X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?p=controller.git;a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-akka-raft%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fraft%2Fbehaviors%2FFollower.java;h=4fa0e161db16ffc2ada57fd52cc6af22ce8160fd;hp=7e8a7725910d599f7cd56ef40daaa06de2f8bece;hb=1b0f84c4957e464bad6f7cb7350a8171c3d1621b;hpb=fa96da71c5ab10973a9f93c2e8b35494b96fd7ed diff --git a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Follower.java b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Follower.java index 7e8a772591..4fa0e161db 100644 --- a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Follower.java +++ b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/Follower.java @@ -18,15 +18,18 @@ import akka.cluster.MemberStatus; import akka.japi.Procedure; import com.google.common.annotations.VisibleForTesting; import com.google.common.base.Stopwatch; +import java.io.IOException; import java.util.ArrayList; +import java.util.List; import java.util.Optional; import java.util.Set; import java.util.concurrent.TimeUnit; +import java.util.concurrent.atomic.AtomicBoolean; import javax.annotation.Nullable; +import org.opendaylight.controller.cluster.messaging.MessageAssembler; 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.TimeoutNow; @@ -38,6 +41,7 @@ import org.opendaylight.controller.cluster.raft.messages.RaftRPC; import org.opendaylight.controller.cluster.raft.messages.RequestVote; import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply; import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload; +import org.opendaylight.controller.cluster.raft.persisted.Snapshot; /** * The behavior of a RaftActor in the Follower raft state. @@ -49,30 +53,33 @@ import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPay * */ public class Follower extends AbstractRaftActorBehavior { - private static final int SYNC_THRESHOLD = 10; - private static final long MAX_ELECTION_TIMEOUT_FACTOR = 18; private final SyncStatusTracker initialSyncStatusTracker; - private final Procedure appendAndPersistCallback = - logEntry -> context.getReplicatedLog().captureSnapshotIfReady(logEntry); + private final MessageAssembler appendEntriesMessageAssembler; private final Stopwatch lastLeaderMessageTimer = Stopwatch.createStarted(); private SnapshotTracker snapshotTracker = null; private String leaderId; private short leaderPayloadVersion; - public Follower(RaftActorContext context) { + public Follower(final RaftActorContext context) { this(context, null, (short)-1); } - public Follower(RaftActorContext context, String initialLeaderId, short initialLeaderPayloadVersion) { + public Follower(final RaftActorContext context, final String initialLeaderId, + final short initialLeaderPayloadVersion) { super(context, RaftState.Follower); this.leaderId = initialLeaderId; this.leaderPayloadVersion = initialLeaderPayloadVersion; - initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), SYNC_THRESHOLD); + initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), context.getConfigParams() + .getSyncIndexThreshold()); + + appendEntriesMessageAssembler = MessageAssembler.builder().logContext(logName()) + .fileBackedStreamFactory(context.getFileBackedOutputStreamFactory()) + .assembledMessageCallback((message, sender) -> handleMessage(sender, message)).build(); if (context.getPeerIds().isEmpty() && getLeaderId() == null) { actor().tell(TimeoutNow.INSTANCE, actor()); @@ -97,7 +104,7 @@ public class Follower extends AbstractRaftActorBehavior { } @VisibleForTesting - protected final void setLeaderPayloadVersion(short leaderPayloadVersion) { + protected final void setLeaderPayloadVersion(final short leaderPayloadVersion) { this.leaderPayloadVersion = leaderPayloadVersion; } @@ -109,7 +116,7 @@ public class Follower extends AbstractRaftActorBehavior { lastLeaderMessageTimer.start(); } - private boolean isLogEntryPresent(long index) { + private boolean isLogEntryPresent(final long index) { if (context.getReplicatedLog().isInSnapshot(index)) { return true; } @@ -119,28 +126,23 @@ public class Follower extends AbstractRaftActorBehavior { } - private void updateInitialSyncStatus(long currentLeaderCommit, String newLeaderId) { + private void updateInitialSyncStatus(final long currentLeaderCommit, final String newLeaderId) { initialSyncStatusTracker.update(newLeaderId, currentLeaderCommit, context.getCommitIndex()); } @Override - protected RaftActorBehavior handleAppendEntries(ActorRef sender, AppendEntries appendEntries) { - - int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0; + protected RaftActorBehavior handleAppendEntries(final ActorRef sender, final AppendEntries