* 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 static java.util.Objects.requireNonNull;
+
import akka.actor.ActorRef;
import akka.actor.ActorSelection;
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.protobuf.ByteString;
+import com.google.common.io.ByteSource;
+import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
import java.io.IOException;
+import java.io.ObjectOutputStream;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
import java.util.Map;
import java.util.Map.Entry;
import java.util.Queue;
-import javax.annotation.Nullable;
+import java.util.concurrent.TimeUnit;
+import org.eclipse.jdt.annotation.Nullable;
+import org.opendaylight.controller.cluster.io.SharedFileBackedOutputStream;
+import org.opendaylight.controller.cluster.messaging.MessageSlicer;
+import org.opendaylight.controller.cluster.messaging.SliceOptions;
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.CheckConsensusReached;
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;
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.RequestVote;
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 org.opendaylight.controller.cluster.raft.persisted.Snapshot;
import scala.concurrent.duration.FiniteDuration;
/**
- * The behavior of a RaftActor when it is in the Leader state
- * <p/>
+ * The behavior of a RaftActor when it is in the Leader state.
+ *
+ * <p>
* Leaders:
* <ul>
* <li> Upon election: send initial empty AppendEntries RPCs
* respond after entry applied to state machine (§5.3)
* <li> If last log index ≥ nextIndex for a follower: send
* AppendEntries RPC with log entries starting at nextIndex
- * <ul>
* <li> If successful: update nextIndex and matchIndex for
* follower (§5.3)
* <li> If AppendEntries fails because of log inconsistency:
* decrement nextIndex and retry (§5.3)
- * </ul>
- * <li> If there exists an N such that N > commitIndex, a majority
+ * <li> 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).
+ * </ul>
*/
public abstract class AbstractLeader extends AbstractRaftActorBehavior {
private final Map<String, FollowerLogInformation> followerToLog = 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<ClientRequestTracker> trackers = new LinkedList<>();
+ /**
+ * Map of serialized AppendEntries output streams keyed by log index. This is used in conjunction with the
+ * appendEntriesMessageSlicer for slicing single ReplicatedLogEntry payloads that exceed the message size threshold.
+ * This Map allows the SharedFileBackedOutputStreams to be reused for multiple followers.
+ */
+ private final Map<Long, SharedFileBackedOutputStream> sharedSerializedAppendEntriesStreams = new HashMap<>();
+ private final MessageSlicer appendEntriesMessageSlicer;
+
private Cancellable heartbeatSchedule = null;
- private Optional<SnapshotHolder> snapshot = Optional.absent();;
+ private Optional<SnapshotHolder> snapshotHolder = Optional.absent();
private int minReplicationCount;
- protected AbstractLeader(RaftActorContext context, RaftState state,
- @Nullable AbstractLeader initializeFromLeader) {
+ protected AbstractLeader(final RaftActorContext context, final RaftState state,
+ final @Nullable AbstractLeader initializeFromLeader) {
super(context, state);
- if(initializeFromLeader != null) {
+ appendEntriesMessageSlicer = MessageSlicer.builder().logContext(logName())
+ .messageSliceSize(context.getConfigParams().getSnapshotChunkSize())
+ .expireStateAfterInactivity(context.getConfigParams().getElectionTimeOutInterval().toMillis() * 3,
+ TimeUnit.MILLISECONDS).build();
+
+ if (initializeFromLeader != null) {
followerToLog.putAll(initializeFromLeader.followerToLog);
- snapshot = initializeFromLeader.snapshot;
+ snapshotHolder = initializeFromLeader.snapshotHolder;
trackers.addAll(initializeFromLeader.trackers);
} else {
- for(PeerInfo peerInfo: context.getPeers()) {
- FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(peerInfo, -1, context);
+ for (PeerInfo peerInfo: context.getPeers()) {
+ FollowerLogInformation followerLogInformation = new FollowerLogInformation(peerInfo, context);
followerToLog.put(peerInfo.getId(), followerLogInformation);
}
}
- LOG.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
+ log.debug("{}: Election: Leader has following peers: {}", logName(), getFollowerIds());
updateMinReplicaCount();
scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
}
- protected AbstractLeader(RaftActorContext context, RaftState state) {
+ protected AbstractLeader(final RaftActorContext context, final RaftState state) {
this(context, state, null);
}
return followerToLog.keySet();
}
- public void addFollower(String followerId) {
- FollowerLogInformation followerLogInformation = new FollowerLogInformationImpl(
- context.getPeerInfo(followerId), -1, context);
+ public void addFollower(final String followerId) {
+ FollowerLogInformation followerLogInformation = new FollowerLogInformation(context.getPeerInfo(followerId),
+ context);
followerToLog.put(followerId, followerLogInformation);
- if(heartbeatSchedule == null) {
+ if (heartbeatSchedule == null) {
scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
}
}
- public void removeFollower(String followerId) {
+ public void removeFollower(final String followerId) {
followerToLog.remove(followerId);
}
public void updateMinReplicaCount() {
int numVoting = 0;
- for(PeerInfo peer: context.getPeers()) {
- if(peer.isVoting()) {
+ for (PeerInfo peer: context.getPeers()) {
+ if (peer.isVoting()) {
numVoting++;
}
}
minReplicationCount = getMajorityVoteCount(numVoting);
}
- protected int getMinIsolatedLeaderPeerCount(){
+ 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
//2 peers = 2 votesRequired , minIsolatedLeaderPeerCount = 1
//4 peers = 3 votesRequired, minIsolatedLeaderPeerCount = 2
