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%2FAbstractLeader.java;h=5076a8a38f891c894b9e01fdc61db5c0d5fa11a9;hp=bdfdd9b3765c576495e5bbf96dcdb19958c73cd5;hb=95d3c7975a423951dcbdecfbfa4cb6b7a23591cc;hpb=2727bea09c83646b6cbd2ef9672d0b7f6cf3b22f diff --git a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java index bdfdd9b376..5076a8a38f 100644 --- a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java +++ b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java @@ -14,8 +14,6 @@ import akka.actor.Cancellable; import com.google.common.annotations.VisibleForTesting; import com.google.common.base.Optional; import com.google.common.base.Preconditions; -import com.google.common.collect.ImmutableMap; -import com.google.common.collect.ImmutableMap.Builder; import com.google.protobuf.ByteString; import java.io.IOException; import java.util.Collection; @@ -26,13 +24,18 @@ import java.util.LinkedList; import java.util.List; import java.util.Map; import java.util.Map.Entry; +import java.util.Queue; +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.FollowerLogInformationImpl; +import org.opendaylight.controller.cluster.raft.PeerInfo; 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.VotingState; 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; @@ -42,6 +45,8 @@ import org.opendaylight.controller.cluster.raft.messages.InstallSnapshot; import org.opendaylight.controller.cluster.raft.messages.InstallSnapshotReply; import org.opendaylight.controller.cluster.raft.messages.RaftRPC; import org.opendaylight.controller.cluster.raft.messages.RequestVoteReply; +import org.opendaylight.controller.cluster.raft.messages.UnInitializedFollowerSnapshotReply; +import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload; import scala.concurrent.duration.FiniteDuration; /** @@ -67,56 +72,39 @@ import scala.concurrent.duration.FiniteDuration; * set commitIndex = N (§5.3, §5.4). */ public abstract class AbstractLeader extends AbstractRaftActorBehavior { + private final Map followerToLog = new HashMap<>(); - // The index of the first chunk that is sent when installing a snapshot - public static final int FIRST_CHUNK_INDEX = 1; - - // The index that the follower should respond with if it needs the install snapshot to be reset - public static final int INVALID_CHUNK_INDEX = -1; - - // This would be passed as the hash code of the last chunk when sending the first chunk - public static final int INITIAL_LAST_CHUNK_HASH_CODE = -1; - - private final Map followerToLog; - private final Map mapFollowerToSnapshot = new HashMap<>(); + /** + * Lookup table for request contexts based on journal index. We could use a {@link Map} here, but we really + * expect the entries to be modified in sequence, hence we open-code the lookup. + * + * TODO: Evaluate the use of ArrayDeque(), as that has lower memory overhead. Non-head removals are more costly, + * but we already expect those to be far from frequent. + */ + private final Queue trackers = new LinkedList<>(); private Cancellable heartbeatSchedule = null; + private Optional snapshot = Optional.absent();; + private int minReplicationCount; - private final Collection trackerList = new LinkedList<>(); - - protected final int minReplicationCount; - - protected final int minIsolatedLeaderPeerCount; - - private Optional snapshot; - - public AbstractLeader(RaftActorContext context) { - super(context, RaftState.Leader); - - final Builder ftlBuilder = ImmutableMap.builder(); - for (String followerId : context.getPeerAddresses().keySet()) { - FollowerLogInformation followerLogInformation = - new FollowerLogInformationImpl(followerId, -1, context); + protected AbstractLeader(RaftActorContext context, RaftState state, + @Nullable AbstractLeader initializeFromLeader) { + super(context, state); - ftlBuilder.put(followerId, followerLogInformation); + if(initializeFromLeader != null) { + followerToLog.putAll(initializeFromLeader.followerToLog); + snapshot = initializeFromLeader.snapshot; + trackers.addAll(initializeFromLeader.trackers); + } else { + for(PeerInfo peerInfo: context.getPeers()) { + FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(peerInfo, -1, context); + followerToLog.put(peerInfo.getId(), followerLogInformation); + } } - followerToLog = ftlBuilder.build(); - - leaderId = context.getId(); LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds()); - minReplicationCount = getMajorityVoteCount(getFollowerIds().size()); - - // the isolated Leader peer count will be 1 less than the majority vote count. - // this is because the vote count has the self vote counted in it - // for e.g - // 0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0 - // 2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1 - // 4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2 - minIsolatedLeaderPeerCount = minReplicationCount > 0 ? (minReplicationCount - 1) : 0; - - snapshot = Optional.absent(); + updateMinReplicaCount(); // Immediately schedule a heartbeat // Upon election: send initial empty AppendEntries RPCs @@ -128,6 +116,10 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval()); } + protected AbstractLeader(RaftActorContext context, RaftState state) { + this(context, state, null); + } + /** * Return an immutable collection of follower identifiers. * @@ -137,9 +129,54 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { return followerToLog.keySet(); } + public void addFollower(String followerId) { + FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl( + context.getPeerInfo(followerId), -1, context); + followerToLog.put(followerId, followerLogInformation); + + if(heartbeatSchedule == null) { + scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval()); + } + } + + public void removeFollower(String followerId) { + followerToLog.remove(followerId); + } + + public void updateMinReplicaCount() { + int numVoting = 0; + for(PeerInfo peer: context.getPeers()) { + if(peer.isVoting()) { + numVoting++; + } + } + + minReplicationCount = getMajorityVoteCount(numVoting); + } + + protected int getMinIsolatedLeaderPeerCount(){ + //the isolated Leader peer count will be 1 less than the majority vote count. + //this is because the vote count has the self vote counted in it + //for e.g + //0 peers = 1 votesRequired , minIsolatedLeaderPeerCount = 0 + //2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1 + //4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2 + + return minReplicationCount > 0 ? (minReplicationCount - 1) : 0; + } + + @VisibleForTesting + void setSnapshot(@Nullable Snapshot snapshot) { + if(snapshot != null) { + this.snapshot = Optional.of(new SnapshotHolder(snapshot)); + } else { + this.snapshot = Optional.absent(); + } + } + @VisibleForTesting - void setSnapshot(Optional snapshot) { - this.snapshot = snapshot; + boolean hasSnapshot() { + return snapshot.isPresent(); } @Override @@ -178,48 +215,116 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { } followerLogInformation.markFollowerActive(); + followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion()); + followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion()); boolean updated = false; - if (appendEntriesReply.isSuccess()) { - 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(), - followerId, followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex()); - } + if(appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) { + // The follower's log is actually ahead of the leader's log. Normally this doesn't happen + // in raft as a node cannot become leader if it's log is behind another's. However, the + // non-voting semantics deviate a bit from raft. Only voting members participate in + // elections and can become leader so it's possible for a non-voting follower to be ahead + // of the leader. This can happen if persistence is disabled and all voting members are + // restarted. In this case, the voting leader will start out with an empty log however + // the non-voting followers still retain the previous data in memory. On the first + // AppendEntries, the non-voting follower returns a successful reply b/c the prevLogIndex + // sent by the leader is -1 and thus the integrity checks pass. However the follower's returned + // lastLogIndex may be higher in which case we want to reset the follower by installing a + // snapshot. It's also possible that the follower's last log index is behind the leader's. + // However in this case the log terms won't match and the logs will conflict - this is handled + // elsewhere. + LOG.debug("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} - forcing install snaphot", + logName(), followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(), + context.getReplicatedLog().lastIndex()); + + followerLogInformation.setMatchIndex(-1); + followerLogInformation.setNextIndex(-1); + + initiateCaptureSnapshot(followerId); + updated = true; + } else if (appendEntriesReply.isSuccess()) { + updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply); } else { LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply); - // 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 + long followerLastLogIndex = appendEntriesReply.getLogLastIndex(); + long followersLastLogTerm = getLogEntryTerm(followerLastLogIndex); + if(appendEntriesReply.isForceInstallSnapshot()) { + // Reset the followers match and next index. This is to signal that this follower has nothing + // in common with this Leader and so would require a snapshot to be installed + followerLogInformation.setMatchIndex(-1); + followerLogInformation.setNextIndex(-1); + + // Force initiate a snapshot capture + initiateCaptureSnapshot(followerId); + } else if(followerLastLogIndex < 0 || (followersLastLogTerm >= 0 && + followersLastLogTerm == 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 { + // The follower's log conflicts with leader's log so decrement follower's next index by 1 + // in an attempt to find where the logs match. + + LOG.debug("{}: follower's last log term {} conflicts with the leader's {} - dec next index", + logName(), appendEntriesReply.getLogLastTerm(), followersLastLogTerm); - followerLogInformation.decrNextIndex(); + followerLogInformation.decrNextIndex(); + } } // Now figure out if this reply warrants a change in the commitIndex // If there exists an N such that N > commitIndex, a majority // of matchIndex[i] ≥ N, and log[N].term == currentTerm: // set commitIndex = N (§5.3, §5.4). + if(LOG.isTraceEnabled()) { + LOG.trace("{}: handleAppendEntriesReply from {}: commitIndex: {}, lastAppliedIndex: {}, currentTerm: {}", + logName(), followerId, context.getCommitIndex(), context.getLastApplied(), currentTerm()); + } + for (long N = context.getCommitIndex() + 1; ; N++) { int replicatedCount = 1; + LOG.trace("{}: checking Nth index {}", logName(), N); for (FollowerLogInformation info : followerToLog.values()) { - if (info.getMatchIndex() >= N) { + final PeerInfo peerInfo = context.getPeerInfo(info.getId()); + if(info.getMatchIndex() >= N && peerInfo != null && peerInfo.isVoting()) { replicatedCount++; + } else if(LOG.isTraceEnabled()) { + LOG.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(), + info.getMatchIndex(), peerInfo); } } + if(LOG.isTraceEnabled()) { + LOG.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount, minReplicationCount); + } + if (replicatedCount >= minReplicationCount) { ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N); - if (replicatedLogEntry != null && - replicatedLogEntry.getTerm() == currentTerm()) { + if (replicatedLogEntry == null) { + LOG.debug("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}", + logName(), N, context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().size()); + break; + } + + // Don't update the commit index if the log entry is from a previous term, as per §5.4.1: + // "Raft never commits log entries from previous terms by counting replicas". + // However we keep looping so we can make progress when new entries in the current term + // reach consensus, as per §5.4.1: "once an entry from the current term is committed by + // counting replicas, then all prior entries are committed indirectly". + if (replicatedLogEntry.getTerm() == currentTerm()) { + LOG.trace("{}: Setting commit index to {}", logName(), N); context.setCommitIndex(N); + } else { + LOG.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, term {} does not match the current term {}", + logName(), N, replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm()); } } else { + LOG.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount); break; } } @@ -240,9 +345,23 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { //Send the next log entry immediately, if possible, no need to wait for heartbeat to trigger that event 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 @@ -258,7 +377,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { @Override protected ClientRequestTracker removeClientRequestTracker(long logIndex) { - final Iterator it = trackerList.iterator(); + final Iterator it = trackers.iterator(); while (it.hasNext()) { final ClientRequestTracker t = it.next(); if (t.getIndex() == logIndex) { @@ -270,16 +389,6 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { return null; } - @Override - protected