X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-akka-raft%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fraft%2Fbehaviors%2FFollower.java;h=fa28cbf718748e5800da017d5eed69726fcbd1c9;hb=b8de4ba085d8d165a7bb48511dc374ed278fa8bd;hp=cb141f9f1e5b28f3c7a3f7428ad827d5706881d0;hpb=b2cb02f62ab7c7599e8d94fe92d1ce63e17d599b;p=controller.git
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 cb141f9f1e..fa28cbf718 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
@@ -5,7 +5,6 @@
* terms of the Eclipse Public License v1.0 which accompanies this distribution,
* and is available at http://www.eclipse.org/legal/epl-v10.html
*/
-
package org.opendaylight.controller.cluster.raft.behaviors;
import akka.actor.ActorRef;
@@ -15,19 +14,25 @@ import akka.cluster.Cluster;
import akka.cluster.ClusterEvent.CurrentClusterState;
import akka.cluster.Member;
import akka.cluster.MemberStatus;
-import akka.japi.Procedure;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Stopwatch;
+import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
+import java.io.IOException;
import java.util.ArrayList;
+import java.util.HashSet;
+import java.util.List;
import java.util.Optional;
import java.util.Set;
import java.util.concurrent.TimeUnit;
-import javax.annotation.Nullable;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.function.Consumer;
+import org.eclipse.jdt.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.ApplyState;
import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
import org.opendaylight.controller.cluster.raft.base.messages.TimeoutNow;
import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
@@ -38,10 +43,10 @@ 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 state
- *
+ * The behavior of a RaftActor in the Follower raft state.
*
* - Respond to RPCs from candidates and leaders
*
- If election timeout elapses without receiving AppendEntries
@@ -49,31 +54,37 @@ import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPay
* convert to candidate
*
*/
+// Non-final for testing
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) {
+ @SuppressFBWarnings(value = "MC_OVERRIDABLE_METHOD_CALL_IN_CONSTRUCTOR",
+ justification = "electionDuration() is not final for Candidate override")
+ public Follower(final RaftActorContext context, final String initialLeaderId,
+ final short initialLeaderPayloadVersion) {
super(context, RaftState.Follower);
- this.leaderId = initialLeaderId;
- this.leaderPayloadVersion = initialLeaderPayloadVersion;
+ leaderId = initialLeaderId;
+ 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());
@@ -88,7 +99,7 @@ public class Follower extends AbstractRaftActorBehavior {
}
@VisibleForTesting
- protected final void setLeaderId(@Nullable final String leaderId) {
+ protected final void setLeaderId(final @Nullable String leaderId) {
this.leaderId = leaderId;
}
@@ -98,7 +109,7 @@ public class Follower extends AbstractRaftActorBehavior {
}
@VisibleForTesting
- protected final void setLeaderPayloadVersion(short leaderPayloadVersion) {
+ protected final void setLeaderPayloadVersion(final short leaderPayloadVersion) {
this.leaderPayloadVersion = leaderPayloadVersion;
}
@@ -110,8 +121,8 @@ public class Follower extends AbstractRaftActorBehavior {
lastLeaderMessageTimer.start();
}
- private boolean isLogEntryPresent(long index){
- if(context.getReplicatedLog().isInSnapshot(index)) {
+ private boolean isLogEntryPresent(final long index) {
+ if (context.getReplicatedLog().isInSnapshot(index)) {
return true;
}
@@ -120,38 +131,32 @@ public class Follower extends AbstractRaftActorBehavior {
}
- private void updateInitialSyncStatus(long currentLeaderCommit, String leaderId){
- initialSyncStatusTracker.update(leaderId, currentLeaderCommit, context.getCommitIndex());
+ 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;
- if(LOG.isTraceEnabled()) {
- LOG.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
- } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
- LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
+ 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;
+ log.debug("{}: snapshot install is in progress but the prior snapshot leaderId {} does not match the "
+ + "AppendEntries leaderId {}", logName(), snapshotTracker.getLeaderId(), appendEntries.getLeaderId());
+ closeSnapshotTracker();
}
if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
// 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());
+ lastIndex(), lastTerm(), context.getPayloadVersion(), false, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion());
- if(LOG.isDebugEnabled()) {
- LOG.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
- }
+ log.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
sender.tell(reply, actor());
return this;
@@ -161,242 +166,319 @@ public class Follower extends AbstractRaftActorBehavior {
leaderId = appendEntries.getLeaderId();
leaderPayloadVersion = appendEntries.getPayloadVersion();
- updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
+ if (appendEntries.getLeaderAddress().isPresent()) {
+ final String address = appendEntries.getLeaderAddress().get();
+ log.debug("New leader address: {}", address);
+
+ context.setPeerAddress(leaderId, address);
+ context.getConfigParams().getPeerAddressResolver().setResolved(leaderId, address);
+ }
+
// 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(), false, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion());
+
+ 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;
+ }
+
+ long existingEntryTerm = getLogEntryTerm(matchEntry.getIndex());
+
+ log.debug("{}: matchEntry {} is present: existingEntryTerm: {}", logName(), matchEntry,
+ existingEntryTerm);
- // 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.