- return minReplicationCount > 0 ? (minReplicationCount - 1) : 0;
+ 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();
- }
+ void setSnapshotHolder(final @Nullable SnapshotHolder snapshotHolder) {
+ this.snapshotHolder = Optional.fromNullable(snapshotHolder);
}
@VisibleForTesting
boolean hasSnapshot() {
- return snapshot.isPresent();
+ return snapshotHolder.isPresent();
}
@Override
- protected RaftActorBehavior handleAppendEntries(ActorRef sender,
- AppendEntries appendEntries) {
+ protected RaftActorBehavior handleAppendEntries(final ActorRef sender,
+ final AppendEntries appendEntries) {
- LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
+ log.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
return this;
}
@Override
- protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
- AppendEntriesReply appendEntriesReply) {
-
- if(LOG.isTraceEnabled()) {
- LOG.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
- }
+ protected RaftActorBehavior handleAppendEntriesReply(final ActorRef sender,
+ final AppendEntriesReply appendEntriesReply) {
+ log.trace("{}: handleAppendEntriesReply: {}", logName(), appendEntriesReply);
// Update the FollowerLogInformation
String followerId = appendEntriesReply.getFollowerId();
- FollowerLogInformation followerLogInformation =
- followerToLog.get(followerId);
+ FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
- if(followerLogInformation == null){
- LOG.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
+ if (followerLogInformation == null) {
+ log.error("{}: handleAppendEntriesReply - unknown follower {}", logName(), followerId);
return this;
}
- if(followerLogInformation.timeSinceLastActivity() >
- context.getConfigParams().getElectionTimeOutInterval().toMillis()) {
- LOG.warn("{} : handleAppendEntriesReply delayed beyond election timeout, " +
- "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
- logName(), appendEntriesReply, followerLogInformation.timeSinceLastActivity(),
+ final long lastActivityNanos = followerLogInformation.nanosSinceLastActivity();
+ if (lastActivityNanos > context.getConfigParams().getElectionTimeOutInterval().toNanos()) {
+ log.warn("{} : handleAppendEntriesReply delayed beyond election timeout, "
+ + "appendEntriesReply : {}, timeSinceLastActivity : {}, lastApplied : {}, commitIndex : {}",
+ logName(), appendEntriesReply, TimeUnit.NANOSECONDS.toMillis(lastActivityNanos),
context.getLastApplied(), context.getCommitIndex());
}
followerLogInformation.markFollowerActive();
followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
followerLogInformation.setRaftVersion(appendEntriesReply.getRaftVersion());
+ followerLogInformation.setNeedsLeaderAddress(appendEntriesReply.isNeedsLeaderAddress());
+ long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
boolean updated = false;
- if(appendEntriesReply.getLogLastIndex() > context.getReplicatedLog().lastIndex()) {
+ 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
// 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());
+ log.info("{}: handleAppendEntriesReply: follower {} lastIndex {} is ahead of our lastIndex {} "
+ + "(snapshotIndex {}, snapshotTerm {}) - forcing install snaphot", logName(),
+ followerLogInformation.getId(), appendEntriesReply.getLogLastIndex(),
+ context.getReplicatedLog().lastIndex(), context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm());
followerLogInformation.setMatchIndex(-1);
followerLogInformation.setNextIndex(-1);
initiateCaptureSnapshot(followerId);
+
updated = true;
} else if (appendEntriesReply.isSuccess()) {
- updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
+ long followersLastLogTermInLeadersLog = getLogEntryTerm(followerLastLogIndex);
+ if (followerLastLogIndex >= 0 && followersLastLogTermInLeadersLog >= 0
+ && followersLastLogTermInLeadersLog != appendEntriesReply.getLogLastTerm()) {
+ // The follower's last entry is present in the leader's journal but the terms don't match so the
+ // follower has a conflicting entry. Since the follower didn't report that it's out of sync, this means
+ // either the previous leader entry sent didn't conflict or the previous leader entry is in the snapshot
+ // and no longer in the journal. Either way, we set the follower's next index to 1 less than the last
+ // index reported by the follower. For the former case, the leader will send all entries starting with
+ // the previous follower's index and the follower will remove and replace the conflicting entries as
+ // needed. For the latter, the leader will initiate an install snapshot.
+
+ followerLogInformation.setNextIndex(followerLastLogIndex - 1);
+ updated = true;
+
+ log.info("{}: handleAppendEntriesReply: follower {} last log term {} for index {} conflicts with the "
+ + "leader's {} - set the follower's next index to {}", logName(),
+ followerId, appendEntriesReply.getLogLastTerm(), appendEntriesReply.getLogLastIndex(),
+ followersLastLogTermInLeadersLog, followerLogInformation.getNextIndex());
+ } else {
+ updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
+ }
} else {
- LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
+ log.info("{}: handleAppendEntriesReply - received unsuccessful reply: {}, leader snapshotIndex: {}, "
+ + "snapshotTerm: {}, replicatedToAllIndex: {}", logName(), appendEntriesReply,
+ context.getReplicatedLog().getSnapshotIndex(), context.getReplicatedLog().getSnapshotTerm(),
+ getReplicatedToAllIndex());
- long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
- long followersLastLogTerm = getLogEntryTerm(followerLastLogIndex);
- if(appendEntriesReply.isForceInstallSnapshot()) {
+ long followersLastLogTermInLeadersLogOrSnapshot = getLogEntryOrSnapshotTerm(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);
// 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.
