Bug 2187: Bootstrap EOS shard when no local shards configured
[controller.git] / opendaylight / md-sal / sal-akka-raft / src / main / java / org / opendaylight / controller / cluster / raft / behaviors / Follower.java
index 31b5efbe3878df73f74d46d90d018751dca67bbf..0149b57e6e94f8ddf076a406f14eea1546d2aa1e 100644 (file)
@@ -10,11 +10,11 @@ package org.opendaylight.controller.cluster.raft.behaviors;
 
 import akka.actor.ActorRef;
 import com.google.common.annotations.VisibleForTesting;
-import com.google.protobuf.ByteString;
 import java.util.ArrayList;
 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.ServerConfigurationPayload;
 import org.opendaylight.controller.cluster.raft.Snapshot;
 import org.opendaylight.controller.cluster.raft.base.messages.ApplySnapshot;
 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
@@ -23,6 +23,7 @@ import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
 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;
 
 /**
@@ -39,10 +40,28 @@ public class Follower extends AbstractRaftActorBehavior {
 
     private SnapshotTracker snapshotTracker = null;
 
+    private final SyncStatusTracker initialSyncStatusTracker;
+
+    private static final int SYNC_THRESHOLD = 10;
+
     public Follower(RaftActorContext context) {
-        super(context);
+        this(context, null);
+    }
+
+    public Follower(RaftActorContext context, String initialLeaderId) {
+        super(context, RaftState.Follower);
+        leaderId = initialLeaderId;
+
+        initialSyncStatusTracker = new SyncStatusTracker(context.getActor(), getId(), SYNC_THRESHOLD);
+
+        if(context.getRaftPolicy().automaticElectionsEnabled()) {
+            if (context.getPeerIds().isEmpty() && getLeaderId() == null) {
+                actor().tell(ELECTION_TIMEOUT, actor());
+            } else {
+                scheduleElection(electionDuration());
+            }
+        }
 
-        scheduleElection(electionDuration());
     }
 
     private boolean isLogEntryPresent(long index){
@@ -72,99 +91,62 @@ public class Follower extends AbstractRaftActorBehavior {
         return -1;
     }
 
+    private void updateInitialSyncStatus(long currentLeaderCommit, String leaderId){
+        initialSyncStatusTracker.update(leaderId, currentLeaderCommit, context.getCommitIndex());
+    }
+
     @Override protected RaftActorBehavior handleAppendEntries(ActorRef sender,
                                                               AppendEntries appendEntries) {
 
-        if(appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: handleAppendEntries: {}", context.getId(), appendEntries);
-            }
+        int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
+        if(LOG.isTraceEnabled()) {
+            LOG.trace("{}: handleAppendEntries: {}", logName(), appendEntries);
+        } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
+            LOG.debug("{}: handleAppendEntries: {}", logName(), appendEntries);
         }
 
         // TODO : Refactor this method into a bunch of smaller methods
         // to make it easier to read. Before refactoring ensure tests
         // cover the code properly
 
-        // 1. Reply false if term < currentTerm (§5.1)
-        // This is handled in the appendEntries method of the base class
-
-        // If we got here then we do appear to be talking to the leader
-        leaderId = appendEntries.getLeaderId();
-
-        // 2. Reply false if log doesn’t contain an entry at prevLogIndex
-        // whose term matches prevLogTerm (§5.3)
-
-        long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
-        boolean prevEntryPresent = isLogEntryPresent(appendEntries.getPrevLogIndex());
-
-
-        boolean outOfSync = true;
-
-        // First check if the logs are in sync or not
-        if (lastIndex() == -1
-                && appendEntries.getPrevLogIndex() != -1) {
-
-            // The follower's log is out of sync because the leader does have
-            // an entry at prevLogIndex and this follower has no entries in
-            // it's log.
+        if (snapshotTracker != null || context.getSnapshotManager().isApplying()) {
+            // if snapshot install is in progress, follower should just acknowledge append entries with a reply.
+            AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
+                    lastIndex(), lastTerm(), context.getPayloadVersion());
 
             if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: The followers log is empty and the senders prevLogIndex is {}",
-                        context.getId(), appendEntries.getPrevLogIndex());
+                LOG.debug("{}: snapshot install is in progress, replying immediately with {}", logName(), reply);
             }
+            sender.tell(reply, actor());
 