appendEntries) { + int numLogEntries = appendEntries.getEntries().size(); if (log.isTraceEnabled()) { log.trace("{}: handleAppendEntries: {}", logName(), appendEntries); } else if (log.isDebugEnabled() && numLogEntries > 0) { log.debug("{}: handleAppendEntries: {}", logName(), appendEntries); } - // TODO : Refactor this method into a bunch of smaller methods - // to make it easier to read. Before refactoring ensure tests - // cover the code properly - if (snapshotTracker != null && !snapshotTracker.getLeaderId().equals(appendEntries.getLeaderId())) { log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the " + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId()); - snapshotTracker = null; + closeSnapshotTracker(); } if (snapshotTracker != null || context.getSnapshotManager().isApplying()) { @@ -158,217 +160,268 @@ public class Follower extends AbstractRaftActorBehavior { leaderId = appendEntries.getLeaderId(); leaderPayloadVersion = appendEntries.getPayloadVersion(); - updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId()); // First check if the logs are in sync or not + if (isOutOfSync(appendEntries, sender)) { + updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId()); + return this; + } + + if (!processNewEntries(appendEntries, sender)) { + updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId()); + return this; + } + long lastIndex = lastIndex(); + long prevCommitIndex = context.getCommitIndex(); - if (isOutOfSync(appendEntries)) { - // We found that the log was out of sync so just send a negative - // reply and return + // If leaderCommit > commitIndex, set commitIndex = min(leaderCommit, index of last new entry) + if (appendEntries.getLeaderCommit() > prevCommitIndex) { + context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex)); + } - log.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}", - logName(), lastIndex, lastTerm()); + if (prevCommitIndex != context.getCommitIndex()) { + log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex()); + } - sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex, - lastTerm(), context.getPayloadVersion()), actor()); - return this; + 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); } - if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) { + // Reply to the leader before applying any previous state so as not to hold up leader consensus. + sender.tell(reply, actor()); - log.debug("{}: Number of entries to be appended = {}", logName(), - appendEntries.getEntries().size()); + updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId()); - // 3. If an existing entry conflicts with a new one (same index - // but different terms), delete the existing entry and all that - // follow it (§5.3) - int addEntriesFrom = 0; - if (context.getReplicatedLog().size() > 0) { + // If leaderCommit > lastApplied, increment lastApplied and apply log[lastApplied] to state machine (§5.3). + // lastApplied can be equal to lastIndex. + if (appendEntries.getLeaderCommit() > context.getLastApplied() && context.getLastApplied() < lastIndex) { + if (log.isDebugEnabled()) { + log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {}, " + + "context.getLastApplied(): {}, lastIndex(): {}", logName(), + appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex); + } - // Find the entry up until the one that is not in the follower's log - for (int i = 0;i < appendEntries.getEntries().size(); i++, addEntriesFrom++) { - ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i); + applyLogToStateMachine(appendEntries.getLeaderCommit()); + } - if (!isLogEntryPresent(matchEntry.getIndex())) { - // newEntry not found in the log - break; - } + if (!context.getSnapshotManager().isCapturing()) { + super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex()); + } - long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex()); + appendEntriesMessageAssembler.checkExpiredAssembledMessageState(); - log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry, - existingEntryTerm); + return this; + } - // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know - // what the term was so we'll assume it matches. - if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) { - continue; - } + private boolean processNewEntries(final AppendEntries appendEntries, final ActorRef sender) { + int numLogEntries = appendEntries.getEntries().size(); + if (numLogEntries == 0) { + return true; + } - if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) { + log.debug("{}: Number of entries to be appended = {}", logName(), numLogEntries); - log.debug("{}: Removing entries from log starting at {}", logName(), - matchEntry.getIndex()); + long lastIndex = lastIndex(); + int