diff options
author | Tor Brede Vekterli <vekterli@verizonmedia.com> | 2019-07-05 16:42:39 +0200 |
---|---|---|
committer | Tor Brede Vekterli <vekterli@verizonmedia.com> | 2019-07-05 16:42:39 +0200 |
commit | e363dc115148782e882ae9789344f2fa4997c120 (patch) | |
tree | 15ff3ca123d0e3cca0fbbb38ab7a788832eadfaa | |
parent | 23303b94e954b29d10350a5ad3dc585daac95a48 (diff) |
Do not allow states to be published when they have pending ZK writes
Avoids a race condition where a bundle ZK write fails but we have not yet
detected that ZK connectivity has been lost. This could lead to violating
the invariant that published state versions are strictly increasing.
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) { |