+ } else if (followerLastLogIndex < 0 || followersLastLogTermInLeadersLogOrSnapshot >= 0
+ && followersLastLogTermInLeadersLogOrSnapshot == appendEntriesReply.getLogLastTerm()) {
+ // The follower's log is empty or the follower's last entry is present in the leader's journal or
+ // snapshot 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);
+
+ log.info("{}: follower {} appears to be behind the leader from the last snapshot - "
+ + "updated: matchIndex: {}, nextIndex: {}", logName(), followerId,
+ followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
} else {
- // The follower's log conflicts with leader's log so decrement follower's next index by 1
+ // The follower's log conflicts with leader's log so decrement follower's next index
// in an attempt to find where the logs match.
+ if (followerLogInformation.decrNextIndex(appendEntriesReply.getLogLastIndex())) {
+ updated = true;
- LOG.debug("{}: follower's last log term {} conflicts with the leader's {} - dec next index",
- logName(), appendEntriesReply.getLogLastTerm(), followersLastLogTerm);
-
- followerLogInformation.decrNextIndex();
+ log.info("{}: follower {} last log term {} conflicts with the leader's {} - dec next index to {}",
+ logName(), followerId, appendEntriesReply.getLogLastTerm(),
+ followersLastLogTermInLeadersLogOrSnapshot, followerLogInformation.getNextIndex());
+ }
}
}
- // 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: {}",
+ 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;
+ possiblyUpdateCommitIndex();
- LOG.trace("{}: checking Nth index {}", logName(), N);
+ //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 void possiblyUpdateCommitIndex() {
+ // Figure out if we can update the the commitIndex as follows:
+ // If there exists an index N such that N > commitIndex, a majority of matchIndex[i] ≥ N,
+ // and log[N].term == currentTerm:
+ // set commitIndex = N (§5.3, §5.4).
+ for (long index = context.getCommitIndex() + 1; ; index++) {
+ ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(index);
+ if (replicatedLogEntry == null) {
+ log.trace("{}: ReplicatedLogEntry not found for index {} - snapshotIndex: {}, journal size: {}",
+ logName(), index, context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().size());
+ break;
+ }
+
+ // Count our entry if it has been persisted.
+ int replicatedCount = replicatedLogEntry.isPersistencePending() ? 0 : 1;
+
+ if (replicatedCount == 0) {
+ // We don't commit and apply a log entry until we've gotten the ack from our local persistence,
+ // even though there *shouldn't* be any issue with updating the commit index if we get a consensus
+ // amongst the followers w/o the local persistence ack.
+ break;
+ }
+
+ log.trace("{}: checking Nth index {}", logName(), index);
for (FollowerLogInformation info : followerToLog.values()) {
final PeerInfo peerInfo = context.getPeerInfo(info.getId());
- if(info.getMatchIndex() >= N && peerInfo != null && peerInfo.isVoting()) {
+ if (info.getMatchIndex() >= index && peerInfo != null && peerInfo.isVoting()) {
replicatedCount++;
- } else if(LOG.isTraceEnabled()) {
- LOG.trace("{}: Not counting follower {} - matchIndex: {}, {}", logName(), info.getId(),
+ } 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 (log.isTraceEnabled()) {
+ log.trace("{}: replicatedCount {}, minReplicationCount: {}", logName(), replicatedCount,
+ minReplicationCount);
}
if (replicatedCount >= minReplicationCount) {
- ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
- 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);
+ log.trace("{}: Setting commit index to {}", logName(), index);
+ context.setCommitIndex(index);
} 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());
+ log.debug("{}: Not updating commit index to {} - retrieved log entry with index {}, "
+ + "term {} does not match the current term {}", logName(), index,
+ replicatedLogEntry.getIndex(), replicatedLogEntry.getTerm(), currentTerm());
}
} else {
- LOG.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
+ log.trace("{}: minReplicationCount not reached, actual {} - breaking", logName(), replicatedCount);
break;
}
}
// Apply the change to the state machine
if (context.getCommitIndex() > context.getLastApplied()) {
- if(LOG.isDebugEnabled()) {
- LOG.debug("{}: handleAppendEntriesReply from {}: applying to log - commitIndex: {}, lastAppliedIndex: {}",
- logName(), followerId, context.getCommitIndex(), context.getLastApplied());
- }
+ log.debug("{}: Applying to log - commitIndex: {}, lastAppliedIndex: {}", logName(),
+ context.getCommitIndex(), context.getLastApplied());
applyLogToStateMachine(context.getCommitIndex());
}
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, !updated);
-
- return this;
}
- private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
- AppendEntriesReply appendEntriesReply) {
+ private boolean updateFollowerLogInformation(final FollowerLogInformation followerLogInformation,
+ final 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());
+ if (updated && log.isDebugEnabled()) {
+ log.debug(
+ "{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
+ logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
+ followerLogInformation.getNextIndex());
}
return updated;
}
}
@Override
- protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
+ protected ClientRequestTracker removeClientRequestTracker(final long logIndex) {
final Iterator<ClientRequestTracker> it = trackers.iterator();
while (it.hasNext()) {
final ClientRequestTracker t = it.next();
}
@Override
- protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
- RequestVoteReply requestVoteReply) {
+ protected RaftActorBehavior handleRequestVoteReply(final ActorRef sender,
+ final RequestVoteReply requestVoteReply) {
return this;
}
protected void beforeSendHeartbeat(){}
@Override
- public RaftActorBehavior handleMessage(ActorRef sender, Object message) {
- Preconditions.checkNotNull(sender, "sender should not be null");
+ public RaftActorBehavior handleMessage(final ActorRef sender, final Object message) {
+ requireNonNull(sender, "sender should not be null");
+
+ if (appendEntriesMessageSlicer.handleMessage(message)) {
+ return this;
+ }
if (message instanceof RaftRPC) {
RaftRPC rpc = (RaftRPC) message;
// 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 leader's term {} - switching to Follower",
+ log.info("{}: Term {} in \"{}\" message is greater than leader's term {} - switching to Follower",
logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
+ // This is a special case. Normally when stepping down as leader we don't process and reply to the
+ // RaftRPC as per raft. But if we're in the process of transferring leadership and we get a
+ // RequestVote, process the RequestVote before switching to Follower. This enables the requesting
+ // candidate node to be elected the leader faster and avoids us possibly timing out in the Follower
+ // state and starting a new election and grabbing leadership back before the other candidate node can
+ // start a new election due to lack of responses. This case would only occur if there isn't a majority
+ // of other nodes available that can elect the requesting candidate. Since we're transferring
+ // leadership, we should make every effort to get the requesting node elected.
