVote);
172 boolean grantVote = canGrantVote(requestVote);
175 context.getTermInformation().updateAndPersist(requestVote.getTerm(), requestVote.getCandidateId());
178 RequestVoteReply reply = new RequestVoteReply(currentTerm(), grantVote);
180 log.debug("{}: requestVote returning: {}", logName(), reply);
182 sender.tell(reply, actor());
187 protected boolean canGrantVote(RequestVote requestVote) {
188 boolean grantVote = false;
190 // Reply false if term < currentTerm (ยง5.1)
191 if (requestVote.getTerm() < currentTerm()) {
194 // If votedFor is null or candidateId, and candidateโs log is at
195 // least as up-to-date as receiverโs log, grant vote (ยง5.2, ยง5.4)
196 } else if (votedFor() == null || votedFor()
197 .equals(requestVote.getCandidateId())) {
199 boolean candidateLatest = false;
202 // Raft determines which of two logs is more up-to-date
203 // by comparing the index and term of the last entries in the
204 // logs. If the logs have last entries with different terms, then
205 // the log with the later term is more up-to-date. If the logs
206 // end with the same term, then whichever log is longer is
208 if (requestVote.getLastLogTerm() > lastTerm()) {
209 candidateLatest = true;
210 } else if (requestVote.getLastLogTerm() == lastTerm()
211 && requestVote.getLastLogIndex() >= lastIndex()) {
212 candidateLatest = true;
215 if (candidateLatest) {
223 * Derived classes should not directly handle RequestVoteReply messages it
224 * should let the base class handle it first. Once the base class handles
225 * the RequestVoteReply message and does the common actions that are
226 * applicable in all RaftState's it will delegate the handling of the
227 * RequestVoteReply message to the derived class to do more state specific
228 * handling by calling this method
230 * @param sender The actor that sent this message
231 * @param requestVoteReply The RequestVoteReply message
232 * @return a new behavior if it was changed or the current behavior
234 protected abstract RaftActorBehavior handleRequestVoteReply(ActorRef sender,
235 RequestVoteReply requestVoteReply);
238 * Returns a duration for election with an additional variance for randomness.
240 * @return a random election duration
242 protected FiniteDuration electionDuration() {
243 long variance = new Random().nextInt(context.getConfigParams().getElectionTimeVariance());
244 return context.getConfigParams().getElectionTimeOutInterval().$plus(
245 new FiniteDuration(variance, TimeUnit.MILLISECONDS));
249 * Stops the currently scheduled election.
251 protected void stopElection() {
252 if (electionCancel != null && !electionCancel.isCancelled()) {
253 electionCancel.cancel();
257 protected boolean canStartElection() {
258 return context.getRaftPolicy().automaticElectionsEnabled() && context.isVotingMember();
262 * Schedule a new election.
264 * @param interval the duration after which we should trigger a new election
266 protected void scheduleElection(FiniteDuration interval) {
269 // Schedule an election. When the scheduler triggers an ElectionTimeout message is sent to itself
270 electionCancel = context.getActorSystem().scheduler().scheduleOnce(interval, context.getActor(),
271 ElectionTimeout.INSTANCE, context.getActorSystem().dispatcher(), context.getActor());
275 * Returns the current election term.
277 * @return the current term
279 protected long currentTerm() {
280 return context.getTermInformation().getCurrentTerm();
284 * Returns the id of the candidate that this server voted for in current term.
286 * @return the candidate for whom we voted in the current term
288 protected String votedFor() {
289 return context.getTermInformation().getVotedFor();
293 * Returns the actor associated with this behavior.
297 protected ActorRef actor() {
298 return context.getActor();
302 * Returns the term of the last entry in the log.
306 protected long lastTerm() {
307 return context.getReplicatedLog().lastTerm();
311 * Returns the index of the last entry in the log.
315 protected long lastIndex() {
316 return context.getReplicatedLog().lastIndex();
320 * Removes and returns the ClientRequestTracker for the specified log index.
321 * @param logIndex the log index
322 * @return the ClientRequestTracker or null if none available
324 protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
329 * Returns the actual index of the entry in replicated log for the given index or -1 if not found.
331 * @return the log entry index or -1 if not found
333 protected long getLogEntryIndex(long index) {
334 if (index == context.getReplicatedLog().getSnapshotIndex()) {
335 return context.getReplicatedLog().getSnapshotIndex();
338 ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
340 return entry.getIndex();
347 * Returns the actual term of the entry in replicated log for the given index or -1 if not found.
