import java.util.List;
import java.util.Map;
import java.util.Map.Entry;
+import javax.annotation.Nullable;
import org.opendaylight.controller.cluster.raft.ClientRequestTracker;
import org.opendaylight.controller.cluster.raft.ClientRequestTrackerImpl;
import org.opendaylight.controller.cluster.raft.FollowerLogInformation;
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.base.messages.CaptureSnapshot;
+import org.opendaylight.controller.cluster.raft.Snapshot;
import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
protected final int minIsolatedLeaderPeerCount;
- private Optional<ByteString> snapshot;
+ private Optional<SnapshotHolder> snapshot;
public AbstractLeader(RaftActorContext context) {
super(context, RaftState.Leader);
+ setLeaderPayloadVersion(context.getPayloadVersion());
+
final Builder<String, FollowerLogInformation> ftlBuilder = ImmutableMap.builder();
for (String followerId : context.getPeerAddresses().keySet()) {
FollowerLogInformation followerLogInformation =
- new FollowerLogInformationImpl(followerId,
- context.getCommitIndex(), -1,
- context.getConfigParams().getElectionTimeOutInterval());
+ new FollowerLogInformationImpl(followerId, -1, context);
ftlBuilder.put(followerId, followerLogInformation);
}
// (heartbeat) to each server; repeat during idle periods to
// prevent election timeouts (ยง5.2)
sendAppendEntries(0, false);
+
+ // It is important to schedule this heartbeat here
+ scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
}
/**
*
* @return Collection of follower IDs
*/
- protected final Collection<String> getFollowerIds() {
+ public final Collection<String> getFollowerIds() {
return followerToLog.keySet();
}
@VisibleForTesting
- void setSnapshot(Optional<ByteString> snapshot) {
- this.snapshot = snapshot;
+ void setSnapshot(@Nullable Snapshot snapshot) {
+ if(snapshot != null) {
+ this.snapshot = Optional.of(new SnapshotHolder(snapshot));
+ } else {
+ this.snapshot = Optional.absent();
+ }
}
@Override
if(LOG.isTraceEnabled()) {
LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
- } else if(LOG.isDebugEnabled() && !appendEntriesReply.isSuccess()) {
- LOG.debug("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
}
// Update the FollowerLogInformation
if(followerLogInformation.timeSinceLastActivity() >
context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
- LOG.error("{} : handleAppendEntriesReply delayed beyond election timeout, " +
+ LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
"appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
context.getLastApplied(), context.getCommitIndex());
}
followerLogInformation.markFollowerActive();
+ followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
+ boolean updated = false;
if (appendEntriesReply.isSuccess()) {
- followerLogInformation
- .setMatchIndex(appendEntriesReply.getLogLastIndex());
- followerLogInformation
- .setNextIndex(appendEntriesReply.getLogLastIndex() + 1);
+ updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
} else {
+ LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
+
+ long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
+ ReplicatedLogEntry followersLastLogEntry = context.getReplicatedLog().get(followerLastLogIndex);
+ if(followerLastLogIndex < 0 || (followersLastLogEntry != null &&
+ followersLastLogEntry.getTerm() == appendEntriesReply.getLogLastTerm())) {
+ // The follower's log is empty or the last entry is present in the leader's journal
+ // and the terms match so the follower is just behind the leader's journal from
+ // the last snapshot, if any. We'll catch up the follower quickly by starting at the
+ // follower's last log index.
+
+ updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
+ } else {
+ // TODO: When we find that the follower is out of sync with the
+ // Leader we simply decrement that followers next index by 1.
+ // Would it be possible to do better than this? The RAFT spec
+ // does not explicitly deal with it but may be something for us to
+ // think about.
- // TODO: When we find that the follower is out of sync with the
- // Leader we simply decrement that followers next index by 1.
- // Would it be possible to do better than this? The RAFT spec
- // does not explicitly deal with it but may be something for us to
- // think about
-
- followerLogInformation.decrNextIndex();
+ followerLogInformation.decrNextIndex();
+ }
}
// Now figure out if this reply warrants a change in the commitIndex
// Apply the change to the state machine
if (context.getCommitIndex() > context.getLastApplied()) {
- LOG.debug("{}: handleAppendEntriesReply: applying to log - commitIndex: {}, lastAppliedIndex: {}",
- logName(), context.getCommitIndex(), context.getLastApplied());
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
+ logName(), followerId, context.getCommitIndex(), context.getLastApplied());
+ }
applyLogToStateMachine(context.getCommitIndex());
}
- if (!context.isSnapshotCaptureInitiated()) {
+ if (!context.getSnapshotManager().isCapturing()) {
purgeInMemoryLog();
}
//Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event
- sendUpdatesToFollower(followerId, followerLogInformation, false, false);
+ sendUpdatesToFollower(followerId, followerLogInformation, false, !updated);
return this;
}
+ private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
+ AppendEntriesReply appendEntriesReply) {
+ boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
+ updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
+
+ if(updated && LOG.isDebugEnabled()) {
+ LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
+ logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
+ followerLogInformation.getNextIndex());
+ }
+ return updated;
+ }
+
private void purgeInMemoryLog() {
//find the lowest index across followers which has been replicated to all.