-        } else if (lastIndex() > -1
-                && appendEntries.getPrevLogIndex() != -1
-                && !prevEntryPresent) {
-
-            // The follower's log is out of sync because the Leader's
-            // prevLogIndex entry was not found in it's log
-
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: The log is not empty but the prevLogIndex {} was not found in it",
-                        context.getId(), appendEntries.getPrevLogIndex());
-            }
+            return this;
+        }
 
-        } else if (lastIndex() > -1
-                && prevEntryPresent
-                && prevLogTerm != appendEntries.getPrevLogTerm()) {
+        // If we got here then we do appear to be talking to the leader
+        leaderId = appendEntries.getLeaderId();
 
-            // The follower's log is out of sync because the Leader's
-            // prevLogIndex entry does exist in the follower's log but it has
-            // a different term in it
+        setLeaderPayloadVersion(appendEntries.getPayloadVersion());
 
-            if (LOG.isDebugEnabled()) {
-                LOG.debug(
-                        "{}: Cannot append entries because previous entry term {}  is not equal to append entries prevLogTerm {}"
-                        , context.getId(), prevLogTerm
-                        , appendEntries.getPrevLogTerm());
-            }
-        } else {
-            outOfSync = false;
-        }
+        updateInitialSyncStatus(appendEntries.getLeaderCommit(), appendEntries.getLeaderId());
+        // First check if the logs are in sync or not
+        long lastIndex = lastIndex();
 
-        if (outOfSync) {
+        if (isOutOfSync(appendEntries)) {
             // We found that the log was out of sync so just send a negative
             // reply and return
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: Follower ({}) is out-of-sync, " +
-                        "so sending negative reply, lastIndex():{}, lastTerm():{}",
-                        context.getId(), context.getId(), lastIndex(), lastTerm()
-                );
-            }
-            sender.tell(
-                new AppendEntriesReply(context.getId(), currentTerm(), false,
-                    lastIndex(), lastTerm()), actor()
-            );
+
+            LOG.debug("{}: Follower is out-of-sync, so sending negative reply, lastIndex: {}, lastTerm: {}",
+                        logName(), lastIndex, lastTerm());
+
+            sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
+                    lastTerm(), context.getPayloadVersion()), actor());
             return this;
         }
 
-        if (appendEntries.getEntries() != null
-            && appendEntries.getEntries().size() > 0) {
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: Number of entries to be appended = {}", context.getId(),
+        if (appendEntries.getEntries() != null && appendEntries.getEntries().size() > 0) {
+
+            LOG.debug("{}: Number of entries to be appended = {}", logName(),
                         appendEntries.getEntries().size());
-            }
 
             // 3. If an existing entry conflicts with a new one (same index
             // but different terms), delete the existing entry and all that
@@ -172,7 +154,7 @@ public class Follower extends AbstractRaftActorBehavior {
             int addEntriesFrom = 0;
             if (context.getReplicatedLog().size() > 0) {
 
-                // Find the entry up until which the one that is not in the follower's log
+                // Find the entry up until the one that is not in the follower's log
                 for (int i = 0;i < appendEntries.getEntries().size(); i++, addEntriesFrom++) {
                     ReplicatedLogEntry matchEntry = appendEntries.getEntries().get(i);
                     ReplicatedLogEntry newEntry = context.getReplicatedLog().get(matchEntry.getIndex());
@@ -182,81 +164,144 @@ public class Follower extends AbstractRaftActorBehavior {
                         break;
                     }
 
-                    if (newEntry.getTerm() == matchEntry
-                        .getTerm()) {
+                    if (newEntry.getTerm() == matchEntry.getTerm()) {
                         continue;
                     }
 
-                    if(LOG.isDebugEnabled()) {
-                        LOG.debug("{}: Removing entries from log starting at {}", context.getId(),
+                    if(!context.getRaftPolicy().applyModificationToStateBeforeConsensus()) {
+
+                        LOG.debug("{}: Removing entries from log starting at {}", logName(),
                                 matchEntry.getIndex());
-                    }
 
-                    // Entries do not match so remove all subsequent entries
-                    context.getReplicatedLog()
-                        .removeFromAndPersist(matchEntry.getIndex());
-                    break;
+                        // Entries do not match so remove all subsequent entries
+                        context.getReplicatedLog().removeFromAndPersist(matchEntry.getIndex());
+                        break;
+                    } else {
+                        sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), false, lastIndex,
+                                lastTerm(), context.getPayloadVersion(), true), actor());
+                        return this;
+                    }
                 }
             }
 
