package org.opendaylight.controller.cluster.raft.behaviors;
import akka.actor.ActorRef;
-import akka.event.LoggingAdapter;
-import com.google.protobuf.ByteString;
+import com.google.common.annotations.VisibleForTesting;
+import java.util.ArrayList;
import org.opendaylight.controller.cluster.raft.RaftActorContext;
import org.opendaylight.controller.cluster.raft.RaftState;
import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
import org.opendaylight.controller.cluster.raft.Snapshot;
import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
+import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot;
import org.opendaylight.controller.cluster.raft.messages.RaftRPC;
import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply;
-import java.util.ArrayList;
-
/**
* The behavior of a RaftActor in the Follower state
* <p/>
* </ul>
*/
public class Follower extends AbstractRaftActorBehavior {
- private ByteString snapshotChunksCollected = ByteString.EMPTY;
- private final LoggingAdapter LOG;
- public Follower(RaftActorContext context) {
- super(context);
- LOG = context.getLogger();
+ private SnapshotTracker snapshotTracker = null;
- scheduleElection(electionDuration());
- }
+ private final InitialSyncStatusTracker initialSyncStatusTracker;
- @Override protected RaftState handleAppendEntries(ActorRef sender,
- AppendEntries appendEntries) {
+ public Follower(RaftActorContext context) {
+ super(context, RaftState.Follower);
- if(appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
- if(LOG.isDebugEnabled()) {
- LOG.debug(appendEntries.toString());
+ initialSyncStatusTracker = new InitialSyncStatusTracker(context.getActor());
+
+ if(context.getRaftPolicy().automaticElectionsEnabled()) {
+ if (context.getPeerAddresses().isEmpty()) {
+ actor().tell(ELECTION_TIMEOUT, actor());
+ } else {
+ scheduleElection(electionDuration());
}
}
- // TODO : Refactor this method into a bunch of smaller methods
- // to make it easier to read. Before refactoring ensure tests
- // cover the code properly
+ }
- // 1. Reply false if term < currentTerm (§5.1)
- // This is handled in the appendEntries method of the base class
+ private boolean isLogEntryPresent(long index){
+ if(index == context.getReplicatedLog().getSnapshotIndex()){
+ return true;
+ }
- // If we got here then we do appear to be talking to the leader
- leaderId = appendEntries.getLeaderId();
+ ReplicatedLogEntry previousEntry = context.getReplicatedLog()
+ .get(index);
+
+ return previousEntry != null;
+
+ }
- // 2. Reply false if log doesn’t contain an entry at prevLogIndex
- // whose term matches prevLogTerm (§5.3)
+ private long getLogEntryTerm(long index){
+ if(index == context.getReplicatedLog().getSnapshotIndex()){
+ return context.getReplicatedLog().getSnapshotTerm();
+ }
ReplicatedLogEntry previousEntry = context.getReplicatedLog()
- .get(appendEntries.getPrevLogIndex());
+ .get(index);
+ if(previousEntry != null){
+ return previousEntry.getTerm();
+ }
- boolean outOfSync = true;
+ return -1;
+ }
- // First check if the logs are in sync or not
- if (lastIndex() == -1
- && appendEntries.getPrevLogIndex() != -1) {
+ private void updateInitialSyncStatus(long currentLeaderCommit, String leaderId){
+ initialSyncStatusTracker.update(leaderId, currentLeaderCommit, context.getCommitIndex());
+ }
- // The follower's log is out of sync because the leader does have
- // an entry at prevLogIndex and this follower has no entries in
- // it's log.
