Bug in AbstractLeader replication consensus

[controller.git] / opendaylight / md-sal / sal-akka-raft / src / main / java / org / opendaylight / controller / cluster / raft / behaviors / AbstractLeader.java

diff --git a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java
index f1a178d0cac9fb51737b449edf7cbc3593cd745f..31ee9d2a7a61d3e2fa10fc6eb65544dbf7907b86 100644 (file)
--- a/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java
+++ b/opendaylight/md-sal/sal-akka-raft/src/main/java/org/opendaylight/controller/cluster/raft/behaviors/AbstractLeader.java
@@ -261,11 +261,18 @@ public abstract class AbstractLeader extends AbstractRaftActorBehavior {
 
             if (replicatedCount >= minReplicationCount) {
                 ReplicatedLogEntry replicatedLogEntry = context.getReplicatedLog().get(N);
-                if (replicatedLogEntry != null && replicatedLogEntry.getTerm() == currentTerm()) {
-                    context.setCommitIndex(N);
-                } else {
+                if (replicatedLogEntry == null) {
                     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()) {
+                    context.setCommitIndex(N);
+                }
             } else {
                 break;
             }