+ if (message instanceof RequestVote && context.getRaftActorLeadershipTransferCohort() != null) {
+ log.debug("{}: Leadership transfer in progress - processing RequestVote", logName());
+ super.handleMessage(sender, message);
+ }
+
return internalSwitchBehavior(RaftState.Follower);
}
}
beforeSendHeartbeat();
sendHeartBeat();
scheduleHeartBeat(context.getConfigParams().getHeartBeatInterval());
- } else if(message instanceof SendInstallSnapshot) {
- // received from RaftActor
- setSnapshot(((SendInstallSnapshot) message).getSnapshot());
+ } else if (message instanceof SendInstallSnapshot) {
+ SendInstallSnapshot sendInstallSnapshot = (SendInstallSnapshot) message;
+ setSnapshotHolder(new SnapshotHolder(sendInstallSnapshot.getSnapshot(),
+ sendInstallSnapshot.getSnapshotBytes()));
sendInstallSnapshot();
} else if (message instanceof Replicate) {
replicate((Replicate) message);
} else if (message instanceof InstallSnapshotReply) {
handleInstallSnapshotReply((InstallSnapshotReply) message);
+ } else if (message instanceof CheckConsensusReached) {
+ possiblyUpdateCommitIndex();
} else {
return super.handleMessage(sender, message);
}
return this;
}
- private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
- LOG.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
+ @SuppressFBWarnings(value = "NP_NULL_PARAM_DEREF_ALL_TARGETS_DANGEROUS",
+ justification = "JDT nullness with SpotBugs at setSnapshotHolder(null)")
+ private void handleInstallSnapshotReply(final InstallSnapshotReply reply) {
+ log.debug("{}: handleInstallSnapshotReply: {}", logName(), reply);
String followerId = reply.getFollowerId();
FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
- if(followerLogInformation == null) {
+ if (followerLogInformation == null) {
// This can happen during AddServer if it times out.
- LOG.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
+ log.error("{}: FollowerLogInformation not found for follower {} in InstallSnapshotReply",
logName(), followerId);
return;
}
LeaderInstallSnapshotState installSnapshotState = followerLogInformation.getInstallSnapshotState();
if (installSnapshotState == null) {
- LOG.error("{}: LeaderInstallSnapshotState not found for follower {} in InstallSnapshotReply",
+ log.error("{}: LeaderInstallSnapshotState not found for follower {} in InstallSnapshotReply",
logName(), followerId);
return;
}
if (installSnapshotState.getChunkIndex() == reply.getChunkIndex()) {
boolean wasLastChunk = false;
if (reply.isSuccess()) {
- if(installSnapshotState.isLastChunk(reply.getChunkIndex())) {
+ if (installSnapshotState.isLastChunk(reply.getChunkIndex())) {
//this was the last chunk reply
- if(LOG.isDebugEnabled()) {
- LOG.debug("{}: InstallSnapshotReply received, " +
- "last chunk received, Chunk: {}. Follower: {} Setting nextIndex: {}",
- logName(), reply.getChunkIndex(), followerId,
- context.getReplicatedLog().getSnapshotIndex() + 1
- );
- }
- long followerMatchIndex = snapshot.get().getLastIncludedIndex();
+ long followerMatchIndex = snapshotHolder.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());
+ log.info("{}: Snapshot successfully installed on follower {} (last chunk {}) - "
+ + "matchIndex set to {}, nextIndex set to {}", logName(), followerId, reply.getChunkIndex(),
+ followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
if (!anyFollowersInstallingSnapshot()) {
// once there are no pending followers receiving snapshots
// we can remove snapshot from the memory
- setSnapshot(null);
+ setSnapshotHolder(null);
}
wasLastChunk = true;
- if(context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED){
+ 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");
+ log.debug("Sent message UnInitializedFollowerSnapshotReply to self");
}
} else {
installSnapshotState.markSendStatus(true);
}
} else {
- LOG.info("{}: InstallSnapshotReply received sending snapshot chunk failed, Will retry, Chunk: {}",
- logName(), reply.getChunkIndex());
+ log.warn("{}: Received failed InstallSnapshotReply - will retry: {}", logName(), reply);
installSnapshotState.markSendStatus(false);
}
- if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
- // Since the follower is now caught up try to purge the log.
