Bug 2787: Batch AppendEntries to speed up follower sync
[controller.git] / opendaylight / md-sal / sal-akka-raft / src / main / java / org / opendaylight / controller / cluster / raft / behaviors / AbstractLeader.java
index a63c62fa30740b5830676ab6f15f3de9c1988e7b..2eb3b32c6f13d0c01db85fe7fa141a1eb6c2aac9 100644 (file)
@@ -33,7 +33,6 @@ import org.opendaylight.controller.cluster.raft.FollowerLogInformationImpl;
 import org.opendaylight.controller.cluster.raft.RaftActorContext;
 import org.opendaylight.controller.cluster.raft.RaftState;
 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
-import org.opendaylight.controller.cluster.raft.base.messages.CaptureSnapshot;
 import org.opendaylight.controller.cluster.raft.base.messages.Replicate;
 import org.opendaylight.controller.cluster.raft.base.messages.SendHeartBeat;
 import org.opendaylight.controller.cluster.raft.base.messages.SendInstallSnapshot;
@@ -179,26 +178,33 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
         }
 
         followerLogInformation.markFollowerActive();
+        followerLogInformation.setPayloadVersion(appendEntriesReply.getPayloadVersion());
 
         boolean updated = false;
         if (appendEntriesReply.isSuccess()) {
-            updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
-            updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
-
-            if(updated && LOG.isDebugEnabled()) {
-                LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}", logName(),
-                        followerId, followerLogInformation.getMatchIndex(), followerLogInformation.getNextIndex());
-            }
+            updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
         } else {
             LOG.debug("{}: handleAppendEntriesReply: received unsuccessful reply: {}", logName(), appendEntriesReply);
 
-            // TODO: When we find that the follower is out of sync with the
-            // Leader we simply decrement that followers next index by 1.
-            // Would it be possible to do better than this? The RAFT spec
-            // does not explicitly deal with it but may be something for us to
-            // think about
+            long followerLastLogIndex = appendEntriesReply.getLogLastIndex();
+            ReplicatedLogEntry followersLastLogEntry = context.getReplicatedLog().get(followerLastLogIndex);
+            if(followerLastLogIndex < 0 || (followersLastLogEntry != null &&
+                    followersLastLogEntry.getTerm() == appendEntriesReply.getLogLastTerm())) {
+                // The follower's log is empty or the last entry is present in the leader's journal
+                // and the terms match so the follower is just behind the leader's journal from
+                // the last snapshot, if any. We'll catch up the follower quickly by starting at the
+                // follower's last log index.
 
-            followerLogInformation.decrNextIndex();
+                updated = updateFollowerLogInformation(followerLogInformation, appendEntriesReply);
+            } else {
+                // TODO: When we find that the follower is out of sync with the
+                // Leader we simply decrement that followers next index by 1.
+                // Would it be possible to do better than this? The RAFT spec
+                // does not explicitly deal with it but may be something for us to
+                // think about.
+
+                followerLogInformation.decrNextIndex();
+            }
         }
 
         // Now figure out if this reply warrants a change in the commitIndex
@@ -235,7 +241,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
             applyLogToStateMachine(context.getCommitIndex());
         }
 
-        if (!context.isSnapshotCaptureInitiated()) {
+        if (!context.getSnapshotManager().isCapturing()) {
             purgeInMemoryLog();
         }
 
@@ -244,6 +250,19 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
         return this;
     }
 
+    private boolean updateFollowerLogInformation(FollowerLogInformation followerLogInformation,
+            AppendEntriesReply appendEntriesReply) {
+        boolean updated = followerLogInformation.setMatchIndex(appendEntriesReply.getLogLastIndex());
+        updated = followerLogInformation.setNextIndex(appendEntriesReply.getLogLastIndex() + 1) || updated;
+
+        if(updated && LOG.isDebugEnabled()) {
+            LOG.debug("{}: handleAppendEntriesReply - FollowerLogInformation for {} updated: matchIndex: {}, nextIndex: {}",
+                    logName(), followerLogInformation.getId(), followerLogInformation.getMatchIndex(),
+                    followerLogInformation.getNextIndex());
+        }
+        return updated;
+    }
+
     private void purgeInMemoryLog() {
         //find the lowest index across followers which has been replicated to all.
         // lastApplied if there are no followers, so that we keep clearing the log for single-node
@@ -388,7 +407,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
                 followerToSnapshot.markSendStatus(false);
             }
 