ClientRequestTracker findClientRequestTracker(long logIndex) { - for (ClientRequestTracker tracker : trackerList) { - if (tracker.getIndex() == logIndex) { - return tracker; - } - } - return null; - } - @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender, RequestVoteReply requestVoteReply) { @@ -289,11 +398,9 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { protected void beforeSendHeartbeat(){} @Override - public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) { + public RaftActorBehavior handleMessage(ActorRef sender, Object message) { Preconditions.checkNotNull(sender, "sender should not be null"); - Object message = fromSerializableMessage(originalMessage); - if (message instanceof RaftRPC) { RaftRPC rpc = (RaftRPC) message; // If RPC request or response contains term T > currentTerm: @@ -305,7 +412,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { context.getTermInformation().updateAndPersist(rpc.getTerm(), null); - return switchBehavior(new Follower(context)); + return internalSwitchBehavior(RaftState.Follower); } } @@ -313,44 +420,46 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { beforeSendHeartbeat(); sendHeartBeat(); scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval()); - return this; - } else if(message instanceof SendInstallSnapshot) { // received from RaftActor - setSnapshot(Optional.of(((SendInstallSnapshot) message).getSnapshot())); + setSnapshot(((SendInstallSnapshot) message).getSnapshot()); sendInstallSnapshot(); - } else if (message instanceof Replicate) { replicate((Replicate) message); - - } else if (message instanceof InstallSnapshotReply){ + } else if (message instanceof InstallSnapshotReply) { handleInstallSnapshotReply((InstallSnapshotReply) message); - + } else { + return super.handleMessage(sender, message); } - - return super.handleMessage(sender, message); + return this; } private void handleInstallSnapshotReply(InstallSnapshotReply reply) { LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply); String followerId = reply.getFollowerId(); - FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId); + FollowerLogInformation followerLogInformation = followerToLog.get(followerId); + if(followerLogInformation == null) { + // This can happen during AddServer if it times out. + LOG.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply", + logName(), followerId); + return; + } - if (followerToSnapshot == null) { - LOG.error("{}: FollowerId {} in InstallSnapshotReply not known to Leader", + LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState(); + if (installSnapshotState == null) { + LOG.error("{}: LeaderInstallSnapshotState not found for follower {} in InstallSnapshotReply", logName(), followerId); return; } - FollowerLogInformation followerLogInformation = followerToLog.get(followerId); followerLogInformation.markFollowerActive(); - if (followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) { + if (installSnapshotState.getChunkIndex() == reply.getChunkIndex()) { boolean wasLastChunk = false; if (reply.isSuccess()) { - if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) { + if(installSnapshotState.isLastChunk(reply.getChunkIndex())) { //this was the last chunk reply if(LOG.isDebugEnabled()) { LOG.debug("{}: InstallSnapshotReply received, " + @@ -360,37 +469,42 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { ); } - followerLogInformation.setMatchIndex( - context.getReplicatedLog().getSnapshotIndex()); - followerLogInformation.setNextIndex( - context.getReplicatedLog().getSnapshotIndex() + 1); - mapFollowerToSnapshot.remove(followerId); + long followerMatchIndex = snapshot.get().getLastIncludedIndex(); + followerLogInformation.setMatchIndex(followerMatchIndex); + followerLogInformation.setNextIndex(followerMatchIndex + 1); + followerLogInformation.clearLeaderInstallSnapshotState(); LOG.debug("{}: follower: {}, matchIndex set to {}, nextIndex set to {}", - logName(), followerId, followerLogInformation.getMatchIndex(), - followerLogInformation.getNextIndex()); + logName(), followerId, followerLogInformation.getMatchIndex(), + followerLogInformation.getNextIndex()); - if (mapFollowerToSnapshot.isEmpty()) { + if (!anyFollowersInstallingSnapshot()) { // once there are no pending followers receiving snapshots // we can remove snapshot from the memory - setSnapshot(Optional.absent()); + setSnapshot(null); } - wasLastChunk = true; + wasLastChunk = true; + if(context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED){ + UnInitializedFollowerSnapshotReply unInitFollowerSnapshotSuccess = + new UnInitializedFollowerSnapshotReply(followerId); + context.getActor().tell(unInitFollowerSnapshotSuccess, context.getActor()); + LOG.debug("Sent message UnInitializedFollowerSnapshotReply to self"); + } } else { - followerToSnapshot.markSendStatus(true); + installSnapshotState.markSendStatus(true); } } else { LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}", logName(), reply.getChunkIndex()); - followerToSnapshot.markSendStatus(false); + installSnapshotState.markSendStatus(false); } if (wasLastChunk && !context.getSnapshotManager().isCapturing()) { // Since the follower is now caught up try to purge the log. purgeInMemoryLog(); - } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) { + } else if (!wasLastChunk && installSnapshotState.canSendNextChunk()) { ActorSelection followerActor = context.getPeerActorSelection(followerId); if(followerActor != null) { sendSnapshotChunk(followerActor, followerId); @@ -400,39 +514,54 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { } else { LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}", logName(), reply.getChunkIndex(), followerId, - followerToSnapshot.getChunkIndex()); + installSnapshotState.getChunkIndex()); - if(reply.getChunkIndex() == INVALID_CHUNK_INDEX){ + if(reply.getChunkIndex() == LeaderInstallSnapshotState.INVALID_CHUNK_INDEX){ // Since the Follower did not find this index to be valid we should reset the follower snapshot // so that Installing the snapshot can resume from the beginning - followerToSnapshot.reset(); + installSnapshotState.reset(); } } } + private boolean anyFollowersInstallingSnapshot() { + for(FollowerLogInformation info: followerToLog.values()) { + if(info.getInstallSnapshotState() != null) { + return true; + } + + } + + return false; + } + private void replicate(Replicate replicate) { long logIndex = replicate.getReplicatedLogEntry().getIndex(); - LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}", logName(), - replicate.getIdentifier(), logIndex); + LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}", logName(), + replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass()); // Create a tracker entry we will use this later to notify the // client actor - trackerList.add( - new ClientRequestTrackerImpl(replicate.getClientActor(), - replicate.getIdentifier(), - logIndex) - ); + if(replicate.getClientActor() != null) { + trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(), + logIndex)); + } - if (followerToLog.isEmpty()) { + boolean applyModificationToState = !context.anyVotingPeers() + || context.getRaftPolicy().applyModificationToStateBeforeConsensus(); + + if(applyModificationToState){ context.setCommitIndex(logIndex); applyLogToStateMachine(logIndex); - } else { + } + + if (!followerToLog.isEmpty()) { sendAppendEntries(0, false); } } - private void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) { + protected void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) { // Send an AppendEntries to all followers for (Entry e : followerToLog.entrySet()) { final String followerId = e.getKey(); @@ -462,9 +591,10 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { boolean sendAppendEntries = false; List entries = Collections.emptyList(); - if (mapFollowerToSnapshot.get(followerId) != null) { + LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState(); + if (installSnapshotState != null) { // if install snapshot is in process , then sent next chunk if possible - if (isFollowerActive && mapFollowerToSnapshot.get(followerId).canSendNextChunk()) { + if (isFollowerActive && installSnapshotState.canSendNextChunk()) { sendSnapshotChunk(followerActor, followerId); } else if(sendHeartbeat) { // we send a heartbeat even if we have not received a reply for the last chunk @@ -475,8 +605,8 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex(); if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) { - LOG.debug("{}: Checking sendAppendEntries for follower {}, followerNextIndex {}, leaderLastIndex: {}, leaderSnapShotIndex: {}", - logName(), followerId, followerNextIndex, leaderLastIndex, leaderSnapShotIndex); + LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}", + logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex); } if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) { @@ -484,9 +614,12 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(), followerNextIndex, followerId); - // FIXME : Sending one entry at a time if(followerLogInformation.okToReplicate()) { - entries = context.getReplicatedLog().getFrom(followerNextIndex, 1); + // 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 && @@ -496,7 +629,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { // then snapshot should be sent if (LOG.isDebugEnabled()) { - LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s," + + LOG.debug(String.format("%s: InitiateInstallSnapshot to follower: %s, " + "follower-nextIndex: %d, leader-snapshot-index: %d, " + "leader-last-index: %d", logName(), followerId, followerNextIndex, leaderSnapShotIndex, leaderLastIndex)); @@ -504,7 +637,9 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { // Send heartbeat to follower whenever install snapshot is initiated. sendAppendEntries = true; - initiateCaptureSnapshot(followerId, followerNextIndex); + if (canInstallSnapshot(followerNextIndex)) { + initiateCaptureSnapshot(followerId); + } } else if(sendHeartbeat) { // we send an AppendEntries, even if the follower is inactive @@ -524,16 +659,16 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex, List entries, String followerId) { AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(), - prevLogIndex(followerNextIndex), - prevLogTerm(followerNextIndex), entries, - context.getCommitIndex(), super.getReplicatedToAllIndex()); + getLogEntryIndex(followerNextIndex - 1), + getLogEntryTerm(followerNextIndex - 1), entries, + context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion()); if(!entries.isEmpty() || LOG.isTraceEnabled()) { LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId, appendEntries); } - followerActor.tell(appendEntries.toSerializable(), actor()); + followerActor.tell(appendEntries, actor()); } /** @@ -548,39 +683,44 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { * 6. If another follower requires a snapshot and a snapshot has been collected (via CaptureSnapshotReply) * then send the existing snapshot in chunks to the follower. * @param followerId - * @param followerNextIndex */ - private void initiateCaptureSnapshot(String followerId, long followerNextIndex) { - if (!context.getReplicatedLog().isPresent(followerNextIndex) && - context.getReplicatedLog().isInSnapshot(followerNextIndex)) { - - if (snapshot.isPresent()) { - // if a snapshot is present in the memory, most likely another install is in progress - // no need to capture snapshot. - // This could happen if another follower needs an install when one is going on. - final ActorSelection followerActor = context.getPeerActorSelection(followerId); - sendSnapshotChunk(followerActor, followerId); + public boolean initiateCaptureSnapshot(String followerId) { + if (snapshot.isPresent()) { + // if a snapshot is present in the memory, most likely another install is in progress + // no need to capture snapshot. + // This could happen if another follower needs an install when one is going on. + final ActorSelection followerActor = context.getPeerActorSelection(followerId); + sendSnapshotChunk(followerActor, followerId); + return true; + } else { + return context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(), + this.getReplicatedToAllIndex(), followerId); + } + } + private boolean canInstallSnapshot(long nextIndex){ + // If the follower's nextIndex is -1 then we might as well send it a snapshot + // Otherwise send it a snapshot only if the nextIndex is not present in the log but is present + // in the snapshot + return nextIndex == -1 || + (!context.getReplicatedLog().isPresent(nextIndex) + && context.getReplicatedLog().isInSnapshot(nextIndex)); - } else { - context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(), - this.getReplicatedToAllIndex(), followerId); - } - } } private void sendInstallSnapshot() { LOG.debug("{}: sendInstallSnapshot", logName()); for (Entry e : followerToLog.entrySet()) { - ActorSelection followerActor = context.getPeerActorSelection(e.getKey()); + String followerId = e.getKey(); + ActorSelection followerActor = context.getPeerActorSelection(followerId); + FollowerLogInformation followerLogInfo = e.getValue(); if (followerActor != null) { - long nextIndex = e.getValue().getNextIndex(); - - if (!context.getReplicatedLog().isPresent(nextIndex) && - context.getReplicatedLog().isInSnapshot(nextIndex)) { - sendSnapshotChunk(followerActor, e.getKey()); + long nextIndex = followerLogInfo.getNextIndex(); + if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED || + canInstallSnapshot(nextIndex)) { + sendSnapshotChunk(followerActor, followerId); } } } @@ -593,28 +733,35 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { private void sendSnapshotChunk(ActorSelection followerActor, String followerId) { try { if (snapshot.isPresent()) { - ByteString nextSnapshotChunk = getNextSnapshotChunk(followerId,snapshot.get()); + byte[] nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes()); // Note: the previous call to getNextSnapshotChunk has the side-effect of adding // followerId to the followerToSnapshot map. - FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId); + LeaderInstallSnapshotState installSnapshotState = followerToLog.get(followerId).getInstallSnapshotState(); + + int nextChunkIndex = installSnapshotState.incrementChunkIndex(); + Optional serverConfig = Optional.absent(); + if(installSnapshotState.isLastChunk(nextChunkIndex)) { + serverConfig = Optional.fromNullable(context.getPeerServerInfo(true)); + } followerActor.tell( new InstallSnapshot(currentTerm(), context.getId(), - context.getReplicatedLog().getSnapshotIndex(), - context.getReplicatedLog().getSnapshotTerm(), + snapshot.get().getLastIncludedIndex(), + snapshot.get().getLastIncludedTerm(), nextSnapshotChunk, - followerToSnapshot.incrementChunkIndex(), - followerToSnapshot.getTotalChunks(), - Optional.of(followerToSnapshot.getLastChunkHashCode()) - ).toSerializable(), + nextChunkIndex, + installSnapshotState.getTotalChunks(), + Optional.of(installSnapshotState.getLastChunkHashCode()), + serverConfig + ).toSerializable(followerToLog.get(followerId).getRaftVersion()), actor() ); if(LOG.isDebugEnabled()) { LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}", - logName(), followerActor.path(), followerToSnapshot.getChunkIndex(), - followerToSnapshot.getTotalChunks()); + logName(), followerActor.path(), installSnapshotState.getChunkIndex(), + installSnapshotState.getTotalChunks()); } } } catch (IOException e) { @@ -626,15 +773,16 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { * Acccepts snaphot as ByteString, enters into map for future chunks * creates and return a ByteString chunk */ - private ByteString getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException { - FollowerToSnapshot followerToSnapshot = mapFollowerToSnapshot.get(followerId); - if (followerToSnapshot == null) { - followerToSnapshot = new FollowerToSnapshot(snapshotBytes); - mapFollowerToSnapshot.put(followerId, followerToSnapshot); + private byte[] getNextSnapshotChunk(String followerId, ByteString snapshotBytes) throws IOException { + LeaderInstallSnapshotState installSnapshotState = followerToLog.get(followerId).getInstallSnapshotState(); + if (installSnapshotState == null) { + installSnapshotState = new LeaderInstallSnapshotState(snapshotBytes, context.getConfigParams().getSnapshotChunkSize(), + logName()); + followerToLog.get(followerId).setLeaderInstallSnapshotState(installSnapshotState); } - ByteString nextChunk = followerToSnapshot.getNextChunk(); + byte[] nextChunk = installSnapshotState.getNextChunk(); - LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.size()); + LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.length); return nextChunk; } @@ -667,147 +815,37 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { // need to be sent if there are other messages being sent to the remote // actor. heartbeatSchedule = context.getActorSystem().scheduler().scheduleOnce( - interval, context.getActor(), new SendHeartBeat(), + interval, context.getActor(), SendHeartBeat.INSTANCE, context.getActorSystem().dispatcher(), context.getActor()); } @Override - public void close() throws Exception { + public void close() { stopHeartBeat(); } @Override - public String getLeaderId() { + public final String getLeaderId() { return context.getId(); } + @Override + public final short getLeaderPayloadVersion() { + return context.getPayloadVersion(); + } + protected boolean isLeaderIsolated() { - int minPresent = minIsolatedLeaderPeerCount; + int minPresent = getMinIsolatedLeaderPeerCount(); for (FollowerLogInformation followerLogInformation : followerToLog.values()) { - if (followerLogInformation.isFollowerActive()) { + final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId()); + if(peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) { --minPresent; if (minPresent == 0) { - break; + return false; } } } - return (minPresent != 0); - } - - /** - * Encapsulates the snapshot bytestring and handles the logic of sending - * snapshot chunks - */ - protected class FollowerToSnapshot { - private final ByteString snapshotBytes; - private int offset = 0; - // the next snapshot chunk is sent only if the replyReceivedForOffset matches offset - private int replyReceivedForOffset; - // if replyStatus is false, the previous chunk is attempted - private boolean replyStatus = false; - private int chunkIndex; - private final int totalChunks; - private int lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE; - private int nextChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE; - - public FollowerToSnapshot(ByteString snapshotBytes) { - this.snapshotBytes = snapshotBytes; - int size = snapshotBytes.size(); - totalChunks = ( size / context.getConfigParams().getSnapshotChunkSize()) + - ((size % context.getConfigParams().getSnapshotChunkSize()) > 0 ? 1 : 0); - if(LOG.isDebugEnabled()) { - LOG.debug("{}: Snapshot {} bytes, total chunks to send:{}", - logName(), size, totalChunks); - } - replyReceivedForOffset = -1; - chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX; - } - - public ByteString getSnapshotBytes() { - return snapshotBytes; - } - - public int incrementOffset() { - if(replyStatus) { - // if prev chunk failed, we would want to sent the same chunk again - offset = offset + context.getConfigParams().getSnapshotChunkSize(); - } - return offset; - } - - public int incrementChunkIndex() { - if (replyStatus) { - // if prev chunk failed, we would want to sent the same chunk again - chunkIndex = chunkIndex + 1; - } - return chunkIndex; - } - - public int getChunkIndex() { - return chunkIndex; - } - - public int getTotalChunks() { - return totalChunks; - } - - public boolean canSendNextChunk() { - // we only send a false if a chunk is sent but we have not received a reply yet - return replyReceivedForOffset == offset; - } - - public boolean isLastChunk(int chunkIndex) { - return totalChunks == chunkIndex; - } - - public void markSendStatus(boolean success) { - if (success) { - // if the chunk sent was successful - replyReceivedForOffset = offset; - replyStatus = true; - lastChunkHashCode = nextChunkHashCode; - } else { - // if the chunk sent was failure - replyReceivedForOffset = offset; - replyStatus = false; - } - } - - public ByteString getNextChunk() { - int snapshotLength = getSnapshotBytes().size(); - int start = incrementOffset(); - int size = context.getConfigParams().getSnapshotChunkSize(); - if (context.getConfigParams().getSnapshotChunkSize() > snapshotLength) { - size = snapshotLength; - } else { - if ((start + context.getConfigParams().getSnapshotChunkSize()) > snapshotLength) { - size = snapshotLength - start; - } - } - - - LOG.debug("{}: Next chunk: length={}, offset={},size={}", logName(), - snapshotLength, start, size); - - ByteString substring = getSnapshotBytes().substring(start, start + size); - nextChunkHashCode = substring.hashCode(); - return substring; - } - - /** - * reset should be called when the Follower needs to be sent the snapshot from the beginning - */ - public void reset(){ - offset = 0; - replyStatus = false; - replyReceivedForOffset = offset; - chunkIndex = AbstractLeader.FIRST_CHUNK_INDEX; - lastChunkHashCode = AbstractLeader.INITIAL_LAST_CHUNK_HASH_CODE; - } - - public int getLastChunkHashCode() { - return lastChunkHashCode; - } + return minPresent != 0; } // called from example-actor for printing the follower-states @@ -833,17 +871,31 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior { } @VisibleForTesting - protected void setFollowerSnapshot(String followerId, FollowerToSnapshot snapshot) { - mapFollowerToSnapshot.put(followerId, snapshot); + public int followerLogSize() { + return followerToLog.size(); } - @VisibleForTesting - public int followerSnapshotSize() { - return mapFollowerToSnapshot.size(); - } + private static class SnapshotHolder { + private final long lastIncludedTerm; + private final long lastIncludedIndex; + private final ByteString snapshotBytes; - @VisibleForTesting - public int followerLogSize() { - return followerToLog.size(); + 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; + } } }