- purgeInMemoryLog();
- } else if (!wasLastChunk && installSnapshotState.canSendNextChunk()) {
+ if (wasLastChunk) {
+ if (!context.getSnapshotManager().isCapturing()) {
+ // Since the follower is now caught up try to purge the log.
+ purgeInMemoryLog();
+ }
+ } else {
ActorSelection followerActor = context.getPeerActorSelection(followerId);
- if(followerActor != null) {
- sendSnapshotChunk(followerActor, followerId);
+ if (followerActor != null) {
+ sendSnapshotChunk(followerActor, followerLogInformation);
}
}
} else {
- LOG.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
+ log.error("{}: Chunk index {} in InstallSnapshotReply from follower {} does not match expected index {}",
logName(), reply.getChunkIndex(), followerId,
installSnapshotState.getChunkIndex());
- if(reply.getChunkIndex() == LeaderInstallSnapshotState.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
installSnapshotState.reset();
}
private boolean anyFollowersInstallingSnapshot() {
- for(FollowerLogInformation info: followerToLog.values()) {
- if(info.getInstallSnapshotState() != null) {
+ for (FollowerLogInformation info: followerToLog.values()) {
+ if (info.getInstallSnapshotState() != null) {
return true;
}
return false;
}
- private void replicate(Replicate replicate) {
+ private void replicate(final Replicate replicate) {
long logIndex = replicate.getReplicatedLogEntry().getIndex();
- LOG.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}", logName(),
- replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass());
+ log.debug("{}: Replicate message: identifier: {}, logIndex: {}, payload: {}, isSendImmediate: {}", logName(),
+ replicate.getIdentifier(), logIndex, replicate.getReplicatedLogEntry().getData().getClass(),
+ replicate.isSendImmediate());
// Create a tracker entry we will use this later to notify the
// client actor
- if(replicate.getClientActor() != null) {
+ if (replicate.getClientActor() != null) {
trackers.add(new ClientRequestTrackerImpl(replicate.getClientActor(), replicate.getIdentifier(),
logIndex));
}
boolean applyModificationToState = !context.anyVotingPeers()
|| context.getRaftPolicy().applyModificationToStateBeforeConsensus();
- if(applyModificationToState){
+ if (applyModificationToState) {
context.setCommitIndex(logIndex);
applyLogToStateMachine(logIndex);
}
- if (!followerToLog.isEmpty()) {
+ if (replicate.isSendImmediate() && !followerToLog.isEmpty()) {
sendAppendEntries(0, false);
}
}
- protected void sendAppendEntries(long timeSinceLastActivityInterval, boolean isHeartbeat) {
+ protected void sendAppendEntries(final long timeSinceLastActivityIntervalNanos, final boolean isHeartbeat) {
// Send an AppendEntries to all followers
for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
final String followerId = e.getKey();
final FollowerLogInformation followerLogInformation = e.getValue();
// This checks helps not to send a repeat message to the follower
- if(!followerLogInformation.isFollowerActive() ||
- followerLogInformation.timeSinceLastActivity() >= timeSinceLastActivityInterval) {
+ if (!followerLogInformation.isFollowerActive()
+ || followerLogInformation.nanosSinceLastActivity() >= timeSinceLastActivityIntervalNanos) {
sendUpdatesToFollower(followerId, followerLogInformation, true, isHeartbeat);
}
}
}
/**
- *
* This method checks if any update needs to be sent to the given follower. This includes append log entries,
* sending next snapshot chunk, and initiating a snapshot.
- * @return true if any update is sent, false otherwise
*/
-
- private void sendUpdatesToFollower(String followerId, FollowerLogInformation followerLogInformation,
- boolean sendHeartbeat, boolean isHeartbeat) {
+ private void sendUpdatesToFollower(final String followerId, final FollowerLogInformation followerLogInformation,
+ final boolean sendHeartbeat, final boolean isHeartbeat) {
ActorSelection followerActor = context.getPeerActorSelection(followerId);
if (followerActor != null) {
if (installSnapshotState != null) {
// if install snapshot is in process , then sent next chunk if possible
if (isFollowerActive && installSnapshotState.canSendNextChunk()) {
- sendSnapshotChunk(followerActor, followerId);
- } else if(sendHeartbeat) {
+ sendSnapshotChunk(followerActor, followerLogInformation);
+ } else if (sendHeartbeat) {
// we send a heartbeat even if we have not received a reply for the last chunk
sendAppendEntries = true;
}
+ } else if (followerLogInformation.isLogEntrySlicingInProgress()) {
+ sendAppendEntries = sendHeartbeat;
} else {
long leaderLastIndex = context.getReplicatedLog().lastIndex();
long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
- if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
- LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
- logName(), followerId, isFollowerActive, followerNextIndex, 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(),
+ log.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
followerNextIndex, 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());
+ if (followerLogInformation.okToReplicate()) {
+ entries = getEntriesToSend(followerLogInformation, followerActor);
sendAppendEntries = true;
}
- } else if (isFollowerActive && followerNextIndex >= 0 &&
- leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
+ } else if (isFollowerActive && followerNextIndex >= 0
+ && 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
- if (LOG.isDebugEnabled()) {
- 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));
- }
-
// Send heartbeat to follower whenever install snapshot is initiated.
sendAppendEntries = true;
if (canInstallSnapshot(followerNextIndex)) {
+ log.info("{}: Initiating install snapshot to follower {}: follower nextIndex: {}, leader "
+ + "snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(), followerId,
+ followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
+ context.getReplicatedLog().size());
+
initiateCaptureSnapshot(followerId);
+ } else {
+ // It doesn't seem like we should ever reach here - most likely indicates sonething is
+ // wrong.
