Merge pull request #9979 from vespa-engine/vekterli/guard-state-version-publishing-with-pending-zk-writes

Do not allow states to be published when they have pending ZK writes

2 files changed, 15 insertions, 0 deletions

diff --git a/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/FleetController.java b/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/FleetController.java
index ba35243c14d..364184331a8 100644
--- a/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/FleetController.java
+++ b/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/FleetController.java
@@ -661,6 +661,12 @@ public class FleetController implements NodeStateOrHostInfoChangeHandler, NodeAd
     }
 
     private boolean broadcastClusterStateToEligibleNodes() {
+        // If there's a pending DB store we have not yet been able to store the
+        // current state bundle to ZK and must therefore _not_ allow it to be published.
+        if (database.hasPendingClusterStateMetaDataStore()) {
+            log.log(LogLevel.DEBUG, "Can't publish current cluster state as it has one or more pending ZooKeeper stores");
+            return false;
+        }
         boolean sentAny = false;
         // Give nodes a fair chance to respond first time to state gathering requests, so we don't
         // disturb system when we take over. Allow anyways if we have states from all nodes.
diff --git a/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/database/DatabaseHandler.java b/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/database/DatabaseHandler.java
index f2b1b523aba..f30b86130c2 100644
--- a/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/database/DatabaseHandler.java
+++ b/clustercontroller-core/src/main/java/com/yahoo/vespa/clustercontroller/core/database/DatabaseHandler.java
@@ -352,6 +352,15 @@ public class DatabaseHandler {
         doNextZooKeeperTask(context);
     }
 
+    // TODO should we expand this to cover _any_ pending ZK write?
+    public boolean hasPendingClusterStateMetaDataStore() {
+        synchronized (databaseMonitor) {
+            return ((zooKeeperAddress != null) &&
+                    ((pendingStore.clusterStateBundle != null) ||
+                     (pendingStore.lastSystemStateVersion != null)));
+        }
+    }
+
     public ClusterStateBundle getLatestClusterStateBundle() throws InterruptedException {
         log.log(LogLevel.DEBUG, () -> String.format("Fleetcontroller %d: Retrieving latest cluster state bundle from ZooKeeper", nodeIndex));
         synchronized (databaseMonitor) {