Examples with ZKHelixDataAccessor org.apache.helix.manager.zk.ZKHelixDataAccessor used on opensource projects

Example 26 with ZKHelixDataAccessor

use of org.apache.helix.manager.zk.ZKHelixDataAccessor in project helix by apache.

the class TestPauseSignal method testPauseSignal.

@Test()
public void testPauseSignal() throws Exception {
    // Logger.getRootLogger().setLevel(Level.INFO);
    String className = TestHelper.getTestClassName();
    String methodName = TestHelper.getTestMethodName();
    final String clusterName = className + "_" + methodName;
    System.out.println("START " + clusterName + " at " + new Date(System.currentTimeMillis()));
    MockParticipantManager[] participants = new MockParticipantManager[5];
    // participant port
    TestHelper.setupCluster(// participant port
    clusterName, // participant port
    ZK_ADDR, // participant port
    12918, // participant name prefix
    "localhost", // resource name prefix
    "TestDB", // resources
    1, // partitions per resource
    10, // number of nodes
    5, // replicas
    3, "MasterSlave", // do rebalance
    true);
    // start controller
    ClusterControllerManager controller = new ClusterControllerManager(ZK_ADDR, clusterName, "controller_0");
    controller.syncStart();
    // start participants
    for (int i = 0; i < 5; i++) {
        String instanceName = "localhost_" + (12918 + i);
        participants[i] = new MockParticipantManager(ZK_ADDR, clusterName, instanceName);
        participants[i].syncStart();
    }
    boolean result = ClusterStateVerifier.verifyByZkCallback(new BestPossAndExtViewZkVerifier(ZK_ADDR, clusterName));
    Assert.assertTrue(result);
    // pause the cluster and make sure pause is persistent
    ZkClient zkClient = new ZkClient(ZK_ADDR);
    zkClient.setZkSerializer(new ZNRecordSerializer());
    final HelixDataAccessor tmpAccessor = new ZKHelixDataAccessor(clusterName, new ZkBaseDataAccessor<ZNRecord>(zkClient));
    String cmd = "-zkSvr " + ZK_ADDR + " -enableCluster " + clusterName + " false";
    ClusterSetup.processCommandLineArgs(cmd.split(" "));
    tmpAccessor.setProperty(tmpAccessor.keyBuilder().pause(), new PauseSignal("pause"));
    zkClient.close();
    // wait for controller to be signaled by pause
    Thread.sleep(1000);
    // add a new resource group
    ClusterSetup setupTool = new ClusterSetup(ZK_ADDR);
    setupTool.addResourceToCluster(clusterName, "TestDB1", 10, "MasterSlave");
    setupTool.rebalanceStorageCluster(clusterName, "TestDB1", 3);
    // make sure TestDB1 external view is empty
    TestHelper.verifyWithTimeout("verifyEmptyCurStateAndExtView", 1000, clusterName, "TestDB1", TestHelper.<String>setOf("localhost_12918", "localhost_12919", "localhost_12920", "localhost_12921", "localhost_12922"), ZK_ADDR);
    // resume controller
    final HelixDataAccessor accessor = new ZKHelixDataAccessor(clusterName, new ZkBaseDataAccessor<ZNRecord>(_gZkClient));
    cmd = "-zkSvr " + ZK_ADDR + " -enableCluster " + clusterName + " true";
    ClusterSetup.processCommandLineArgs(cmd.split(" "));
    result = ClusterStateVerifier.verifyByZkCallback(new BestPossAndExtViewZkVerifier(ZK_ADDR, clusterName));
    Assert.assertTrue(result);
    // clean up
    controller.syncStop();
    for (int i = 0; i < 5; i++) {
        participants[i].syncStop();
    }
    System.out.println("END " + clusterName + " at " + new Date(System.currentTimeMillis()));
}

Example 27 with ZKHelixDataAccessor

use of org.apache.helix.manager.zk.ZKHelixDataAccessor in project helix by apache.

the class TestRebalancePipeline method testMsgTriggeredRebalance.

