/** decommissioning using a exclude hosts file */
@Test
public void testDecommissionWithExcludeHosts() throws Exception {
Configuration conf = new Configuration();
conf.set("yarn.resourcemanager.nodes.exclude-path", hostFile.getAbsolutePath());
writeToHostsFile("");
rm = new MockRM(conf);
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = rm.registerNode("host2:5678", 10240);
int initialMetricCount = ClusterMetrics.getMetrics().getNumDecommisionedNMs();
HeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
writeToHostsFile("host2");
rm.getNodesListManager().refreshNodes();
nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(
"The decommisioned metrics are not updated",
NodeAction.SHUTDOWN.equals(nodeHeartbeat.getNodeAction()));
checkDecommissionedNMCount(rm, ++initialMetricCount);
}
/** Decommissioning using a pre-configured include hosts file */
@Test
public void testDecommissionWithIncludeHosts() throws Exception {
writeToHostsFile("localhost", "host1", "host2");
Configuration conf = new Configuration();
conf.set(YarnConfiguration.RM_NODES_INCLUDE_FILE_PATH, hostFile.getAbsolutePath());
rm = new MockRM(conf);
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = rm.registerNode("host2:5678", 10240);
MockNM nm3 = rm.registerNode("localhost:4433", 1024);
ClusterMetrics metrics = ClusterMetrics.getMetrics();
assert (metrics != null);
int metricCount = metrics.getNumDecommisionedNMs();
NodeHeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm3.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
// To test that IPs also work
String ip = NetUtils.normalizeHostName("localhost");
writeToHostsFile("host1", ip);
rm.getNodesListManager().refreshNodes(conf);
checkDecommissionedNMCount(rm, ++metricCount);
nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
Assert.assertEquals(1, ClusterMetrics.getMetrics().getNumDecommisionedNMs());
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(
"Node is not decommisioned.", NodeAction.SHUTDOWN.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm3.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
Assert.assertEquals(metricCount, ClusterMetrics.getMetrics().getNumDecommisionedNMs());
}
@Test
public void testReboot() throws Exception {
Configuration conf = new Configuration();
rm = new MockRM(conf);
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = rm.registerNode("host2:1234", 2048);
int initialMetricCount = ClusterMetrics.getMetrics().getNumRebootedNMs();
NodeHeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat =
nm2.nodeHeartbeat(new HashMap<ApplicationId, List<ContainerStatus>>(), true, -100);
Assert.assertTrue(NodeAction.RESYNC.equals(nodeHeartbeat.getNodeAction()));
Assert.assertEquals(
"Too far behind rm response id:0 nm response id:-100",
nodeHeartbeat.getDiagnosticsMessage());
checkRebootedNMCount(rm, ++initialMetricCount);
}
@Test
public void testReboot() throws Exception {
Configuration conf = new Configuration();
rm = new MockRM(conf);
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = new MockNM("host2:1234", 2048, rm.getResourceTrackerService());
int initialMetricCount = ClusterMetrics.getMetrics().getNumRebootedNMs();
HeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat =
nm2.nodeHeartbeat(
new HashMap<ApplicationId, List<ContainerStatus>>(),
true,
recordFactory.newRecordInstance(NodeId.class));
Assert.assertTrue(NodeAction.REBOOT.equals(nodeHeartbeat.getNodeAction()));
checkRebootedNMCount(rm, ++initialMetricCount);
}
@Test
public void testNMUnregistration() throws Exception {
Configuration conf = new Configuration();
rm = new MockRM(conf);
rm.start();
ResourceTrackerService resourceTrackerService = rm.getResourceTrackerService();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
int shutdownNMsCount = ClusterMetrics.getMetrics().getNumShutdownNMs();
NodeHeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
UnRegisterNodeManagerRequest request = Records.newRecord(UnRegisterNodeManagerRequest.class);
request.setNodeId(nm1.getNodeId());
resourceTrackerService.unRegisterNodeManager(request);
checkShutdownNMCount(rm, ++shutdownNMsCount);
// The RM should remove the node after unregistration, hence send a reboot
// command.
nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.RESYNC.equals(nodeHeartbeat.getNodeAction()));
}
@Test
public void testReconnectNode() throws Exception {
final DrainDispatcher dispatcher = new DrainDispatcher();
rm =
new MockRM() {
@Override
protected EventHandler<SchedulerEvent> createSchedulerEventDispatcher() {
return new SchedulerEventDispatcher(this.scheduler) {
@Override
public void handle(SchedulerEvent event) {
scheduler.handle(event);
}
};
}
@Override
protected Dispatcher createDispatcher() {
return dispatcher;
}
};
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = rm.registerNode("host2:5678", 5120);
nm1.nodeHeartbeat(true);
nm2.nodeHeartbeat(false);
dispatcher.await();
checkUnealthyNMCount(rm, nm2, true, 1);
final int expectedNMs = ClusterMetrics.getMetrics().getNumActiveNMs();
QueueMetrics metrics = rm.getResourceScheduler().getRootQueueMetrics();
// TODO Metrics incorrect in case of the FifoScheduler
Assert.assertEquals(5120, metrics.getAvailableMB());
// reconnect of healthy node
nm1 = rm.registerNode("host1:1234", 5120);
NodeHeartbeatResponse response = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(response.getNodeAction()));
dispatcher.await();
Assert.assertEquals(expectedNMs, ClusterMetrics.getMetrics().getNumActiveNMs());
checkUnealthyNMCount(rm, nm2, true, 1);
// reconnect of unhealthy node
nm2 = rm.registerNode("host2:5678", 5120);
response = nm2.nodeHeartbeat(false);
Assert.assertTrue(NodeAction.NORMAL.equals(response.getNodeAction()));
dispatcher.await();
Assert.assertEquals(expectedNMs, ClusterMetrics.getMetrics().getNumActiveNMs());
checkUnealthyNMCount(rm, nm2, true, 1);
// unhealthy node changed back to healthy
nm2 = rm.registerNode("host2:5678", 5120);
dispatcher.await();
response = nm2.nodeHeartbeat(true);
response = nm2.nodeHeartbeat(true);
dispatcher.await();
Assert.assertEquals(5120 + 5120, metrics.getAvailableMB());
// reconnect of node with changed capability
nm1 = rm.registerNode("host2:5678", 10240);
dispatcher.await();
response = nm1.nodeHeartbeat(true);
dispatcher.await();
Assert.assertTrue(NodeAction.NORMAL.equals(response.getNodeAction()));
Assert.assertEquals(5120 + 10240, metrics.getAvailableMB());
// reconnect of node with changed capability and running applications
List<ApplicationId> runningApps = new ArrayList<ApplicationId>();
runningApps.add(ApplicationId.newInstance(1, 0));
nm1 = rm.registerNode("host2:5678", 15360, 2, runningApps);
dispatcher.await();
response = nm1.nodeHeartbeat(true);
dispatcher.await();
Assert.assertTrue(NodeAction.NORMAL.equals(response.getNodeAction()));
Assert.assertEquals(5120 + 15360, metrics.getAvailableMB());
// reconnect healthy node changing http port
nm1 = new MockNM("host1:1234", 5120, rm.getResourceTrackerService());
nm1.setHttpPort(3);
nm1.registerNode();
dispatcher.await();
response = nm1.nodeHeartbeat(true);
response = nm1.nodeHeartbeat(true);
dispatcher.await();
RMNode rmNode = rm.getRMContext().getRMNodes().get(nm1.getNodeId());
Assert.assertEquals(3, rmNode.getHttpPort());
Assert.assertEquals(5120, rmNode.getTotalCapability().getMemory());
Assert.assertEquals(5120 + 15360, metrics.getAvailableMB());
}
/** Decommissioning using a pre-configured exclude hosts file */
@Test
public void testDecommissionWithExcludeHosts() throws Exception {
Configuration conf = new Configuration();
conf.set(YarnConfiguration.RM_NODES_EXCLUDE_FILE_PATH, hostFile.getAbsolutePath());
writeToHostsFile("");
final DrainDispatcher dispatcher = new DrainDispatcher();
rm =
new MockRM(conf) {
@Override
protected Dispatcher createDispatcher() {
return dispatcher;
}
};
rm.start();
MockNM nm1 = rm.registerNode("host1:1234", 5120);
MockNM nm2 = rm.registerNode("host2:5678", 10240);
MockNM nm3 = rm.registerNode("localhost:4433", 1024);
dispatcher.await();
int metricCount = ClusterMetrics.getMetrics().getNumDecommisionedNMs();
NodeHeartbeatResponse nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
dispatcher.await();
// To test that IPs also work
String ip = NetUtils.normalizeHostName("localhost");
writeToHostsFile("host2", ip);
rm.getNodesListManager().refreshNodes(conf);
checkDecommissionedNMCount(rm, metricCount + 2);
nodeHeartbeat = nm1.nodeHeartbeat(true);
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm2.nodeHeartbeat(true);
Assert.assertTrue(
"The decommisioned metrics are not updated",
NodeAction.SHUTDOWN.equals(nodeHeartbeat.getNodeAction()));
nodeHeartbeat = nm3.nodeHeartbeat(true);
Assert.assertTrue(
"The decommisioned metrics are not updated",
NodeAction.SHUTDOWN.equals(nodeHeartbeat.getNodeAction()));
dispatcher.await();
writeToHostsFile("");
rm.getNodesListManager().refreshNodes(conf);
nm3 = rm.registerNode("localhost:4433", 1024);
dispatcher.await();
nodeHeartbeat = nm3.nodeHeartbeat(true);
dispatcher.await();
Assert.assertTrue(NodeAction.NORMAL.equals(nodeHeartbeat.getNodeAction()));
// decommissined node is 1 since 1 node is rejoined after updating exclude
// file
checkDecommissionedNMCount(rm, metricCount + 1);
}