+ // 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;
+ }
+
+ if (!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
+ log.info("{}: Removing entries from log starting at {}, commitIndex: {}, lastApplied: {}",
+ logName(), matchEntry.getIndex(), context.getCommitIndex(), context.getLastApplied());
- 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;
- }
+ // 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.
- break;
- } else {
+ 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;
+ lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion()), actor());
+ return false;
}
+
+ break;
+ } else {
+ sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
+ lastTerm(), context.getPayloadVersion(), true, needsLeaderAddress(),
+ appendEntries.getLeaderRaftVersion()), actor());
+ return false;
}
}
+ }
+
+ 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 Consumer 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());
+ }
+ };
- lastIndex = lastIndex();
- LOG.debug("{}: After cleanup, lastIndex: {}, entries to be added from: {}", logName(),
- lastIndex, addEntriesFrom);
+ // Append any new entries not already in the log
+ for (int i = addEntriesFrom; i < numLogEntries; i++) {
+ ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
- // 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);
+ 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);
+ shouldCaptureSnapshot.compareAndSet(false,
+ context.getReplicatedLog().shouldCaptureSnapshot(entry.getIndex()));
- if(entry.getData() instanceof ServerConfigurationPayload) {
- context.updatePeerIds((ServerConfigurationPayload)entry.getData());
- }
+ if (entry.getData() instanceof ServerConfigurationPayload serverConfiguration) {
+ context.updatePeerIds(serverConfiguration);
}
-
- LOG.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
}
- // 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());
- }
-
- // 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);
- }
+ final long lastIndex = lastIndex();
+ if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
- applyLogToStateMachine(appendEntries.getLeaderCommit());
- }
+ // 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.
- AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
- lastIndex, lastTerm(), context.getPayloadVersion());
+ log.info("{}: The followers log is empty and the senders prevLogIndex is {}", logName(),
+ appendEntries.getPrevLogIndex());
- if(LOG.isTraceEnabled()) {
- LOG.trace("{}: handleAppendEntries returning : {}", logName(), reply);
- } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
- LOG.debug("{}: handleAppendEntries returning : {}", logName(), reply);
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
+ return true;
}
- sender.tell(reply, actor());
+ if (lastIndex > -1) {
+ if (isLogEntryPresent(appendEntries.getPrevLogIndex())) {
+ final long leadersPrevLogTermInFollowersLogOrSnapshot =
+ getLogEntryOrSnapshotTerm(appendEntries.getPrevLogIndex());
+ if (leadersPrevLogTermInFollowersLogOrSnapshot != appendEntries.getPrevLogTerm()) {
- if (!context.getSnapshotManager().isCapturing()) {
- super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
+ // The follower's log is out of sync because the Leader's prevLogIndex entry does exist
+ // in the follower's log or snapshot but it has a different term.
+
+ log.info("{}: The prevLogIndex {} was found in the log but the term {} is not equal to the append "
+ + "entries prevLogTerm {} - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
+ appendEntries.getPrevLogIndex(), leadersPrevLogTermInFollowersLogOrSnapshot,
+ appendEntries.getPrevLogTerm(), lastIndex, context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm());
+
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
+ 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: {}, snapshotTerm: {}", logName(), appendEntries.getPrevLogIndex(), lastIndex,
+ context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
+
+ sendOutOfSyncAppendEntriesReply(sender, false, appendEntries.getLeaderRaftVersion());
+ 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 - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
+ appendEntries.getReplicatedToAllIndex(), lastIndex,
+ context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm());
- 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, appendEntries.getLeaderRaftVersion());
+ 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 - lastIndex: {}, snapshotIndex: {}, snapshotTerm: {}", logName(),
+ entries.get(0).getIndex() - 1, lastIndex, context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm());
- // 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, appendEntries.getLeaderRaftVersion());
+ return true;
+ }
}
- return outOfSync;
+
+ return false;
+ }
+
+ private void sendOutOfSyncAppendEntriesReply(final ActorRef sender, final boolean forceInstallSnapshot,
+ final short leaderRaftVersion) {
+ // 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, needsLeaderAddress(),
+ leaderRaftVersion);
+
