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
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
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
156 protected abstract RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
157 AppendEntriesReply appendEntriesReply);
160 * requestVote handles the RequestVote message. This logic is common
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
226 protected abstract RaftActorBehavior handleRequestVoteReply(ActorRef sender,
227 RequestVoteReply requestVoteReply);
230 * Creates a random election duration
234 protected FiniteDuration electionDuration() {
235 long variance = new Random().nextInt(context.getConfigParams().getElectionTimeVariance());
236 return context.getConfigParams().getElectionTimeOutInterval().$plus(
237 new FiniteDuration(variance, TimeUnit.MILLISECONDS));
241 * stop the scheduled election
243 protected void stopElection() {
244 if (electionCancel != null && !electionCancel.isCancelled()) {
245 electionCancel.cancel();
250 * schedule a new election
254 protected void scheduleElection(FiniteDuration interval) {
257 // Schedule an election. When the scheduler triggers an ElectionTimeout
258 // message is sent to itself
260 context.getActorSystem().scheduler().scheduleOnce(interval,
261 context.getActor(), ELECTION_TIMEOUT,
262 context.getActorSystem().dispatcher(), context.getActor());
266 * Get the current term
269 protected long currentTerm() {
270 return context.getTermInformation().getCurrentTerm();
274 * Get the candidate for whom we voted in the current term
277 protected String votedFor() {
278 return context.getTermInformation().getVotedFor();
282 * Get the actor associated with this behavior
285 protected ActorRef actor() {
286 return context.getActor();
290 * Get the term from the last entry in the log
294 protected long lastTerm() {
295 return context.getReplicatedLog().lastTerm();
299 * Get the index from the last entry in the log
303 protected long lastIndex() {
304 return context.getReplicatedLog().lastIndex();
308 * Find the client request tracker for a specific logIndex
313 protected ClientRequestTracker findClientRequestTracker(long logIndex) {
318 * Find the client request tracker for a specific logIndex
323 protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
329 * Find the log index from the previous to last entry in the log
333 protected long prevLogIndex(long index){
334 ReplicatedLogEntry prevEntry =
335 context.getReplicatedLog().get(index - 1);
336 if (prevEntry != null) {
337 return prevEntry.getIndex();
343 * Find the log term from the previous to last entry in the log
346 protected long prevLogTerm(long index){
347 ReplicatedLogEntry prevEntry =
348 context.getReplicatedLog().get(index - 1);
349 if (prevEntry != null) {
350 return prevEntry.getTerm();
356 * Apply the provided index to the state machine
358 * @param index a log index that is known to be committed
360 protected void applyLogToStateMachine(final long index) {
361 long newLastApplied = context.getLastApplied();
362 // Now maybe we apply to the state machine
363 for (long i = context.getLastApplied() + 1;
364 i < index + 1; i++) {
365 ActorRef clientActor = null;
366 String identifier = null;
367 ClientRequestTracker tracker = removeClientRequestTracker(i);
369 if (tracker != null) {
370 clientActor = tracker.getClientActor();
371 identifier = tracker.getIdentifier();
373 ReplicatedLogEntry replicatedLogEntry =
374 context.getReplicatedLog().get(i);
376 if (replicatedLogEntry != null) {
377 // Send a local message to the local RaftActor (it's derived class to be
378 // specific to apply the log to it's index)
379 actor().tell(new ApplyState(clientActor, identifier,
380 replicatedLogEntry), actor());
383 //if one index is not present in the log, no point in looping
384 // around as the rest wont be present either
386 "{}: Missing index {} from log. Cannot apply state. Ignoring {} to {}",
387 logName(), i, i, index);
391 if(LOG.isDebugEnabled()) {
392 LOG.debug("{}: Setting last applied to {}", logName(), newLastApplied);
394 context.setLastApplied(newLastApplied);
396 // send a message to persist a ApplyLogEntries marker message into akka's persistent journal
397 // will be used during recovery
398 //in case if the above code throws an error and this message is not sent, it would be fine
399 // as the append entries received later would initiate add this message to the journal
400 actor().tell(new ApplyJournalEntries(context.getLastApplied()), actor());
403 protected Object fromSerializableMessage(Object serializable){
404 return SerializationUtils.fromSerializable(serializable);
408 public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
409 if (message instanceof AppendEntries) {
410 return appendEntries(sender, (AppendEntries) message);
411 } else if (message instanceof AppendEntriesReply) {
412 return handleAppendEntriesReply(sender, (AppendEntriesReply) message);
413 } else if (message instanceof RequestVote) {
414 return requestVote(sender, (RequestVote) message);
415 } else if (message instanceof RequestVoteReply) {
416 return handleRequestVoteReply(sender, (RequestVoteReply) message);
421 @Override public String getLeaderId() {
426 public short getLeaderPayloadVersion() {
427 return leaderPayloadVersion;
430 public void setLeaderPayloadVersion(short leaderPayloadVersion) {
431 this.leaderPayloadVersion = leaderPayloadVersion;
434 protected RaftActorBehavior switchBehavior(RaftActorBehavior behavior) {
435 LOG.info("{} :- Switching from behavior {} to {}", logName(), this.state(), behavior.state());
438 } catch (Exception e) {
439 LOG.error("{}: Failed to close behavior : {}", logName(), this.state(), e);
445 protected int getMajorityVoteCount(int numPeers) {
446 // Votes are required from a majority of the peers including self.
447 // The numMajority field therefore stores a calculated value
448 // of the number of votes required for this candidate to win an
449 // election based on it's known peers.
450 // If a peer was added during normal operation and raft replicas
451 // came to know about them then the new peer would also need to be
452 // taken into consideration when calculating this value.
453 // Here are some examples for what the numMajority would be for n
455 // 0 peers = 1 numMajority -: (0 + 1) / 2 + 1 = 1
456 // 2 peers = 2 numMajority -: (2 + 1) / 2 + 1 = 2
457 // 4 peers = 3 numMajority -: (4 + 1) / 2 + 1 = 3
462 numMajority = (numPeers + self) / 2 + 1;
470 * Performs a snapshot with no capture on the replicated log.
471 * It clears the log from the supplied index or last-applied-1 which ever is minimum.
473 * @param snapshotCapturedIndex
475 protected void performSnapshotWithoutCapture(final long snapshotCapturedIndex) {
476 long actualIndex = context.getSnapshotManager().trimLog(snapshotCapturedIndex, this);
478 if(actualIndex != -1){
479 setReplicatedToAllIndex(actualIndex);
483 protected String getId(){
484 return context.getId();