+ @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
+ AppendEntries appendEntries) {
- if(LOG.isDebugEnabled()) {
- LOG.debug("The followers log is empty and the senders prevLogIndex is {}",
- appendEntries.getPrevLogIndex());
- }
+ int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
+ if(LOG.isTraceEnabled()) {
+ LOG.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
+ } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
+ LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
+ }
- } else if (lastIndex() > -1
- && appendEntries.getPrevLogIndex() != -1
- && previousEntry == null) {
+ // TODO : Refactor this method into a bunch of smaller methods
+ // to make it easier to read. Before refactoring ensure tests
+ // cover the code properly
- // The follower's log is out of sync because the Leader's
- // prevLogIndex entry was not found in it's log
+ if (snapshotTracker != null) {
+ // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
+ AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
+ lastIndex(), lastTerm(), context.getPayloadVersion());
if(LOG.isDebugEnabled()) {
- LOG.debug("The log is not empty but the prevLogIndex {} was not found in it",
- appendEntries.getPrevLogIndex());
+ LOG.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
}
+ sender.tell(reply, actor());
- } else if (lastIndex() > -1
- && previousEntry != null
- && previousEntry.getTerm()!= appendEntries.getPrevLogTerm()) {
+ return this;
+ }
- // The follower's log is out of sync because the Leader's
- // prevLogIndex entry does exist in the follower's log but it has
- // a different term in it
+ // If we got here then we do appear to be talking to the leader
+ leaderId = appendEntries.getLeaderId();
- if(LOG.isDebugEnabled()) {
- LOG.debug(
- "Cannot append entries because previous entry term {} is not equal to append entries prevLogTerm {}"
- , previousEntry.getTerm()
- , appendEntries.getPrevLogTerm());
- }
- } else {
- outOfSync = false;
- }
+ setLeaderPayloadVersion(appendEntries.getPayloadVersion());
+
+ updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
+ // First check if the logs are in sync or not
+ long lastIndex = lastIndex();
- if (outOfSync) {
+ if (isOutOfSync(appendEntries)) {
// We found that the log was out of sync so just send a negative
// reply and return
- if(LOG.isDebugEnabled()) {
- LOG.debug("Follower is out-of-sync, " +
- "so sending negative reply, lastIndex():{}, lastTerm():{}",
- lastIndex(), lastTerm()
- );
- }
- sender.tell(
- new AppendEntriesReply(context.getId(), currentTerm(), false,
- lastIndex(), lastTerm()), actor()
- );
- return state();
+
+ LOG.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}",
+ logName(), lastIndex, lastTerm());
+
+ sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
+ lastTerm(), context.getPayloadVersion()), actor());
+ return this;
}
- if (appendEntries.getEntries() != null
- && appendEntries.getEntries().size() > 0) {
- if(LOG.isDebugEnabled()) {
- LOG.debug(
- "Number of entries to be appended = " + appendEntries
- .getEntries().size()
- );
- }
+ if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
+
+ LOG.debug("{}: Number of entries to be appended = {}", logName(),
+ appendEntries.getEntries().size());
// 3. If an existing entry conflicts with a new one (same index
// but different terms), delete the existing entry and all that
break;
}
- if (newEntry.getTerm() == matchEntry
- .getTerm()) {
+ if (newEntry.getTerm() == matchEntry.getTerm()) {
continue;
}
- if(LOG.isDebugEnabled()) {
- LOG.debug(
- "Removing entries from log starting at "
- + matchEntry.getIndex()
- );
- }
+ LOG.debug("{}: Removing entries from log starting at {}", logName(),
+ matchEntry.getIndex());
// Entries do not match so remove all subsequent entries
- context.getReplicatedLog()
- .removeFromAndPersist(matchEntry.getIndex());
+ context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex());
break;
}
}
- if(LOG.isDebugEnabled()) {
- context.getLogger().debug(
- "After cleanup entries to be added from = " + (addEntriesFrom
- + lastIndex())
- );
- }
+ lastIndex = lastIndex();
+ LOG.debug("{}: After cleanup entries to be added from = {}", logName(),
+ (addEntriesFrom + lastIndex));
// 4. Append any new entries not already in the log
- for (int i = addEntriesFrom;
- i < appendEntries.getEntries().size(); i++) {
-
- context.getLogger().info(
- "Append entry to log " + appendEntries.getEntries().get(
- i).getData()
- .toString()
- );
- context.getReplicatedLog()
- .appendAndPersist(appendEntries.getEntries().get(i));
- }
+ for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) {
+ ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
- if(LOG.isDebugEnabled()) {
- LOG.debug("Log size is now " + context.getReplicatedLog().size());
+ LOG.debug("{}: Append entry to log {}", logName(), entry.getData());
+
+ context.getReplicatedLog().appendAndPersist(entry);
}
- }
+ LOG.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
+ }
// 5. If leaderCommit > commitIndex, set commitIndex =