// lastApplied if there are no followers, so that we keep clearing the log for single-node
}
}
- try {
- if (message instanceof SendHeartBeat) {
- beforeSendHeartbeat();
- sendHeartBeat();
- return this;
+ if (message instanceof SendHeartBeat) {
+ beforeSendHeartbeat();
+ sendHeartBeat();
+ scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
+ return this;
- } else if(message instanceof SendInstallSnapshot) {
- // received from RaftActor
- setSnapshot(Optional.of(((SendInstallSnapshot) message).getSnapshot()));
- sendInstallSnapshot();
+ } else if(message instanceof SendInstallSnapshot) {
+ // received from RaftActor
+ setSnapshot(((SendInstallSnapshot) message).getSnapshot());
+ sendInstallSnapshot();
- } else if (message instanceof Replicate) {
- replicate((Replicate) message);
+ } else if (message instanceof Replicate) {
+ replicate((Replicate) message);
- } else if (message instanceof InstallSnapshotReply){
- handleInstallSnapshotReply((InstallSnapshotReply) message);
+ } else if (message instanceof InstallSnapshotReply){
+ handleInstallSnapshotReply((InstallSnapshotReply) message);
- }
- } finally {
- scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
}
+
return super.handleMessage(sender, message);
}
);
}
- followerLogInformation.setMatchIndex(
- context.getReplicatedLog().getSnapshotIndex());
- followerLogInformation.setNextIndex(
- context.getReplicatedLog().getSnapshotIndex() + 1);
+ long followerMatchIndex = snapshot.get().getLastIncludedIndex();
+ followerLogInformation.setMatchIndex(followerMatchIndex);
+ followerLogInformation.setNextIndex(followerMatchIndex + 1);
mapFollowerToSnapshot.remove(followerId);
LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}",
if (mapFollowerToSnapshot.isEmpty()) {
// once there are no pending followers receiving snapshots
// we can remove snapshot from the memory
- setSnapshot(Optional.<ByteString>absent());
+ setSnapshot(null);
}
wasLastChunk = true;
followerToSnapshot.markSendStatus(false);
}
- if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
+ if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
+ // Since the follower is now caught up try to purge the log.
+ purgeInMemoryLog();
+ } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
ActorSelection followerActor = context.getPeerActorSelection(followerId);
if(followerActor != null) {
sendSnapshotChunk(followerActor, followerId);
if (followerActor != null) {
long followerNextIndex = followerLogInformation.getNextIndex();
boolean isFollowerActive = followerLogInformation.isFollowerActive();
+ boolean sendAppendEntries = false;
+ List<ReplicatedLogEntry> entries = Collections.emptyList();
if (mapFollowerToSnapshot.get(followerId) != null) {
// if install snapshot is in process , then sent next chunk if possible
sendSnapshotChunk(followerActor, followerId);
} else if(sendHeartbeat) {
// we send a heartbeat even if we have not received a reply for the last chunk
- sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
- Collections.<ReplicatedLogEntry>emptyList(), followerId);
+ sendAppendEntries = true;
}
} else {
long leaderLastIndex = context.getReplicatedLog().lastIndex();
long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
- if(!isHeartbeat || LOG.isTraceEnabled()) {
- LOG.debug("{}: Checking sendAppendEntries for follower {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
- logName(), followerId, leaderLastIndex, leaderSnapShotIndex);
+ if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
+ LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
+ logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
}
if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
followerNextIndex, followerId);
- // FIXME : Sending one entry at a time
- final List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
-
- sendAppendEntriesToFollower(followerActor, followerNextIndex, entries, followerId);
-
+ if(followerLogInformation.okToReplicate()) {
+ // Try to send all the entries in the journal but not exceeding the max data size
+ // for a single AppendEntries message.
