import akka.actor.ActorSelection;
import akka.actor.Cancellable;
import com.google.common.base.Preconditions;
+import com.google.protobuf.ByteString;
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.messages.RequestVoteReply;
import scala.concurrent.duration.FiniteDuration;
+import java.io.IOException;
import java.util.ArrayList;
import java.util.Collections;
import java.util.HashMap;
public class Leader extends AbstractRaftActorBehavior {
- private final Map<String, FollowerLogInformation> followerToLog =
+ protected final Map<String, FollowerLogInformation> followerToLog =
new HashMap();
+ protected final Map<String, FollowerToSnapshot> mapFollowerToSnapshot = new HashMap<>();
private final Set<String> followers;
private Cancellable heartbeatSchedule = null;
- private Cancellable appendEntriesSchedule = null;
private Cancellable installSnapshotSchedule = null;
private List<ClientRequestTracker> trackerList = new ArrayList<>();
public Leader(RaftActorContext context) {
super(context);
- if (lastIndex() >= 0) {
- context.setCommitIndex(lastIndex());
- }
-
followers = context.getPeerAddresses().keySet();
for (String followerId : followers) {
FollowerLogInformation followerLogInformation =
new FollowerLogInformationImpl(followerId,
- new AtomicLong(lastIndex()),
+ new AtomicLong(context.getCommitIndex()),
new AtomicLong(-1));
followerToLog.put(followerId, followerLogInformation);
}
- context.getLogger().debug("Election:Leader has following peers:"+ followers);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("Election:Leader has following peers: {}", followers);
+ }
if (followers.size() > 0) {
minReplicationCount = (followers.size() + 1) / 2 + 1;
}
- @Override protected RaftState handleAppendEntries(ActorRef sender,
+ @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
AppendEntries appendEntries) {
- context.getLogger().debug(appendEntries.toString());
+ if(LOG.isDebugEnabled()) {
+ LOG.debug(appendEntries.toString());
+ }
- return state();
+ return this;
}
- @Override protected RaftState handleAppendEntriesReply(ActorRef sender,
+ @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
AppendEntriesReply appendEntriesReply) {
if(! appendEntriesReply.isSuccess()) {
- context.getLogger()
- .debug(appendEntriesReply.toString());
+ if(LOG.isDebugEnabled()) {
+ LOG.debug(appendEntriesReply.toString());
+ }
}
// Update the FollowerLogInformation
followerToLog.get(followerId);
if(followerLogInformation == null){
- context.getLogger().error("Unknown follower {}", followerId);
- return state();
+ LOG.error("Unknown follower {}", followerId);
+ return this;
}
if (appendEntriesReply.isSuccess()) {
applyLogToStateMachine(context.getCommitIndex());
}
- return state();
+ return this;
+ }
+
+ protected ClientRequestTracker removeClientRequestTracker(long logIndex) {
+
+ ClientRequestTracker toRemove = findClientRequestTracker(logIndex);
+ if(toRemove != null) {
+ trackerList.remove(toRemove);
+ }
+
+ return toRemove;
}
protected ClientRequestTracker findClientRequestTracker(long logIndex) {
return null;
}
- @Override protected RaftState handleRequestVoteReply(ActorRef sender,
+ @Override protected RaftActorBehavior handleRequestVoteReply(ActorRef sender,
RequestVoteReply requestVoteReply) {
- return state();
+ return this;
}
@Override public RaftState state() {
return RaftState.Leader;
}
- @Override public RaftState handleMessage(ActorRef sender, Object originalMessage) {
+ @Override public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
Preconditions.checkNotNull(sender, "sender should not be null");
Object message = fromSerializableMessage(originalMessage);
// This applies to all RPC messages and responses
if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
- return RaftState.Follower;
+
+ return switchBehavior(new Follower(context));
}
}
try {
if (message instanceof SendHeartBeat) {
- return sendHeartBeat();
+ sendHeartBeat();
+ return this;