// min(leaderCommit, index of last new entry)
+ lastIndex = lastIndex();
long prevCommitIndex = context.getCommitIndex();
- context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(),
- context.getReplicatedLog().lastIndex()));
+ context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
if (prevCommitIndex != context.getCommitIndex()) {
- if(LOG.isDebugEnabled()) {
- LOG.debug("Commit index set to " + context.getCommitIndex());
- }
+ LOG.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
}
// If commitIndex > lastApplied: increment lastApplied, apply
// log[lastApplied] to state machine (§5.3)
// check if there are any entries to be applied. last-applied can be equal to last-index
if (appendEntries.getLeaderCommit() > context.getLastApplied() &&
- context.getLastApplied() < lastIndex()) {
+ context.getLastApplied() < lastIndex) {
if(LOG.isDebugEnabled()) {
- LOG.debug("applyLogToStateMachine, " +
- "appendEntries.getLeaderCommit():{}," +
- "context.getLastApplied():{}, lastIndex():{}",
- appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex()
- );
+ LOG.debug("{}: applyLogToStateMachine, " +
+ "appendEntries.getLeaderCommit(): {}," +
+ "context.getLastApplied(): {}, lastIndex(): {}", logName(),
+ appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
}
applyLogToStateMachine(appendEntries.getLeaderCommit());
}
- sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), true,
- lastIndex(), lastTerm()), actor());
+ AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
+ lastIndex, lastTerm(), context.getPayloadVersion());
+
+ if(LOG.isTraceEnabled()) {
+ LOG.trace("{}: handleAppendEntries returning : {}", logName(), reply);
+ } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
+ LOG.debug("{}: handleAppendEntries returning : {}", logName(), reply);
+ }
+
+ sender.tell(reply, actor());
- return state();
+ if (!context.getSnapshotManager().isCapturing()) {
+ super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
+ }
+
+ return this;
}
- @Override protected RaftState handleAppendEntriesReply(ActorRef sender,
- AppendEntriesReply appendEntriesReply) {
- return state();
+ private boolean isOutOfSync(AppendEntries appendEntries) {
+
+ long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
+ boolean prevEntryPresent = isLogEntryPresent(appendEntries.getPrevLogIndex());
+ long lastIndex = lastIndex();
+ int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
+ boolean outOfSync = true;
+
+ if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
+
+ // The follower's log is out of sync because the leader does have
+ // an entry at prevLogIndex and this follower has no entries in
+ // it's log.
+
+ LOG.debug("{}: The followers log is empty and the senders prevLogIndex is {}",
+ logName(), appendEntries.getPrevLogIndex());
+ } else if (lastIndex > -1 && appendEntries.getPrevLogIndex() != -1 && !prevEntryPresent) {
+
+ // The follower's log is out of sync because the Leader's
+ // prevLogIndex entry was not found in it's log
+
+ LOG.debug("{}: The log is not empty but the prevLogIndex {} was not found in it",
+ logName(), appendEntries.getPrevLogIndex());
+ } else if (lastIndex > -1 && prevEntryPresent && prevLogTerm != appendEntries.getPrevLogTerm()) {
+
+ // The follower's log is out of sync because the Leader's
+ // prevLogIndex entry does exist in the follower's log but it has
+ // a different term in it
+
+ LOG.debug(
+ "{}: Cannot append entries because previous entry term {} is not equal to append entries prevLogTerm {}",
+ logName(), prevLogTerm, appendEntries.getPrevLogTerm());
+ } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
+ && appendEntries.getReplicatedToAllIndex() != -1
+ && !isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
+ // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
+ // the previous entry in it's in-memory journal
+
+ LOG.debug(
+ "{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the in-memory journal",
+ logName(), appendEntries.getReplicatedToAllIndex());
+ } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
+ && appendEntries.getReplicatedToAllIndex() != -1 && numLogEntries > 0 &&
+ !isLogEntryPresent(appendEntries.getEntries().get(0).getIndex() - 1)){
+ LOG.debug(
+ "{}: Cannot append entries because the calculated previousIndex {} was not found in the in-memory journal",
+ logName(), appendEntries.getEntries().get(0).getIndex() - 1);
+ } else {
+ outOfSync = false;
+ }
+ return outOfSync;
}
- @Override protected RaftState handleRequestVoteReply(ActorRef sender,
- RequestVoteReply requestVoteReply) {
- return state();
+ @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
+ AppendEntriesReply appendEntriesReply) {
+ return this;
}
- @Override public RaftState state() {
- return RaftState.Follower;
+ @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
+ RequestVoteReply requestVoteReply) {
+ return this;
}
- @Override public RaftState handleMessage(ActorRef sender, Object originalMessage) {