+ log.info("{}: Follower {} is behind but cannot install snapshot: follower nextIndex: {}, "
+ + "leader snapshotIndex: {}, leader lastIndex: {}, leader log size: {}", logName(),
+ followerId, followerNextIndex, leaderSnapShotIndex, leaderLastIndex,
+ context.getReplicatedLog().size());
}
- } else if(sendHeartbeat) {
+ } else if (sendHeartbeat) {
// we send an AppendEntries, even if the follower is inactive
// in-order to update the followers timestamp, in case it becomes active again
sendAppendEntries = true;
}
- if(sendAppendEntries) {
- sendAppendEntriesToFollower(followerActor, followerNextIndex,
- entries, followerId);
+ if (sendAppendEntries) {
+ sendAppendEntriesToFollower(followerActor, entries, followerLogInformation);
}
}
}
- private void sendAppendEntriesToFollower(ActorSelection followerActor, long followerNextIndex,
- List<ReplicatedLogEntry> entries, String followerId) {
+ private List<ReplicatedLogEntry> getEntriesToSend(final FollowerLogInformation followerLogInfo,
+ final ActorSelection followerActor) {
+ // Try to get all the entries in the journal but not exceeding the max data size for a single AppendEntries
+ // message.
+ int maxEntries = (int) context.getReplicatedLog().size();
+ final int maxDataSize = context.getConfigParams().getSnapshotChunkSize();
+ final long followerNextIndex = followerLogInfo.getNextIndex();
+ List<ReplicatedLogEntry> entries = context.getReplicatedLog().getFrom(followerNextIndex,
+ maxEntries, maxDataSize);
+
+ // If the first entry's size exceeds the max data size threshold, it will be returned from the call above. If
+ // that is the case, then we need to slice it into smaller chunks.
+ if (!(entries.size() == 1 && entries.get(0).getData().size() > maxDataSize)) {
+ // Don't need to slice.
+ return entries;
+ }
+
+ log.debug("{}: Log entry size {} exceeds max payload size {}", logName(), entries.get(0).getData().size(),
+ maxDataSize);
+
+ // If an AppendEntries has already been serialized for the log index then reuse the
+ // SharedFileBackedOutputStream.
+ final Long logIndex = entries.get(0).getIndex();
+ SharedFileBackedOutputStream fileBackedStream = sharedSerializedAppendEntriesStreams.get(logIndex);
+ if (fileBackedStream == null) {
+ fileBackedStream = context.getFileBackedOutputStreamFactory().newSharedInstance();
+
+ final AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
+ getLogEntryIndex(followerNextIndex - 1), getLogEntryTerm(followerNextIndex - 1), entries,
+ context.getCommitIndex(), getReplicatedToAllIndex(), context.getPayloadVersion());
+
+ log.debug("{}: Serializing {} for slicing for follower {}", logName(), appendEntries,
+ followerLogInfo.getId());
+
+ try (ObjectOutputStream out = new ObjectOutputStream(fileBackedStream)) {
+ out.writeObject(appendEntries);
+ } catch (IOException e) {
+ log.error("{}: Error serializing {}", logName(), appendEntries, e);
+ fileBackedStream.cleanup();
+ return Collections.emptyList();
+ }
+
+ sharedSerializedAppendEntriesStreams.put(logIndex, fileBackedStream);
+
+ fileBackedStream.setOnCleanupCallback(index -> {
+ log.debug("{}: On SharedFileBackedOutputStream cleanup for index {}", logName(), index);
+ sharedSerializedAppendEntriesStreams.remove(index);
+ }, logIndex);
+ } else {
+ log.debug("{}: Reusing SharedFileBackedOutputStream for follower {}", logName(), followerLogInfo.getId());
+ fileBackedStream.incrementUsageCount();
+ }
+
+ log.debug("{}: Slicing stream for index {}, follower {}", logName(), logIndex, followerLogInfo.getId());
+
+ // Record that slicing is in progress for the follower.
+ followerLogInfo.setSlicedLogEntryIndex(logIndex);
+
+ final FollowerIdentifier identifier = new FollowerIdentifier(followerLogInfo.getId());
+ appendEntriesMessageSlicer.slice(SliceOptions.builder().identifier(identifier)
+ .fileBackedOutputStream(fileBackedStream).sendTo(followerActor).replyTo(actor())
+ .onFailureCallback(failure -> {
+ log.error("{}: Error slicing AppendEntries for follower {}", logName(),
+ followerLogInfo.getId(), failure);
+ followerLogInfo.setSlicedLogEntryIndex(FollowerLogInformation.NO_INDEX);
+ }).build());
+
+ return Collections.emptyList();
+ }
+
+ private void sendAppendEntriesToFollower(final ActorSelection followerActor, final List<ReplicatedLogEntry> entries,
+ final FollowerLogInformation followerLogInformation) {
+ // In certain cases outlined below we don't want to send the actual commit index to prevent the follower from
+ // possibly committing and applying conflicting entries (those with same index, different term) from a prior
+ // term that weren't replicated to a majority, which would be a violation of raft.
