@Test
public void testMailboxTracker() throws Exception {
ZooKeeper zk = getClient(0);
MailboxTracker tracker = new MailboxTracker(zk, handler);
MailboxPublisher publisher = new MailboxPublisher(VoltZK.mailboxes + "/1");
VoltZK.createPersistentZKNodes(zk);
publisher.registerMailbox(MailboxType.ExecutionSite, new MailboxNodeContent(1L, 0));
publisher.registerMailbox(MailboxType.ExecutionSite, new MailboxNodeContent(2L, 1));
publisher.publish(zk);
// start the mailbox tracker and watch all the changes
tracker.start();
while (handler.m_handleCount.get() == 0) {
Thread.sleep(1);
}
Map<MailboxType, List<MailboxNodeContent>> value = handler.m_mailboxes;
assertTrue(value.containsKey(MailboxType.ExecutionSite));
List<MailboxNodeContent> list = value.get(MailboxType.ExecutionSite);
assertEquals(2, list.size());
MailboxNodeContent node1 = list.get(0);
assertEquals(1, node1.HSId.longValue());
assertEquals(0, node1.partitionId.intValue());
MailboxNodeContent node2 = list.get(1);
assertEquals(2, node2.HSId.longValue());
assertEquals(1, node2.partitionId.intValue());
tracker.shutdown();
}
@Test
public void testUpdate() throws Exception {
ZooKeeper zk = getClient(0);
ZooKeeper zk2 = getClient(0);
MailboxTracker tracker = new MailboxTracker(zk, handler);
MailboxPublisher publisher = new MailboxPublisher(VoltZK.mailboxes + "/1");
VoltZK.createPersistentZKNodes(zk);
publisher.registerMailbox(MailboxType.ExecutionSite, new MailboxNodeContent(1L, 0));
publisher.publish(zk2);
publisher = new MailboxPublisher(VoltZK.mailboxes + "/2");
publisher.registerMailbox(MailboxType.ExecutionSite, new MailboxNodeContent(2L, 1));
publisher.publish(zk);
tracker.start();
// The ephemaral node just created will disappear and we should get an update
zk2.close();
while (handler.m_handleCount.get() < 2) {
Thread.sleep(1);
}
Map<MailboxType, List<MailboxNodeContent>> value = handler.m_mailboxes;
assertTrue(value.containsKey(MailboxType.ExecutionSite));
List<MailboxNodeContent> list = value.get(MailboxType.ExecutionSite);
assertEquals(1, list.size());
assertEquals(2, list.get(0).HSId.longValue());
assertEquals(1, list.get(0).partitionId.intValue());
tracker.shutdown();
}
@Override
public void run(List<String> children) {
List<Long> updatedHSIds = VoltZK.childrenToReplicaHSIds(children);
// compute previously unseen HSId set in the callback list
Set<Long> newHSIds = new HashSet<Long>(updatedHSIds);
newHSIds.removeAll(m_replicas);
tmLog.debug("Newly seen replicas: " + CoreUtils.hsIdCollectionToString(newHSIds));
// compute previously seen but now vanished from the callback list HSId set
Set<Long> missingHSIds = new HashSet<Long>(m_replicas);
missingHSIds.removeAll(updatedHSIds);
tmLog.debug("Newly dead replicas: " + CoreUtils.hsIdCollectionToString(missingHSIds));
tmLog.debug(
"Handling babysitter callback for partition "
+ m_partitionId
+ ": children: "
+ CoreUtils.hsIdCollectionToString(updatedHSIds));
if (m_state.get() == AppointerState.CLUSTER_START) {
// We can't yet tolerate a host failure during startup. Crash it all
if (missingHSIds.size() > 0) {
VoltDB.crashGlobalVoltDB("Node failure detected during startup.", false, null);
}
// ENG-3166: Eventually we would like to get rid of the extra replicas beyond k_factor,
// but for now we just look to see how many replicas of this partition we actually expect
// and gate leader assignment on that many copies showing up.
int replicaCount = m_kfactor + 1;
JSONArray parts;
try {
parts = m_topo.getJSONArray("partitions");
for (int p = 0; p < parts.length(); p++) {
JSONObject aPartition = parts.getJSONObject(p);
int pid = aPartition.getInt("partition_id");
if (pid == m_partitionId) {
replicaCount = aPartition.getJSONArray("replicas").length();
}
}
} catch (JSONException e) {
// Ignore and just assume the normal number of replicas
}
if (children.size() == replicaCount) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
} else {
tmLog.info(
"Waiting on "
+ ((m_kfactor + 1) - children.size())
+ " more nodes "
+ "for k-safety before startup");
}
} else {
// Check for k-safety
if (!isClusterKSafe()) {
VoltDB.crashGlobalVoltDB(
"Some partitions have no replicas. Cluster has become unviable.", false, null);
}
// Check if replay has completed
if (m_replayComplete.get() == false) {
VoltDB.crashGlobalVoltDB(
"Detected node failure during command log replay. Cluster will shut down.",
false,
null);
}
// Check to see if there's been a possible network partition and we're not already handling
// it
if (m_partitionDetectionEnabled && !m_partitionDetected) {
doPartitionDetectionActivities();
}
// If we survived the above gauntlet of fail, appoint a new leader for this partition.
if (missingHSIds.contains(m_currentLeader)) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
}
}
m_replicas.clear();
m_replicas.addAll(updatedHSIds);
}