+ int maxEntries = (int) context.getReplicatedLog().size();
+ entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
+ context.getConfigParams().getSnapshotChunkSize());
+ sendAppendEntries = true;
+ }
} else if (isFollowerActive && followerNextIndex >= 0 &&
- leaderLastIndex > followerNextIndex && !context.isSnapshotCaptureInitiated()) {
+ leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
// if the followers next index is not present in the leaders log, and
// if the follower is just not starting and if leader's index is more than followers index
// then snapshot should be sent
}
// Send heartbeat to follower whenever install snapshot is initiated.
- sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
- Collections.<ReplicatedLogEntry>emptyList(), followerId);
-
+ sendAppendEntries = true;
initiateCaptureSnapshot(followerId, followerNextIndex);
} else if(sendHeartbeat) {
- //we send an AppendEntries, even if the follower is inactive
+ // we send an AppendEntries, even if the follower is inactive
// in-order to update the followers timestamp, in case it becomes active again
- sendAppendEntriesToFollower(followerActor, followerLogInformation.getNextIndex(),
- Collections.<ReplicatedLogEntry>emptyList(), followerId);
+ sendAppendEntries = true;
}
}
+
+ if(sendAppendEntries) {
+ sendAppendEntriesToFollower(followerActor, followerNextIndex,
+ entries, followerId);
+ }
}
}
AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
prevLogIndex(followerNextIndex),
prevLogTerm(followerNextIndex), entries,
- context.getCommitIndex(), super.getReplicatedToAllIndex());
+ context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
if(!entries.isEmpty() || LOG.isTraceEnabled()) {
LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
final ActorSelection followerActor = context.getPeerActorSelection(followerId);
sendSnapshotChunk(followerActor, followerId);
- } else if (!context.isSnapshotCaptureInitiated()) {
-
- LOG.info("{}: Initiating Snapshot Capture to Install Snapshot, Leader:{}", logName(), getLeaderId());
- ReplicatedLogEntry lastAppliedEntry = context.getReplicatedLog().get(context.getLastApplied());
- long lastAppliedIndex = -1;
- long lastAppliedTerm = -1;
-
- if (lastAppliedEntry != null) {
- lastAppliedIndex = lastAppliedEntry.getIndex();
- lastAppliedTerm = lastAppliedEntry.getTerm();
- } else if (context.getReplicatedLog().getSnapshotIndex() > -1) {
- lastAppliedIndex = context.getReplicatedLog().getSnapshotIndex();
- lastAppliedTerm = context.getReplicatedLog().getSnapshotTerm();
- }
- boolean isInstallSnapshotInitiated = true;
- long replicatedToAllIndex = super.getReplicatedToAllIndex();
- ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
- actor().tell(new CaptureSnapshot(lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
- (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
- (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1),
- isInstallSnapshotInitiated), actor());
- context.setSnapshotCaptureInitiated(true);
+ } else {
+ context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
+ this.getReplicatedToAllIndex(), followerId);
}
}
}
private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
try {
if (snapshot.isPresent()) {
- ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId,snapshot.get());
+ ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
// Note: the previous call to getNextSnapshotChunk has the side-effect of adding
// followerId to the followerToSnapshot map.
followerActor.tell(
new InstallSnapshot(currentTerm(), context.getId(),
- context.getReplicatedLog().getSnapshotIndex(),
- context.getReplicatedLog().getSnapshotTerm(),
+ snapshot.get().getLastIncludedIndex(),
+ snapshot.get().getLastIncludedTerm(),
nextSnapshotChunk,
followerToSnapshot.incrementChunkIndex(),
followerToSnapshot.getTotalChunks(),
).toSerializable(),
actor()
);
- LOG.info("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
- logName(), followerActor.path(),
- followerToSnapshot.getChunkIndex(),
- followerToSnapshot.getTotalChunks());
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
+ logName(), followerActor.path(), followerToSnapshot.getChunkIndex(),
+ followerToSnapshot.getTotalChunks());
+ }
}
} catch (IOException e) {
LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
public int followerLogSize() {
return followerToLog.size();
}
+
+ private static class SnapshotHolder {
+ private final long lastIncludedTerm;
+ private final long lastIncludedIndex;
+ private final ByteString snapshotBytes;
+
+ SnapshotHolder(Snapshot snapshot) {
+ this.lastIncludedTerm = snapshot.getLastAppliedTerm();
+ this.lastIncludedIndex = snapshot.getLastAppliedIndex();
+ this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
+ }
+
+ long getLastIncludedTerm() {
+ return lastIncludedTerm;
+ }
+
+ long getLastIncludedIndex() {
+ return lastIncludedIndex;
+ }
+
+ ByteString getSnapshotBytes() {
+ return snapshotBytes;
+ }
+ }
}