+ // - if the follower isn't active. In this case we don't know the state of the follower and we send an
+ // empty AppendEntries as a heart beat to prevent election.
+ // - if we're in the process of installing a snapshot. In this case we don't send any new entries but still
+ // need to send AppendEntries to prevent election.
+ // - if we're in the process of slicing an AppendEntries with a large log entry payload. In this case we
+ // need to send an empty AppendEntries to prevent election.
+ boolean isInstallingSnaphot = followerLogInformation.getInstallSnapshotState() != null;
+ long leaderCommitIndex = isInstallingSnaphot || followerLogInformation.isLogEntrySlicingInProgress()
+ || !followerLogInformation.isFollowerActive() ? -1 : context.getCommitIndex();
+
+ long followerNextIndex = followerLogInformation.getNextIndex();
AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
getLogEntryIndex(followerNextIndex - 1),
getLogEntryTerm(followerNextIndex - 1), entries,
- context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
+ leaderCommitIndex, super.getReplicatedToAllIndex(), context.getPayloadVersion(),
+ followerLogInformation.getRaftVersion(), followerLogInformation.needsLeaderAddress(getId()));
- if(!entries.isEmpty() || LOG.isTraceEnabled()) {
- LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
+ if (!entries.isEmpty() || log.isTraceEnabled()) {
+ log.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerLogInformation.getId(),
appendEntries);
}
}
/**
+ * Initiates a snapshot capture to install on a follower.
+ *
+ * <p>
* Install Snapshot works as follows
- * 1. Leader initiates the capture snapshot by sending a CaptureSnapshot message to actor
- * 2. RaftActor on receipt of the CaptureSnapshotReply (from Shard), stores the received snapshot in the replicated log
- * and makes a call to Leader's handleMessage , with SendInstallSnapshot message.
- * 3. Leader , picks the snapshot from im-mem ReplicatedLog and sends it in chunks to the Follower
- * 4. On complete, Follower sends back a InstallSnapshotReply.
- * 5. On receipt of the InstallSnapshotReply for the last chunk, Leader marks the install complete for that follower
- * and replenishes the memory by deleting the snapshot in Replicated log.
- * 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
+ * 1. Leader initiates the capture snapshot by calling createSnapshot on the RaftActor.
+ * 2. On receipt of the CaptureSnapshotReply message, the RaftActor persists the snapshot and makes a call to
+ * the Leader's handleMessage with a SendInstallSnapshot message.
+ * 3. The Leader obtains and stores the Snapshot from the SendInstallSnapshot message and sends it in chunks to
+ * the Follower via InstallSnapshot messages.
+ * 4. For each chunk, the Follower sends back an InstallSnapshotReply.
+ * 5. On receipt of the InstallSnapshotReply for the last chunk, the Leader marks the install complete for that
+ * follower.
+ * 6. If another follower requires a snapshot and a snapshot has been collected (via SendInstallSnapshot)
+ * then send the existing snapshot in chunks to the follower.
+ *
+ * @param followerId the id of the follower.
+ * @return true if capture was initiated, false otherwise.
*/
- 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.
+ public boolean initiateCaptureSnapshot(final String followerId) {
+ FollowerLogInformation followerLogInfo = followerToLog.get(followerId);
+ if (snapshotHolder.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);
+
+ // Note: sendSnapshotChunk will set the LeaderInstallSnapshotState.
+ sendSnapshotChunk(followerActor, followerLogInfo);
return true;
- } else {
- return context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
- this.getReplicatedToAllIndex(), followerId);
}
+
+ boolean captureInitiated = context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
+ this.getReplicatedToAllIndex(), followerId);
+ if (captureInitiated) {
+ followerLogInfo.setLeaderInstallSnapshotState(new LeaderInstallSnapshotState(
+ context.getConfigParams().getSnapshotChunkSize(), logName()));
+ }
+
+ return captureInitiated;
}
- private boolean canInstallSnapshot(long nextIndex){
+ private boolean canInstallSnapshot(final 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));
+ return nextIndex == -1 || !context.getReplicatedLog().isPresent(nextIndex)
+ && context.getReplicatedLog().isInSnapshot(nextIndex);
}
private void sendInstallSnapshot() {
- LOG.debug("{}: sendInstallSnapshot", logName());
+ log.debug("{}: sendInstallSnapshot", logName());
for (Entry<String, FollowerLogInformation> e : followerToLog.entrySet()) {
String followerId = e.getKey();
ActorSelection followerActor = context.getPeerActorSelection(followerId);
if (followerActor != null) {
long nextIndex = followerLogInfo.getNextIndex();
- if (context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED ||
- canInstallSnapshot(nextIndex)) {
- sendSnapshotChunk(followerActor, followerId);
+ if (followerLogInfo.getInstallSnapshotState() != null
+ || context.getPeerInfo(followerId).getVotingState() == VotingState.VOTING_NOT_INITIALIZED
+ || canInstallSnapshot(nextIndex)) {
+ sendSnapshotChunk(followerActor, followerLogInfo);
}
}
}
* Sends a snapshot chunk to a given follower
* InstallSnapshot should qualify as a heartbeat too.