addEntriesFrom = 0; + + // First check for conflicting entries. If an existing entry conflicts with a new one (same index but different + // term), delete the existing entry and all that follow it (§5.3) + if (context.getReplicatedLog().size() > 0) { + // Find the entry up until the one that is not in the follower's log + for (int i = 0;i < numLogEntries; i++, addEntriesFrom++) { + ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i); + + if (!isLogEntryPresent(matchEntry.getIndex())) { + // newEntry not found in the log + break; + } - // Entries do not match so remove all subsequent entries - if (!context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) { - // Could not remove the entries - this means the matchEntry index must be in the - // snapshot and not the log. In this case the prior entries are part of the state - // so we must send back a reply to force a snapshot to completely re-sync the - // follower's log and state. + long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex()); - log.debug("{}: Could not remove entries - sending reply to force snapshot", logName()); - sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex, - lastTerm(), context.getPayloadVersion(), true), actor()); - return this; - } + log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry, + existingEntryTerm); + + // existingEntryTerm == -1 means it's in the snapshot and not in the log. We don't know + // what the term was so we'll assume it matches. + if (existingEntryTerm == -1 || existingEntryTerm == matchEntry.getTerm()) { + continue; + } - break; - } else { + if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) { + log.info("{}: Removing entries from log starting at {}, commitIndex: {}, lastApplied: {}", + logName(), matchEntry.getIndex(), context.getCommitIndex(), context.getLastApplied()); + + // Entries do not match so remove all subsequent entries but only if the existing entries haven't + // been applied to the state yet. + if (matchEntry.getIndex() <= context.getLastApplied() + || !context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex())) { + // Could not remove the entries - this means the matchEntry index must be in the + // snapshot and not the log. In this case the prior entries are part of the state + // so we must send back a reply to force a snapshot to completely re-sync the + // follower's log and state. + + log.info("{}: Could not remove entries - sending reply to force snapshot", logName()); sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex, lastTerm(), context.getPayloadVersion(), true), actor()); - return this; + return false; } + + break; + } else { + sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex, + lastTerm(), context.getPayloadVersion(), true), actor()); + return false; } } + } - lastIndex = lastIndex(); - log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(), - lastIndex, addEntriesFrom); + lastIndex = lastIndex(); + log.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(), lastIndex, + addEntriesFrom); + + // When persistence successfully completes for each new log entry appended, we need to determine if we + // should capture a snapshot to compact the persisted log. shouldCaptureSnapshot tracks whether or not + // one of the log entries has exceeded the log size threshold whereby a snapshot should be taken. However + // we don't initiate the snapshot at that log entry but rather after the last log entry has been persisted. + // This is done because subsequent log entries after the one that tripped the threshold may have been + // applied to the state already, as the persistence callback occurs async, and we want those entries + // purged from the persisted log as well. + final AtomicBoolean shouldCaptureSnapshot = new AtomicBoolean(false); + final Procedure appendAndPersistCallback = logEntry -> { + final List entries = appendEntries.getEntries(); + final ReplicatedLogEntry lastEntryToAppend = entries.get(entries.size() - 1); + if (shouldCaptureSnapshot.get() && logEntry == lastEntryToAppend) { + context.getSnapshotManager().capture(context.getReplicatedLog().last(), getReplicatedToAllIndex()); + } + }; - // 4. Append any new entries not already in the log - for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) { - ReplicatedLogEntry entry = appendEntries.getEntries().get(i); + // Append any new entries not already in the log + for (int i = addEntriesFrom; i < numLogEntries; i++) { + ReplicatedLogEntry entry = appendEntries.getEntries().get(i); - log.debug("{}: Append entry to log {}", logName(), entry.getData()); + log.debug("{}: Append entry to log {}", logName(), entry.getData()); - context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false); + context.getReplicatedLog().appendAndPersist(entry, appendAndPersistCallback, false); - if (entry.getData() instanceof ServerConfigurationPayload) { - context.updatePeerIds((ServerConfigurationPayload)entry.getData()); - } - } + shouldCaptureSnapshot.compareAndSet(false, + context.getReplicatedLog().shouldCaptureSnapshot(entry.getIndex())); - log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size()); + if (entry.getData() instanceof ServerConfigurationPayload) { + context.updatePeerIds((ServerConfigurationPayload)entry.getData()); + } } - // 5. If leaderCommit > commitIndex, set commitIndex = - // min(leaderCommit, index of last new entry) + log.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size()); - lastIndex = lastIndex(); - long prevCommitIndex = context.getCommitIndex(); + return true; + } - if (appendEntries.getLeaderCommit() > prevCommitIndex) { - context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex)); - } + private boolean isOutOfSync(final AppendEntries appendEntries, final ActorRef sender) { - if (prevCommitIndex != context.getCommitIndex()) { - log.debug("{}: Commit index set to {}", logName(), context.getCommitIndex()); - } + final long lastIndex = lastIndex(); + if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) { - // 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) { - if (log.isDebugEnabled()) { - log.debug("{}: applyLogToStateMachine, appendEntries.getLeaderCommit(): {}," - + "context.getLastApplied(): {}, lastIndex(): {}", logName(), - appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex); - } + // 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. - applyLogToStateMachine(appendEntries.getLeaderCommit()); + log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(), + appendEntries.getPrevLogIndex()); + + sendOutOfSyncAppendEntriesReply(sender, false); + return true; } - AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true, - lastIndex, lastTerm(), context.getPayloadVersion()); + if (lastIndex > -1) { + if (isLogEntryPresent(appendEntries.getPrevLogIndex())) { + final long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex()); + if (prevLogTerm != appendEntries.getPrevLogTerm()) { - if (log.isTraceEnabled()) { - log.trace("{}: handleAppendEntries returning : {}", logName(), reply); - } else if (log.isDebugEnabled() && numLogEntries > 0) { - log.debug("{}: handleAppendEntries returning : {}", logName(), reply); - } + // 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 - sender.tell(reply, actor()); + log.info("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append " + + "entries prevLogTerm {} - lastIndex: {}, snapshotIndex: {}", logName(), + appendEntries.getPrevLogIndex(), prevLogTerm, appendEntries.getPrevLogTerm(), lastIndex, + context.getReplicatedLog().getSnapshotIndex()); - if (!context.getSnapshotManager().isCapturing()) { - super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex()); + sendOutOfSyncAppendEntriesReply(sender, false); + return true; + } + } else if (appendEntries.getPrevLogIndex() != -1) { + + // The follower's log is out of sync because the Leader's prevLogIndex entry was not found in it's log + + log.info("{}: The log is not empty but the prevLogIndex {} was not found in it - lastIndex: {}, " + + "snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex, + context.getReplicatedLog().getSnapshotIndex()); + + sendOutOfSyncAppendEntriesReply(sender, false); + return true; + } } - return this; - } + if (appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1 + && appendEntries.getReplicatedToAllIndex() != -1) { + if (!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 - private boolean isOutOfSync(AppendEntries appendEntries) { + log.info("{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the " + + "in-memory journal", logName(), appendEntries.getReplicatedToAllIndex()); - 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; + sendOutOfSyncAppendEntriesReply(sender, false); + return true; + } - if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) { + final List entries = appendEntries.getEntries(); + if (entries.size() > 0 && !isLogEntryPresent(entries.get(0).getIndex() - 1)) { + log.info("{}: Cannot append entries because the calculated previousIndex {} was not found in the " + + "in-memory journal", logName(), entries.get(0).getIndex() - 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 - " - + "lastIndex: {}, snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex, - context.getReplicatedLog().getSnapshotIndex()); - } 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("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append