-            if (wasLastChunk && !context.isSnapshotCaptureInitiated()) {
+            if (wasLastChunk && !context.getSnapshotManager().isCapturing()) {
                 // Since the follower is now caught up try to purge the log.
                 purgeInMemoryLog();
             } else if (!wasLastChunk && followerToSnapshot.canSendNextChunk()) {
@@ -461,7 +480,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
             long followerNextIndex = followerLogInformation.getNextIndex();
             boolean isFollowerActive = followerLogInformation.isFollowerActive();
             boolean sendAppendEntries = false;
-            List<ReplicatedLogEntry> entries = Collections.EMPTY_LIST;
+            List<ReplicatedLogEntry> entries = Collections.emptyList();
 
             if (mapFollowerToSnapshot.get(followerId) != null) {
                 // if install snapshot is in process , then sent next chunk if possible
@@ -476,8 +495,8 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
                 long leaderSnapShotIndex = context.getReplicatedLog().getSnapshotIndex();
 
                 if((!isHeartbeat && LOG.isDebugEnabled()) || LOG.isTraceEnabled()) {
-                    LOG.debug("{}: Checking sendAppendEntries for follower {}, followerNextIndex {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
-                            logName(), followerId, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
+                    LOG.debug("{}: Checking sendAppendEntries for follower {}: active: {}, followerNextIndex: {}, leaderLastIndex: {}, leaderSnapShotIndex: {}",
+                            logName(), followerId, isFollowerActive, followerNextIndex, leaderLastIndex, leaderSnapShotIndex);
                 }
 
                 if (isFollowerActive && context.getReplicatedLog().isPresent(followerNextIndex)) {
@@ -485,13 +504,16 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
                     LOG.debug("{}: sendAppendEntries: {} is present for follower {}", logName(),
                             followerNextIndex, followerId);
 
-                    // FIXME : Sending one entry at a time
                     if(followerLogInformation.okToReplicate()) {
-                        entries = context.getReplicatedLog().getFrom(followerNextIndex, 1);
+                        // Try to send all the entries in the journal but not exceeding the max data size
+                        // for a single AppendEntries message.
+                        int maxEntries = (int) context.getReplicatedLog().size();
+                        entries = context.getReplicatedLog().getFrom(followerNextIndex, maxEntries,
+                                context.getConfigParams().getSnapshotChunkSize());
                         sendAppendEntries = true;
                     }
                 } else if (isFollowerActive && followerNextIndex >= 0 &&
-                    leaderLastIndex > followerNextIndex && !context.isSnapshotCaptureInitiated()) {
+                    leaderLastIndex > followerNextIndex && !context.getSnapshotManager().isCapturing()) {
                     // if the followers next index is not present in the leaders log, and
                     // if the follower is just not starting and if leader's index is more than followers index
                     // then snapshot should be sent
@@ -527,7 +549,7 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
         AppendEntries appendEntries = new AppendEntries(currentTerm(), context.getId(),
             prevLogIndex(followerNextIndex),
             prevLogTerm(followerNextIndex), entries,
-            context.getCommitIndex(), super.getReplicatedToAllIndex());
+            context.getCommitIndex(), super.getReplicatedToAllIndex(), context.getPayloadVersion());
 
         if(!entries.isEmpty() || LOG.isTraceEnabled()) {
             LOG.debug("{}: Sending AppendEntries to follower {}: {}", logName(), followerId,
@@ -562,37 +584,10 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
                 final ActorSelection followerActor = context.getPeerActorSelection(followerId);
                 sendSnapshotChunk(followerActor, followerId);
 
-            } else if (!context.isSnapshotCaptureInitiated()) {
 
-                ReplicatedLogEntry lastAppliedEntry = context.getReplicatedLog().get(context.getLastApplied());
-                long lastAppliedIndex = -1;
-                long lastAppliedTerm = -1;
-
-                if (lastAppliedEntry != null) {
-                    lastAppliedIndex = lastAppliedEntry.getIndex();
-                    lastAppliedTerm = lastAppliedEntry.getTerm();
-                } else if (context.getReplicatedLog().getSnapshotIndex() > -1) {
-                    lastAppliedIndex = context.getReplicatedLog().getSnapshotIndex();
-                    lastAppliedTerm = context.getReplicatedLog().getSnapshotTerm();
-                }
-
-                boolean isInstallSnapshotInitiated = true;
-                long replicatedToAllIndex = super.getReplicatedToAllIndex();
-                ReplicatedLogEntry replicatedToAllEntry = context.getReplicatedLog().get(replicatedToAllIndex);
-
-                CaptureSnapshot captureSnapshot = new CaptureSnapshot(
-                        lastIndex(), lastTerm(), lastAppliedIndex, lastAppliedTerm,
-                        (replicatedToAllEntry != null ? replicatedToAllEntry.getIndex() : -1),
-                        (replicatedToAllEntry != null ? replicatedToAllEntry.getTerm() : -1),
-                        isInstallSnapshotInitiated);
-
-                if(LOG.isDebugEnabled()) {
-                    LOG.debug("{}: Initiating install snapshot to follower {}: {}", logName(), followerId,
-                            captureSnapshot);
-                }
-
-                actor().tell(captureSnapshot, actor());
-                context.setSnapshotCaptureInitiated(true);
+            } else {
+                context.getSnapshotManager().captureToInstall(context.getReplicatedLog().last(),
+                        this.getReplicatedToAllIndex(), followerId);
             }
         }
     }