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: After cleanup entries to be added from = {}", context.getId(),
-                        (addEntriesFrom + lastIndex()));
-            }
+            lastIndex = lastIndex();
+            LOG.debug("{}: After cleanup entries to be added from = {}", logName(),
+                        (addEntriesFrom + lastIndex));
 
             // 4. Append any new entries not already in the log
-            for (int i = addEntriesFrom;
-                 i < appendEntries.getEntries().size(); i++) {
+            for (int i = addEntriesFrom; i < appendEntries.getEntries().size(); i++) {
+                ReplicatedLogEntry entry = appendEntries.getEntries().get(i);
+
+                LOG.debug("{}: Append entry to log {}", logName(), entry.getData());
+
+                context.getReplicatedLog().appendAndPersist(entry);
 
-                if(LOG.isDebugEnabled()) {
-                    LOG.debug("{}: Append entry to log {}", context.getId(),
-                            appendEntries.getEntries().get(i).getData());
+                if(entry.getData() instanceof ServerConfigurationPayload) {
+                    context.updatePeerIds((ServerConfigurationPayload)entry.getData());
                 }
-                context.getReplicatedLog().appendAndPersist(appendEntries.getEntries().get(i));
             }
 
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: Log size is now {}", context.getId(), context.getReplicatedLog().size());
-            }
+            LOG.debug("{}: Log size is now {}", logName(), context.getReplicatedLog().size());
         }
 
-
         // 5. If leaderCommit > commitIndex, set commitIndex =
         // min(leaderCommit, index of last new entry)
 
+        lastIndex = lastIndex();
         long prevCommitIndex = context.getCommitIndex();
 
-        context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(),
-            context.getReplicatedLog().lastIndex()));
+        context.setCommitIndex(Math.min(appendEntries.getLeaderCommit(), lastIndex));
 
         if (prevCommitIndex != context.getCommitIndex()) {
-            if(LOG.isDebugEnabled()) {
-                LOG.debug("{}: Commit index set to {}", context.getId(), context.getCommitIndex());
-            }
+            LOG.debug("{}: Commit index set to {}", logName(), context.getCommitIndex());
         }
 
         // If commitIndex > lastApplied: increment lastApplied, apply
         // log[lastApplied] to state machine (§5.3)
         // check if there are any entries to be applied. last-applied can be equal to last-index
         if (appendEntries.getLeaderCommit() > context.getLastApplied() &&
-            context.getLastApplied() < lastIndex()) {
+            context.getLastApplied() < lastIndex) {
             if(LOG.isDebugEnabled()) {
                 LOG.debug("{}: applyLogToStateMachine, " +
-                        "appendEntries.getLeaderCommit():{}," +
-                        "context.getLastApplied():{}, lastIndex():{}", context.getId(),
-                    appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex()
-                );
+                        "appendEntries.getLeaderCommit(): {}," +
+                        "context.getLastApplied(): {}, lastIndex(): {}", logName(),
+                    appendEntries.getLeaderCommit(), context.getLastApplied(), lastIndex);
             }
 
             applyLogToStateMachine(appendEntries.getLeaderCommit());
         }
 
-        sender.tell(new AppendEntriesReply(context.getId(), currentTerm(), true,
-            lastIndex(), lastTerm()), actor());
+        AppendEntriesReply reply = new AppendEntriesReply(context.getId(), currentTerm(), true,
+            lastIndex, lastTerm(), context.getPayloadVersion());
+
+        if(LOG.isTraceEnabled()) {
+            LOG.trace("{}: handleAppendEntries returning : {}", logName(), reply);
+        } else if(LOG.isDebugEnabled() && numLogEntries > 0) {
+            LOG.debug("{}: handleAppendEntries returning : {}", logName(), reply);
+        }
+
+        sender.tell(reply, actor());
+
+        if (!context.getSnapshotManager().isCapturing()) {
+            super.performSnapshotWithoutCapture(appendEntries.getReplicatedToAllIndex());
+        }
 