*/
- private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
- try {
- if (snapshot.isPresent()) {
- byte[] nextSnapshotChunk = getNextSnapshotChunk(followerId, snapshot.get().getSnapshotBytes());
+ private void sendSnapshotChunk(final ActorSelection followerActor, final FollowerLogInformation followerLogInfo) {
+ if (snapshotHolder.isPresent()) {
+ LeaderInstallSnapshotState installSnapshotState = followerLogInfo.getInstallSnapshotState();
+ if (installSnapshotState == null) {
+ installSnapshotState = new LeaderInstallSnapshotState(context.getConfigParams().getSnapshotChunkSize(),
+ logName());
+ followerLogInfo.setLeaderInstallSnapshotState(installSnapshotState);
+ }
+
+ try {
+ // Ensure the snapshot bytes are set - this is a no-op.
+ installSnapshotState.setSnapshotBytes(snapshotHolder.get().getSnapshotBytes());
- // Note: the previous call to getNextSnapshotChunk has the side-effect of adding
- // followerId to the followerToSnapshot map.
- LeaderInstallSnapshotState installSnapshotState = followerToLog.get(followerId).getInstallSnapshotState();
+ if (!installSnapshotState.canSendNextChunk()) {
+ return;
+ }
+
+ byte[] nextSnapshotChunk = installSnapshotState.getNextChunk();
+
+ log.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerLogInfo.getId(),
+ nextSnapshotChunk.length);
int nextChunkIndex = installSnapshotState.incrementChunkIndex();
Optional<ServerConfigurationPayload> serverConfig = Optional.absent();
- if(installSnapshotState.isLastChunk(nextChunkIndex)) {
+ if (installSnapshotState.isLastChunk(nextChunkIndex)) {
serverConfig = Optional.fromNullable(context.getPeerServerInfo(true));
}
followerActor.tell(
new InstallSnapshot(currentTerm(), context.getId(),
- snapshot.get().getLastIncludedIndex(),
- snapshot.get().getLastIncludedTerm(),
+ snapshotHolder.get().getLastIncludedIndex(),
+ snapshotHolder.get().getLastIncludedTerm(),
nextSnapshotChunk,
nextChunkIndex,
installSnapshotState.getTotalChunks(),
Optional.of(installSnapshotState.getLastChunkHashCode()),
serverConfig
- ).toSerializable(followerToLog.get(followerId).getRaftVersion()),
+ ).toSerializable(followerLogInfo.getRaftVersion()),
actor()
);
- if(LOG.isDebugEnabled()) {
- LOG.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}",
- logName(), followerActor.path(), installSnapshotState.getChunkIndex(),
- installSnapshotState.getTotalChunks());
- }
+ } catch (IOException e) {
+ throw new RuntimeException(e);
}
- } catch (IOException e) {
- LOG.error("{}: InstallSnapshot failed for Leader.", logName(), e);
- }
- }
- /**
- * Acccepts snaphot as ByteString, enters into map for future chunks
- * creates and return a ByteString chunk
- */
- 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);
+ log.debug("{}: InstallSnapshot sent to follower {}, Chunk: {}/{}", logName(), followerActor.path(),
+ installSnapshotState.getChunkIndex(), installSnapshotState.getTotalChunks());
}
- byte[] nextChunk = installSnapshotState.getNextChunk();
-
- LOG.debug("{}: next snapshot chunk size for follower {}: {}", logName(), followerId, nextChunk.length);
-
- return nextChunk;
}
private void sendHeartBeat() {
if (!followerToLog.isEmpty()) {
- LOG.trace("{}: Sending heartbeat", logName());
- sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toMillis(), true);
+ log.trace("{}: Sending heartbeat", logName());
+ sendAppendEntries(context.getConfigParams().getHeartBeatInterval().toNanos(), true);
+
+ appendEntriesMessageSlicer.checkExpiredSlicedMessageState();
}
}
}
}
- private void scheduleHeartBeat(FiniteDuration interval) {
+ private void scheduleHeartBeat(final FiniteDuration interval) {
if (followerToLog.isEmpty()) {
// Optimization - do not bother scheduling a heartbeat as there are
// no followers
@Override
public void close() {
stopHeartBeat();
+ appendEntriesMessageSlicer.close();
}
@Override
int minPresent = getMinIsolatedLeaderPeerCount();
for (FollowerLogInformation followerLogInformation : followerToLog.values()) {
final PeerInfo peerInfo = context.getPeerInfo(followerLogInformation.getId());
- if(peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
+ if (peerInfo != null && peerInfo.isVoting() && followerLogInformation.isFollowerActive()) {
--minPresent;
if (minPresent == 0) {
return false;
}
@VisibleForTesting
- public FollowerLogInformation getFollower(String followerId) {
+ public FollowerLogInformation getFollower(final String followerId) {
return followerToLog.get(followerId);
}
return followerToLog.size();
}
- private static class SnapshotHolder {
+ static class SnapshotHolder {
private final long lastIncludedTerm;
private final long lastIncludedIndex;
- private final ByteString snapshotBytes;
+ private final ByteSource snapshotBytes;
- SnapshotHolder(Snapshot snapshot) {
+ SnapshotHolder(final Snapshot snapshot, final ByteSource snapshotBytes) {
this.lastIncludedTerm = snapshot.getLastAppliedTerm();
this.lastIncludedIndex = snapshot.getLastAppliedIndex();
- this.snapshotBytes = ByteString.copyFrom(snapshot.getState());
+ this.snapshotBytes = snapshotBytes;
}
long getLastIncludedTerm() {
return lastIncludedIndex;
}
- ByteString getSnapshotBytes() {
+ ByteSource getSnapshotBytes() {
return snapshotBytes;
}
}