+ log.info("{}: Follower is out-of-sync so sending negative reply: {}", logName(), reply);
+ sender.tell(reply, actor());
+ }
+
+ private boolean needsLeaderAddress() {
+ return context.getPeerAddress(leaderId) == null;
}
@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) {
+ final ApplyState getApplyStateFor(final ReplicatedLogEntry entry) {
+ return new ApplyState(null, null, entry);
+ }
+
+ @Override
+ public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
if (message instanceof ElectionTimeout || message instanceof TimeoutNow) {
return handleElectionTimeout(message);
}
- if (!(message instanceof RaftRPC)) {
+ if (appendEntriesMessageAssembler.handleMessage(message, actor())) {
+ return this;
+ }
+
+ if (!(message instanceof RaftRPC rpc)) {
// The rest of the processing requires the message to be a RaftRPC
return null;
}
- final RaftRPC rpc = (RaftRPC) message;
// If RPC request or response contains term T > currentTerm:
// 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",
+ if (rpc.getTerm() > context.getTermInformation().getCurrentTerm() && shouldUpdateTerm(rpc)) {
+ 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);
}
- if (rpc instanceof InstallSnapshot) {
- handleInstallSnapshot(sender, (InstallSnapshot) rpc);
+ if (rpc instanceof InstallSnapshot installSnapshot) {
+ handleInstallSnapshot(sender, installSnapshot);
restartLastLeaderMessageTimer();
scheduleElection(electionDuration());
return this;
}
- if (!(rpc instanceof RequestVote) || canGrantVote((RequestVote) rpc)) {
+ if (!(rpc instanceof RequestVote requestVote) || canGrantVote(requestVote)) {
restartLastLeaderMessageTimer();
scheduleElection(electionDuration());
}
@@ -404,7 +486,7 @@ 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
@@ -412,33 +494,37 @@ public class Follower extends AbstractRaftActorBehavior {
// lastLeaderMessageTimer.
long lastLeaderMessageInterval = lastLeaderMessageTimer.elapsed(TimeUnit.MILLISECONDS);
long electionTimeoutInMillis = context.getConfigParams().getElectionTimeOutInterval().toMillis();
- boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning() ||
- lastLeaderMessageInterval >= electionTimeoutInMillis;
+ boolean noLeaderMessageReceived = !lastLeaderMessageTimer.isRunning()
+ || lastLeaderMessageInterval >= electionTimeoutInMillis;
- if(canStartElection()) {
- if(message instanceof TimeoutNow) {
- LOG.debug("{}: Received TimeoutNow - switching to Candidate", logName());
+ if (canStartElection()) {
+ if (message instanceof TimeoutNow) {
+ log.debug("{}: Received TimeoutNow - switching to Candidate", logName());
return internalSwitchBehavior(RaftState.Candidate);
- } else if(noLeaderMessageReceived) {
+ } else if (noLeaderMessageReceived) {
// Check the cluster state to see if the leader is known to be up before we go to Candidate.
// However if we haven't heard from the leader in a long time even though the cluster state
// indicates it's up then something is wrong - leader might be stuck indefinitely - so switch
// to Candidate,
long maxElectionTimeout = electionTimeoutInMillis * MAX_ELECTION_TIMEOUT_FACTOR;
- if(isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
- LOG.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
+ if (isLeaderAvailabilityKnown() && lastLeaderMessageInterval < maxElectionTimeout) {
+ log.debug("{}: Received ElectionTimeout but leader appears to be available", logName());
+ scheduleElection(electionDuration());
+ } else if (isThisFollowerIsolated()) {
+ log.debug("{}: this follower is isolated. Do not switch to Candidate for now.", logName());
+ setLeaderId(null);
scheduleElection(electionDuration());
} else {
- LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
+ log.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
return internalSwitchBehavior(RaftState.Candidate);
}
} else {
- LOG.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
+ log.debug("{}: Received ElectionTimeout but lastLeaderMessageInterval {} < election timeout {}",
logName(), lastLeaderMessageInterval, context.getConfigParams().getElectionTimeOutInterval());
scheduleElection(electionDuration());
}
- } else if(message instanceof ElectionTimeout) {
- if(noLeaderMessageReceived) {
+ } else if (message instanceof ElectionTimeout) {
+ if (noLeaderMessageReceived) {
setLeaderId(null);
}
@@ -449,17 +535,17 @@ public class Follower extends AbstractRaftActorBehavior {
}
private boolean isLeaderAvailabilityKnown() {
- if(leaderId == null) {
+ if (leaderId == null) {
return false;
}
Optional cluster = context.getCluster();
- if(!cluster.isPresent()) {
+ if (!cluster.isPresent()) {
return false;
}
ActorSelection leaderActor = context.getPeerActorSelection(leaderId);
- if(leaderActor == null) {
+ if (leaderActor == null) {
return false;
}
@@ -468,43 +554,78 @@ public class Follower extends AbstractRaftActorBehavior {
CurrentClusterState state = cluster.get().state();
Set unreachable = state.getUnreachable();
- LOG.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
+ log.debug("{}: Checking for leader {} in the cluster unreachable set {}", logName(), leaderAddress,