         return this;
     }
 
+    private boolean isOutOfSync(AppendEntries appendEntries) {
+
+        long prevLogTerm = getLogEntryTerm(appendEntries.getPrevLogIndex());
+        boolean prevEntryPresent = isLogEntryPresent(appendEntries.getPrevLogIndex());
+        long lastIndex = lastIndex();
+        int numLogEntries = appendEntries.getEntries() != null ? appendEntries.getEntries().size() : 0;
+        boolean outOfSync = true;
+
+        if (lastIndex == -1 && appendEntries.getPrevLogIndex() != -1) {
+
+            // The follower's log is out of sync because the leader does have
+            // an entry at prevLogIndex and this follower has no entries in
+            // it's log.
+
+            LOG.debug("{}: The followers log is empty and the senders prevLogIndex is {}",
+                        logName(), appendEntries.getPrevLogIndex());
+        } else if (lastIndex > -1 && appendEntries.getPrevLogIndex() != -1 && !prevEntryPresent) {
+
+            // The follower's log is out of sync because the Leader's
+            // prevLogIndex entry was not found in it's log
+
+            LOG.debug("{}: The log is not empty but the prevLogIndex {} was not found in it",
+                        logName(), appendEntries.getPrevLogIndex());
+        } else if (lastIndex > -1 && prevEntryPresent && prevLogTerm != appendEntries.getPrevLogTerm()) {
+
+            // The follower's log is out of sync because the Leader's
+            // prevLogIndex entry does exist in the follower's log but it has
+            // a different term in it
+
+            LOG.debug(
+                    "{}: Cannot append entries because previous entry term {}  is not equal to append entries prevLogTerm {}",
+                    logName(), prevLogTerm, appendEntries.getPrevLogTerm());
+        } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
+                && appendEntries.getReplicatedToAllIndex() != -1
+                && !isLogEntryPresent(appendEntries.getReplicatedToAllIndex())) {
+            // This append entry comes from a leader who has it's log aggressively trimmed and so does not have
+            // the previous entry in it's in-memory journal
+
+            LOG.debug(
+                    "{}: Cannot append entries because the replicatedToAllIndex {} does not appear to be in the in-memory journal",
+                    logName(), appendEntries.getReplicatedToAllIndex());
+        } else if(appendEntries.getPrevLogIndex() == -1 && appendEntries.getPrevLogTerm() == -1
+                && appendEntries.getReplicatedToAllIndex() != -1 && numLogEntries > 0 &&
+                !isLogEntryPresent(appendEntries.getEntries().get(0).getIndex() - 1)){
+            LOG.debug(
+                    "{}: Cannot append entries because the calculated previousIndex {} was not found in the in-memory journal",
+                    logName(), appendEntries.getEntries().get(0).getIndex() - 1);
+        } else {
+            outOfSync = false;
+        }
+        return outOfSync;
+    }
+
     @Override protected RaftActorBehavior handleAppendEntriesReply(ActorRef sender,
         AppendEntriesReply appendEntriesReply) {
         return this;
@@ -267,10 +312,6 @@ public class Follower extends AbstractRaftActorBehavior {
         return this;
     }
 
-    @Override public RaftState state() {
-        return RaftState.Follower;
-    }
-
     @Override public RaftActorBehavior handleMessage(ActorRef sender, Object originalMessage) {
 
         Object message = fromSerializableMessage(originalMessage);
@@ -281,37 +322,47 @@ public class Follower extends AbstractRaftActorBehavior {
             // set currentTerm = T, convert to follower (§5.1)
             // This applies to all RPC messages and responses
             if (rpc.getTerm() > context.getTermInformation().getCurrentTerm()) {
+                LOG.debug("{}: Term {} in \"{}\" message is greater than follower's term {} - updating term",
+                        logName(), rpc.getTerm(), rpc, context.getTermInformation().getCurrentTerm());
+
                 context.getTermInformation().updateAndPersist(rpc.getTerm(), null);
             }
         }
 
         if (message instanceof ElectionTimeout) {
-            return switchBehavior(new Candidate(context));
+            LOG.debug("{}: Received ElectionTimeout - switching to Candidate", logName());
+            return internalSwitchBehavior(RaftState.Candidate);
 
         } else if (message instanceof InstallSnapshot) {
             InstallSnapshot installSnapshot = (InstallSnapshot) message;
             handleInstallSnapshot(sender, installSnapshot);
         }
 