} else if(message instanceof SendInstallSnapshot) {
installSnapshotIfNeeded();
} else if (message instanceof Replicate) {
return super.handleMessage(sender, message);
}
- private void handleInstallSnapshotReply(InstallSnapshotReply message) {
- InstallSnapshotReply reply = message;
+ private void handleInstallSnapshotReply(InstallSnapshotReply reply) {
String followerId = reply.getFollowerId();
- FollowerLogInformation followerLogInformation =
- followerToLog.get(followerId);
+ FollowerToSnapshot followerToSnapshot =
+ mapFollowerToSnapshot.get(followerId);
+
+ if (followerToSnapshot != null &&
+ followerToSnapshot.getChunkIndex() == reply.getChunkIndex()) {
+
+ if (reply.isSuccess()) {
+ if(followerToSnapshot.isLastChunk(reply.getChunkIndex())) {
+ //this was the last chunk reply
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("InstallSnapshotReply received, " +
+ "last chunk received, Chunk:{}. Follower:{} Setting nextIndex:{}",
+ reply.getChunkIndex(), followerId,
+ context.getReplicatedLog().getSnapshotIndex() + 1
+ );
+ }
+
+ FollowerLogInformation followerLogInformation =
+ followerToLog.get(followerId);
+ followerLogInformation.setMatchIndex(
+ context.getReplicatedLog().getSnapshotIndex());
+ followerLogInformation.setNextIndex(
+ context.getReplicatedLog().getSnapshotIndex() + 1);
+ mapFollowerToSnapshot.remove(followerId);
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("followerToLog.get(followerId).getNextIndex().get()=" +
+ followerToLog.get(followerId).getNextIndex().get());
+ }
+
+ } else {
+ followerToSnapshot.markSendStatus(true);
+ }
+ } else {
+ LOG.info("InstallSnapshotReply received, " +
+ "sending snapshot chunk failed, Will retry, Chunk:{}",
+ reply.getChunkIndex()
+ );
+ followerToSnapshot.markSendStatus(false);
+ }
- followerLogInformation
- .setMatchIndex(context.getReplicatedLog().getSnapshotIndex());
- followerLogInformation
- .setNextIndex(context.getReplicatedLog().getSnapshotIndex() + 1);
+ } else {
+ LOG.error("ERROR!!" +
+ "FollowerId in InstallSnapshotReply not known to Leader" +
+ " or Chunk Index in InstallSnapshotReply not matching {} != {}",
+ followerToSnapshot.getChunkIndex(), reply.getChunkIndex()
+ );
+ }
}
private void replicate(Replicate replicate) {
long logIndex = replicate.getReplicatedLogEntry().getIndex();
- context.getLogger().debug("Replicate message " + logIndex);
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("Replicate message {}", logIndex);
+ }
// Create a tracker entry we will use this later to notify the
// client actor
private void sendAppendEntries() {
// Send an AppendEntries to all followers
for (String followerId : followers) {
- ActorSelection followerActor =
- context.getPeerActorSelection(followerId);
+ ActorSelection followerActor = context.getPeerActorSelection(followerId);
if (followerActor != null) {
- FollowerLogInformation followerLogInformation =
- followerToLog.get(followerId);
-
- long nextIndex = followerLogInformation.getNextIndex().get();
-
+ FollowerLogInformation followerLogInformation = followerToLog.get(followerId);
+ long followerNextIndex = followerLogInformation.getNextIndex().get();
List<ReplicatedLogEntry> entries = Collections.emptyList();
- if (context.getReplicatedLog().isPresent(nextIndex)) {
- // FIXME : Sending one entry at a time
- entries =
- context.getReplicatedLog().getFrom(nextIndex, 1);
+ if (mapFollowerToSnapshot.get(followerId) != null) {
+ if (mapFollowerToSnapshot.get(followerId).canSendNextChunk()) {
+ sendSnapshotChunk(followerActor, followerId);
+ }
+
+ } else {
+
+ if (context.getReplicatedLog().isPresent(followerNextIndex)) {
+ // FIXME : Sending one entry at a time
+ entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
+
+ followerActor.tell(
+ new AppendEntries(currentTerm(), context.getId(),
+ prevLogIndex(followerNextIndex),
+ prevLogTerm(followerNextIndex), entries,
+ context.getCommitIndex()).toSerializable(),