+ @Override public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
Object message = fromSerializableMessage(originalMessage);
// set currentTerm = T, convert to follower (§5.1)
// This applies to all RPC messages and responses
if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
+ LOG.debug("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
+ logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
+
context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
}
}
if (message instanceof ElectionTimeout) {
- return RaftState.Candidate;
+ LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
+ return switchBehavior(new Candidate(context));
} else if (message instanceof InstallSnapshot) {
InstallSnapshot installSnapshot = (InstallSnapshot) message;
private void handleInstallSnapshot(ActorRef sender, InstallSnapshot installSnapshot) {
- if(LOG.isDebugEnabled()) {
- LOG.debug("InstallSnapshot received by follower " +
- "datasize:{} , Chunk:{}/{}", installSnapshot.getData().size(),
- installSnapshot.getChunkIndex(), installSnapshot.getTotalChunks()
- );
- }
+ LOG.debug("{}: InstallSnapshot received from leader {}, datasize: {} , Chunk: {}/{}",
+ logName(), installSnapshot.getLeaderId(), installSnapshot.getData().size(),
+ installSnapshot.getChunkIndex(), installSnapshot.getTotalChunks());
- try {
- if (installSnapshot.getChunkIndex() == installSnapshot.getTotalChunks()) {
- // this is the last chunk, create a snapshot object and apply
+ if(snapshotTracker == null){
+ snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks());
+ }
- snapshotChunksCollected = snapshotChunksCollected.concat(installSnapshot.getData());
- context.getLogger().debug("Last chunk received: snapshotChunksCollected.size:{}",
- snapshotChunksCollected.size());
+ updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
- Snapshot snapshot = Snapshot.create(snapshotChunksCollected.toByteArray(),
- new ArrayList<ReplicatedLogEntry>(),
- installSnapshot.getLastIncludedIndex(),
- installSnapshot.getLastIncludedTerm(),
- installSnapshot.getLastIncludedIndex(),
- installSnapshot.getLastIncludedTerm());
+ try {
+ if(snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
+ installSnapshot.getLastChunkHashCode())){
+ Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
+ new ArrayList<ReplicatedLogEntry>(),
+ installSnapshot.getLastIncludedIndex(),
+ installSnapshot.getLastIncludedTerm(),
+ installSnapshot.getLastIncludedIndex(),
+ installSnapshot.getLastIncludedTerm());
actor().tell(new ApplySnapshot(snapshot), actor());
- } else {
- // we have more to go
- snapshotChunksCollected = snapshotChunksCollected.concat(installSnapshot.getData());
+ snapshotTracker = null;
- if(LOG.isDebugEnabled()) {
- LOG.debug("Chunk={},snapshotChunksCollected.size:{}",
- installSnapshot.getChunkIndex(), snapshotChunksCollected.size());
- }
}
- sender.tell(new InstallSnapshotReply(
- currentTerm(), context.getId(), installSnapshot.getChunkIndex(),
- true), actor());
+ InstallSnapshotReply reply = new InstallSnapshotReply(
+ currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
+
+ LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
+
+ sender.tell(reply, actor());
+
+ } catch (SnapshotTracker.InvalidChunkException e) {
+ LOG.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
+
+ sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
+ -1, false), actor());
+ snapshotTracker = null;
+
+ } catch (Exception e){
+ LOG.error("{}: Exception in InstallSnapshot of follower", logName(), e);
- } catch (Exception e) {
- context.getLogger().error("Exception in InstallSnapshot of follower", e);
//send reply with success as false. The chunk will be sent again on failure
sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
- installSnapshot.getChunkIndex(), false), actor());
+ installSnapshot.getChunkIndex(), false), actor());
+
}
}
- @Override public void close() throws Exception {
+ @Override
+ public void close() throws Exception {
stopElection();
}
+
+ @VisibleForTesting
+ SnapshotTracker getSnapshotTracker(){
+ return snapshotTracker;
+ }
+
+ private class InitialSyncStatusTracker {
+
+ private static final long INVALID_LOG_INDEX = -2L;
+ private long initialLeaderCommit = INVALID_LOG_INDEX;
+ private boolean initialSyncUpDone = false;
+ private String syncedLeaderId = null;
+ private final ActorRef actor;
+
+ public InitialSyncStatusTracker(ActorRef actor) {
+ this.actor = actor;
+ }
+
+ public void update(String leaderId, long leaderCommit, long commitIndex){
+
+ if(!leaderId.equals(syncedLeaderId)){
+ initialSyncUpDone = false;
+ initialLeaderCommit = INVALID_LOG_INDEX;
+ syncedLeaderId = leaderId;
+ }
+
+ if(!initialSyncUpDone){
+ if(initialLeaderCommit == INVALID_LOG_INDEX){
+ actor.tell(new FollowerInitialSyncUpStatus(false, getId()), ActorRef.noSender());
+ initialLeaderCommit = leaderCommit;
+ } else if(commitIndex >= initialLeaderCommit){
+ actor.tell(new FollowerInitialSyncUpStatus(true, getId()), ActorRef.noSender());
+ initialSyncUpDone = true;
+ }
+ }
+ }
+ }
}