-        scheduleElection(electionDuration());
+        if(message instanceof RaftRPC && (!(message instanceof RequestVote) || (canGrantVote((RequestVote) message)))){
+            scheduleElection(electionDuration());
+        }
 
         return super.handleMessage(sender, message);
     }
 
-    private void handleInstallSnapshot(ActorRef sender, InstallSnapshot installSnapshot) {
+    private void handleInstallSnapshot(final ActorRef sender, InstallSnapshot installSnapshot) {
 
-        if(LOG.isDebugEnabled()) {
-            LOG.debug("{}: InstallSnapshot received by follower " +
-                    "datasize:{} , Chunk:{}/{}", context.getId(), installSnapshot.getData().size(),
-                installSnapshot.getChunkIndex(), installSnapshot.getTotalChunks()
-            );
-        }
+        LOG.debug("{}: InstallSnapshot received from leader {}, datasize: {} , Chunk: {}/{}",
+                    logName(), installSnapshot.getLeaderId(), installSnapshot.getData().size(),
+                    installSnapshot.getChunkIndex(), installSnapshot.getTotalChunks());
+
+        leaderId = installSnapshot.getLeaderId();
 
         if(snapshotTracker == null){
             snapshotTracker = new SnapshotTracker(LOG, installSnapshot.getTotalChunks());
         }
 
+        updateInitialSyncStatus(installSnapshot.getLastIncludedIndex(), installSnapshot.getLeaderId());
+
         try {
+            final InstallSnapshotReply reply = new InstallSnapshotReply(
+                    currentTerm(), context.getId(), installSnapshot.getChunkIndex(), true);
+
             if(snapshotTracker.addChunk(installSnapshot.getChunkIndex(), installSnapshot.getData(),
                     installSnapshot.getLastChunkHashCode())){
                 Snapshot snapshot = Snapshot.create(snapshotTracker.getSnapshot(),
@@ -319,26 +370,43 @@ public class Follower extends AbstractRaftActorBehavior {
                         installSnapshot.getLastIncludedIndex(),
                         installSnapshot.getLastIncludedTerm(),
                         installSnapshot.getLastIncludedIndex(),
-                        installSnapshot.getLastIncludedTerm());
+                        installSnapshot.getLastIncludedTerm(),
+                        context.getTermInformation().getCurrentTerm(),
+                        context.getTermInformation().getVotedFor(),
+                        context.getPeerServerInfo());
 
-                actor().tell(new ApplySnapshot(snapshot), actor());
+                ApplySnapshot.Callback applySnapshotCallback = new ApplySnapshot.Callback() {
+                    @Override
+                    public void onSuccess() {
+                        LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 
-                snapshotTracker = null;
+                        sender.tell(reply, actor());
+                    }
 
-            }
+                    @Override
+                    public void onFailure() {
+                        sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(), -1, false), actor());
+                    }
+                };
+
+                actor().tell(new ApplySnapshot(snapshot, applySnapshotCallback), actor());
 
-            sender.tell(new InstallSnapshotReply(
-                    currentTerm(), context.getId(), installSnapshot.getChunkIndex(),
-                    true), actor());
+                snapshotTracker = null;
+            } else {
+                LOG.debug("{}: handleInstallSnapshot returning: {}", logName(), reply);
 
+                sender.tell(reply, actor());
+            }
         } catch (SnapshotTracker.InvalidChunkException e) {
+            LOG.debug("{}: Exception in InstallSnapshot of follower", logName(), e);
 
             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
                     -1, false), actor());
             snapshotTracker = null;
 
         } catch (Exception e){
-            LOG.error(e, "{}: Exception in InstallSnapshot of follower", context.getId());
+            LOG.error("{}: Exception in InstallSnapshot of follower", logName(), e);
+
             //send reply with success as false. The chunk will be sent again on failure
             sender.tell(new InstallSnapshotReply(currentTerm(), context.getId(),
                     installSnapshot.getChunkIndex(), false), actor());
@@ -346,14 +414,13 @@ public class Follower extends AbstractRaftActorBehavior {
         }
     }
 
-    @Override public void close() throws Exception {
+    @Override
+    public void close() throws Exception {
         stopElection();
     }
 
     @VisibleForTesting
-    ByteString getSnapshotChunksCollected(){
-        return snapshotTracker != null ? snapshotTracker.getCollectedChunks() : ByteString.EMPTY;
+    SnapshotTracker getSnapshotTracker(){
+        return snapshotTracker;
     }
-
-
 }