unreachable);
- for(Member m: unreachable) {
- if(leaderAddress.equals(m.address())) {
- LOG.info("{}: Leader {} is unreachable", logName(), leaderAddress);
+ for (Member m: unreachable) {
+ if (leaderAddress.equals(m.address())) {
+ log.info("{}: Leader {} is unreachable", logName(), leaderAddress);
return false;
}
}
- for(Member m: state.getMembers()) {
- if(leaderAddress.equals(m.address())) {
- if(m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
- LOG.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
+ for (Member m: state.getMembers()) {
+ if (leaderAddress.equals(m.address())) {
+ if (m.status() == MemberStatus.up() || m.status() == MemberStatus.weaklyUp()) {
+ log.debug("{}: Leader {} cluster status is {} - leader is available", logName(),
leaderAddress, m.status());
return true;
} else {
- LOG.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
+ log.debug("{}: Leader {} cluster status is {} - leader is unavailable", logName(),
leaderAddress, m.status());
return false;
}
}
}
- LOG.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
+ log.debug("{}: Leader {} not found in the cluster member set", logName(), leaderAddress);
return false;
}
- private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) {
+ private boolean isThisFollowerIsolated() {
+ final Optional maybeCluster = context.getCluster();
+ if (!maybeCluster.isPresent()) {
+ return false;
+ }
+
+ final Cluster cluster = maybeCluster.get();
+ final Member selfMember = cluster.selfMember();
+
+ final CurrentClusterState state = cluster.state();
+ final Set unreachable = state.getUnreachable();
+ final Iterable members = state.getMembers();
+
+ log.debug("{}: Checking if this node is isolated in the cluster unreachable set {},"
+ + "all members {} self member: {}", logName(), unreachable, members, selfMember);
+
+ // no unreachable peers means we cannot be isolated
+ if (unreachable.size() == 0) {
+ return false;
+ }
- LOG.debug("{}: handleInstallSnapshot: {}", logName(), installSnapshot);
+ final Set membersToCheck = new HashSet<>();
+ members.forEach(membersToCheck::add);
+
+ membersToCheck.removeAll(unreachable);
+
+ // check if the only member not unreachable is us
+ if (membersToCheck.size() == 1 && membersToCheck.iterator().next().equals(selfMember)) {
+ return true;
+ }
+
+ return false;
+ }
+
+ 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());
+ if (snapshotTracker == null) {
+ snapshotTracker = new SnapshotTracker(log, installSnapshot.getTotalChunks(), installSnapshot.getLeaderId(),
+ context);
}
updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
@@ -513,9 +634,13 @@ public class Follower extends AbstractRaftActorBehavior {
final InstallSnapshotReply reply = new InstallSnapshotReply(
currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
- if(snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
- installSnapshot.getLastChunkHashCode())){
- Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
+ if (snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
+ installSnapshot.getLastChunkHashCode())) {
+
+ log.info("{}: Snapshot installed from leader: {}", logName(), installSnapshot.getLeaderId());
+
+ Snapshot snapshot = Snapshot.create(
+ context.getSnapshotManager().convertSnapshot(snapshotTracker.getSnapshotBytes()),
new ArrayList<>(),
installSnapshot.getLastIncludedIndex(),
installSnapshot.getLastIncludedTerm(),
@@ -523,12 +648,12 @@ public class Follower extends AbstractRaftActorBehavior {
installSnapshot.getLastIncludedTerm(),
context.getTermInformation().getCurrentTerm(),
context.getTermInformation().getVotedFor(),
- installSnapshot.getServerConfig().orNull());
+ installSnapshot.getServerConfig().orElse(null));
ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
@Override
public void onSuccess() {
- LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
+ log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
sender.tell(reply, actor());
}
@@ -541,36 +666,38 @@ public class Follower extends AbstractRaftActorBehavior {
actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
- snapshotTracker = null;
+ closeSnapshotTracker();
} else {
- LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
+ log.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
sender.tell(reply, actor());
}
- } catch (SnapshotTracker.InvalidChunkException e) {
- LOG.debug("{}: Exception in InstallSnapshot of follower", logName(), 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;
- } catch (Exception e){
- LOG.error("{}: Exception in InstallSnapshot of follower", logName(), 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());
+ closeSnapshotTracker();
+ }
+ }
+ private void closeSnapshotTracker() {
+ if (snapshotTracker != null) {
+ snapshotTracker.close();
+ snapshotTracker = null;
}
}
@Override
public void close() {
+ closeSnapshotTracker();
stopElection();
+ appendEntriesMessageAssembler.close();
}
@VisibleForTesting
- SnapshotTracker getSnapshotTracker(){
+ SnapshotTracker getSnapshotTracker() {
return snapshotTracker;
}
}