349 * @return the log entry term or -1 if not found
351 protected long getLogEntryTerm(long index) {
352 if (index == context.getReplicatedLog().getSnapshotIndex()) {
353 return context.getReplicatedLog().getSnapshotTerm();
356 ReplicatedLogEntry entry = context.getReplicatedLog().get(index);
358 return entry.getTerm();
365 * Applies the log entries up to the specified index that is known to be committed to the state machine.
367 * @param index the log index
369 protected void applyLogToStateMachine(final long index) {
370 long newLastApplied = context.getLastApplied();
371 // Now maybe we apply to the state machine
372 for (long i = context.getLastApplied() + 1; i < index + 1; i++) {
374 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(i);
375 if (replicatedLogEntry != null) {
376 // Send a local message to the local RaftActor (it's derived class to be
377 // specific to apply the log to it's index)
379 final ApplyState msg;
380 final ClientRequestTracker tracker = removeClientRequestTracker(i);
381 if (tracker != null) {
382 msg = new ApplyState(tracker.getClientActor(), tracker.getIdentifier(), replicatedLogEntry);
384 msg = new ApplyState(null, null, replicatedLogEntry);
387 actor().tell(msg, actor());
390 //if one index is not present in the log, no point in looping
391 // around as the rest wont be present either
392 log.warn("{}: Missing index {} from log. Cannot apply state. Ignoring {} to {}",
393 logName(), i, i, index);
398 log.debug("{}: Setting last applied to {}", logName(), newLastApplied);
400 context.setLastApplied(newLastApplied);
402 // send a message to persist a ApplyLogEntries marker message into akka's persistent journal
403 // will be used during recovery
404 //in case if the above code throws an error and this message is not sent, it would be fine
405 // as the append entries received later would initiate add this message to the journal
406 actor().tell(new ApplyJournalEntries(context.getLastApplied()), actor());
410 public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
411 if (message instanceof AppendEntries) {
412 return appendEntries(sender, (AppendEntries) message);
413 } else if (message instanceof AppendEntriesReply) {
414 return handleAppendEntriesReply(sender, (AppendEntriesReply) message);
415 } else if (message instanceof RequestVote) {
416 return requestVote(sender, (RequestVote) message);
417 } else if (message instanceof RequestVoteReply) {
418 return handleRequestVoteReply(sender, (RequestVoteReply) message);
425 public RaftActorBehavior switchBehavior(RaftActorBehavior behavior) {
426 return internalSwitchBehavior(behavior);
429 protected RaftActorBehavior internalSwitchBehavior(RaftState newState) {
430 return internalSwitchBehavior(createBehavior(context, newState));
433 @SuppressWarnings("checkstyle:IllegalCatch")
434 protected RaftActorBehavior internalSwitchBehavior(RaftActorBehavior newBehavior) {
435 if (!context.getRaftPolicy().automaticElectionsEnabled()) {
439 log.info("{} :- Switching from behavior {} to {}", logName(), this.state(), newBehavior.state());
442 } catch (RuntimeException e) {
443 log.error("{}: Failed to close behavior : {}", logName(), this.state(), e);
449 protected int getMajorityVoteCount(int numPeers) {
450 // Votes are required from a majority of the peers including self.
451 // The numMajority field therefore stores a calculated value
452 // of the number of votes required for this candidate to win an
453 // election based on it's known peers.
454 // If a peer was added during normal operation and raft replicas
455 // came to know about them then the new peer would also need to be
456 // taken into consideration when calculating this value.
457 // Here are some examples for what the numMajority would be for n
459 // 0 peers = 1 numMajority -: (0 + 1) / 2 + 1 = 1
460 // 2 peers = 2 numMajority -: (2 + 1) / 2 + 1 = 2
461 // 4 peers = 3 numMajority -: (4 + 1) / 2 + 1 = 3
466 numMajority = (numPeers + self) / 2 + 1;
474 * Performs a snapshot with no capture on the replicated log. It clears the log from the supplied index or
475 * lastApplied-1 which ever is minimum.
477 * @param snapshotCapturedIndex the index from which to clear
479 protected void performSnapshotWithoutCapture(final long snapshotCapturedIndex) {
480 long actualIndex = context.getSnapshotManager().trimLog(snapshotCapturedIndex);
482 if (actualIndex != -1) {
483 setReplicatedToAllIndex(actualIndex);
487 protected String getId() {
488 return context.getId();