+ actor()
+ );
+
+ } else {
+ // if the followers next index is not present in the leaders log, then snapshot should be sent
+ long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
+ long leaderLastIndex = context.getReplicatedLog().lastIndex();
+ if (followerNextIndex >= 0 && leaderLastIndex >= followerNextIndex ) {
+ // if the follower is just not starting and leader's index
+ // is more than followers index
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("SendInstallSnapshot to follower:{}," +
+ "follower-nextIndex:{}, leader-snapshot-index:{}, " +
+ "leader-last-index:{}", followerId,
+ followerNextIndex, leaderSnapShotIndex, leaderLastIndex
+ );
+ }
+
+ actor().tell(new SendInstallSnapshot(), actor());
+ } else {
+ followerActor.tell(
+ new AppendEntries(currentTerm(), context.getId(),
+ prevLogIndex(followerNextIndex),
+ prevLogTerm(followerNextIndex), entries,
+ context.getCommitIndex()).toSerializable(),
+ actor()
+ );
+ }
+ }
}
-
- followerActor.tell(
- new AppendEntries(currentTerm(), context.getId(),
- prevLogIndex(nextIndex),
- prevLogTerm(nextIndex), entries,
- context.getCommitIndex()).toSerializable(),
- actor()
- );
}
}
}
long nextIndex = followerLogInformation.getNextIndex().get();
- if (!context.getReplicatedLog().isPresent(nextIndex) && context
- .getReplicatedLog().isInSnapshot(nextIndex)) {
- followerActor.tell(
- new InstallSnapshot(currentTerm(), context.getId(),
- context.getReplicatedLog().getSnapshotIndex(),
- context.getReplicatedLog().getSnapshotTerm(),
- context.getReplicatedLog().getSnapshot()
- ),
- actor()
- );
+ if (!context.getReplicatedLog().isPresent(nextIndex) &&
+ context.getReplicatedLog().isInSnapshot(nextIndex)) {
+ sendSnapshotChunk(followerActor, followerId);
}
}
}
}
- private RaftState sendHeartBeat() {
+ /**
+ * Sends a snapshot chunk to a given follower
+ * InstallSnapshot should qualify as a heartbeat too.
+ */
+ private void sendSnapshotChunk(ActorSelection followerActor, String followerId) {
+ try {
+ followerActor.tell(
+ new InstallSnapshot(currentTerm(), context.getId(),
+ context.getReplicatedLog().getSnapshotIndex(),
+ context.getReplicatedLog().getSnapshotTerm(),
+ getNextSnapshotChunk(followerId,
+ context.getReplicatedLog().getSnapshot()),
+ mapFollowerToSnapshot.get(followerId).incrementChunkIndex(),
+ mapFollowerToSnapshot.get(followerId).getTotalChunks()
+ ).toSerializable(),
+ actor()
+ );
+ LOG.info("InstallSnapshot sent to follower {}, Chunk: {}/{}",
+ followerActor.path(), mapFollowerToSnapshot.get(followerId).getChunkIndex(),
+ mapFollowerToSnapshot.get(followerId).getTotalChunks());
+ } catch (IOException e) {
+ LOG.error(e, "InstallSnapshot failed for Leader.");
+ }
+ }
+
+ /**
+ * 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);
+ }
+ ByteString nextChunk = followerToSnapshot.getNextChunk();
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("Leader's snapshot nextChunk size:{}", nextChunk.size());
+ }
+
+ return nextChunk;
+ }
+
+ private void sendHeartBeat() {
if (followers.size() > 0) {
sendAppendEntries();
}
- return state();
}
private void stopHeartBeat() {
return context.getId();
}
+ /**
+ * Encapsulates the snapshot bytestring and handles the logic of sending
+ * snapshot chunks
+ */
+ protected class FollowerToSnapshot {
+ private 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 int totalChunks;
+
+ public FollowerToSnapshot(ByteString snapshotBytes) {
+ this.snapshotBytes = snapshotBytes;
+ replyReceivedForOffset = -1;
+ chunkIndex = 1;
+ 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:{}",
+ size, totalChunks);
+ }
+ }
+
+ 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;
+ } 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;
+ }
+ }
+
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("length={}, offset={},size={}",
+ snapshotLength, start, size);
+ }
+ return getSnapshotBytes().substring(start, start + size);
+
+ }
+ }
+
}