entries" - + "prevLogTerm {} - lastIndex: {}, snapshotIndex: {}", logName(), appendEntries.getPrevLogIndex(), - prevLogTerm, appendEntries.getPrevLogTerm(), lastIndex, - context.getReplicatedLog().getSnapshotIndex()); - } 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; + sendOutOfSyncAppendEntriesReply(sender, false); + return true; + } } - return outOfSync; + + return false; + } + + private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, boolean forceInstallSnapshot) { + // We found that the log was out of sync so just send a negative reply. + final AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex(), + lastTerm(), context.getPayloadVersion(), forceInstallSnapshot); + + log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply); + sender.tell(reply, actor()); } @Override - protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender, - AppendEntriesReply appendEntriesReply) { + protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender, + final AppendEntriesReply appendEntriesReply) { return this; } @Override - protected RaftActorBehavior handleRequestVoteReply(ActorRef sender, - RequestVoteReply requestVoteReply) { + protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender, + final RequestVoteReply requestVoteReply) { return this; } @Override - public RaftActorBehavior handleMessage(ActorRef sender, Object message) { + public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) { if (message instanceof ElectionTimeout || message instanceof TimeoutNow) { return handleElectionTimeout(message); } + if (appendEntriesMessageAssembler.handleMessage(message, actor())) { + return this; + } + if (!(message instanceof RaftRPC)) { // The rest of the processing requires the message to be a RaftRPC return null; @@ -379,7 +432,7 @@ public class Follower extends AbstractRaftActorBehavior { // 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", + log.info("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term", logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm()); context.getTermInformation().updateAndPersist(rpc.getTerm(), null); @@ -400,7 +453,7 @@ public class Follower extends AbstractRaftActorBehavior { return super.handleMessage(sender, rpc); } - private RaftActorBehavior handleElectionTimeout(Object message) { + private RaftActorBehavior handleElectionTimeout(final Object message) { // If the message is ElectionTimeout, verify we haven't actually seen a message from the leader // during the election timeout interval. It may that the election timer expired b/c this actor // was busy and messages got delayed, in which case leader messages would be backed up in the @@ -493,14 +546,15 @@ public class Follower extends AbstractRaftActorBehavior { return false; } - private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) { + private void handleInstallSnapshot(final ActorRef sender, final InstallSnapshot installSnapshot) { log.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot); leaderId = installSnapshot.getLeaderId(); if (snapshotTracker == null) { - snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId()); + snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId(), + context); } updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId()); @@ -511,7 +565,11 @@ public class Follower extends AbstractRaftActorBehavior { if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(), installSnapshot.getLastChunkHashCode())) { - Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(), + + log.info("{}: Snapshot installed from leader: {}", logName(), installSnapshot.getLeaderId()); + + Snapshot snapshot = Snapshot.create( + context.getSnapshotManager().convertSnapshot(snapshotTracker.getSnapshotBytes()), new ArrayList<>(), installSnapshot.getLastIncludedIndex(), installSnapshot.getLastIncludedTerm(), @@ -537,25 +595,34 @@ public class Follower extends AbstractRaftActorBehavior { actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor()); - snapshotTracker = null; + closeSnapshotTracker(); } else { log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply); sender.tell(reply, actor()); } - } catch (SnapshotTracker.InvalidChunkException e) { + } catch (IOException e) { log.debug("{}: Exception in InstallSnapshot of follower", logName(), e); sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor()); - snapshotTracker = null; + closeSnapshotTracker(); + } + } + + private void closeSnapshotTracker() { + if (snapshotTracker != null) { + snapshotTracker.close(); + snapshotTracker = null; } } @Override public void close() { + closeSnapshotTracker(); stopElection(); + appendEntriesMessageAssembler.close(); } @VisibleForTesting