@Test
public void testMsgTriggeredRebalance() throws Exception {
    String clusterName = "CLUSTER_" + _className + "_msgTrigger";
    System.out.println("START " + clusterName + " at " + new Date(System.currentTimeMillis()));
    HelixDataAccessor accessor = new ZKHelixDataAccessor(clusterName, new ZkBaseDataAccessor<ZNRecord>(_gZkClient));
    HelixManager manager = new DummyClusterManager(clusterName, accessor);
    ClusterEvent event = new ClusterEvent(ClusterEventType.Unknown);
    refreshClusterConfig(clusterName, accessor);
    final String resourceName = "testResource_dup";
    String[] resourceGroups = new String[] { resourceName };
    TestHelper.setupEmptyCluster(_gZkClient, clusterName);
    // ideal state: node0 is MASTER, node1 is SLAVE
    // replica=2 means 1 master and 1 slave
    setupIdealState(clusterName, new int[] { 0, 1 }, resourceGroups, 1, 2);
    setupStateModel(clusterName);
    setupInstances(clusterName, new int[] { 0, 1 });
    setupLiveInstances(clusterName, new int[] { 0, 1 });
    ClusterControllerManager controller = new ClusterControllerManager(ZK_ADDR, clusterName, "controller_0");
    controller.syncStart();
    // round1: controller sends O->S to both node0 and node1
    Thread.sleep(1000);
    Builder keyBuilder = accessor.keyBuilder();
    List<String> messages = accessor.getChildNames(keyBuilder.messages("localhost_0"));
    Assert.assertEquals(messages.size(), 1);
    messages = accessor.getChildNames(keyBuilder.messages("localhost_1"));
    Assert.assertEquals(messages.size(), 1);
    // round2: node0 and node1 update current states but not removing messages
    // Since controller's rebalancer pipeline will GC pending messages after timeout, and both hosts
    // update current states to SLAVE, controller will send out rebalance message to
    // have one host to become master
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_0", "SLAVE", true);
    setCurrentState(clusterName, "localhost_1", resourceName, resourceName + "_0", "session_1", "SLAVE", true);
    // Controller has 3s timeout, so after 1s, controller should not have GCed message
    Thread.sleep(1000);
    Assert.assertEquals(accessor.getChildValues(keyBuilder.messages("localhost_0")).size(), 1);
    Assert.assertEquals(accessor.getChildValues(keyBuilder.messages("localhost_1")).size(), 1);
    // After another 2 second, controller should cleanup messages and continue to rebalance
    Thread.sleep(3000);
    // Manually trigger another rebalance by touching current state
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_0", "SLAVE");
    Thread.sleep(1000);
    List<Message> host0Msg = accessor.getChildValues(keyBuilder.messages("localhost_0"));
    List<Message> host1Msg = accessor.getChildValues(keyBuilder.messages("localhost_1"));
    List<Message> allMsgs = new ArrayList<>(host0Msg);
    allMsgs.addAll(host1Msg);
    Assert.assertEquals(allMsgs.size(), 1);
    Assert.assertEquals(allMsgs.get(0).getToState(), "MASTER");
    Assert.assertEquals(allMsgs.get(0).getFromState(), "SLAVE");
    // round3: node0 changes state to master, but failed to delete message,
    // controller will clean it up
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_0", "MASTER", true);
    Thread.sleep(3500);
    // touch current state to trigger rebalance
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_0", "MASTER", false);
    Thread.sleep(1000);
    Assert.assertTrue(accessor.getChildNames(keyBuilder.messages("localhost_0")).isEmpty());
    // round4: node0 has duplicated but valid message, i.e. there is a P2P message sent to it
    // due to error in the triggered pipeline, controller should remove duplicated message
    // immediately as the partition has became master 3 sec ago (there is already a timeout)
    Message sourceMsg = allMsgs.get(0);
    Message dupMsg = new Message(sourceMsg.getMsgType(), UUID.randomUUID().toString());
    dupMsg.getRecord().setSimpleFields(sourceMsg.getRecord().getSimpleFields());
    dupMsg.getRecord().setListFields(sourceMsg.getRecord().getListFields());
    dupMsg.getRecord().setMapFields(sourceMsg.getRecord().getMapFields());
    accessor.setProperty(dupMsg.getKey(accessor.keyBuilder(), dupMsg.getTgtName()), dupMsg);
    Thread.sleep(1000);
    messages = accessor.getChildNames(keyBuilder.messages("localhost_0"));
    Assert.assertTrue(messages.isEmpty());
    // round5: node0 has completely invalid message, controller should immediately delete it
    dupMsg.setFromState("SLAVE");
    dupMsg.setToState("OFFLINE");
    accessor.setProperty(dupMsg.getKey(accessor.keyBuilder(), dupMsg.getTgtName()), dupMsg);
    Thread.sleep(1000);
    messages = accessor.getChildNames(keyBuilder.messages("localhost_0"));
    Assert.assertTrue(messages.isEmpty());
    System.out.println("END " + clusterName + " at " + new Date(System.currentTimeMillis()));
}

Example 28 with ZKHelixDataAccessor

use of org.apache.helix.manager.zk.ZKHelixDataAccessor in project helix by apache.

the class TestRebalancePipeline method testDuplicateMsg.

@Test
public void testDuplicateMsg() {
    String clusterName = "CLUSTER_" + _className + "_dup";
    System.out.println("START " + clusterName + " at " + new Date(System.currentTimeMillis()));
    HelixDataAccessor accessor = new ZKHelixDataAccessor(clusterName, new ZkBaseDataAccessor<ZNRecord>(_gZkClient));
    refreshClusterConfig(clusterName, accessor);
    HelixManager manager = new DummyClusterManager(clusterName, accessor);
    ClusterEvent event = new ClusterEvent(ClusterEventType.Unknown);
    event.addAttribute(AttributeName.helixmanager.name(), manager);
    final String resourceName = "testResource_dup";
    String[] resourceGroups = new String[] { resourceName };
    // ideal state: node0 is MASTER, node1 is SLAVE
    // replica=2 means 1 master and 1 slave
    setupIdealState(clusterName, new int[] { 0 }, resourceGroups, 1, 1);
    setupLiveInstances(clusterName, new int[] { 0 });
    setupStateModel(clusterName);
    // cluster data cache refresh pipeline
    Pipeline dataRefresh = new Pipeline();
    dataRefresh.addStage(new ReadClusterDataStage());
    // rebalance pipeline
    Pipeline rebalancePipeline = new Pipeline();
    rebalancePipeline.addStage(new ResourceComputationStage());
    rebalancePipeline.addStage(new CurrentStateComputationStage());
    rebalancePipeline.addStage(new BestPossibleStateCalcStage());
    rebalancePipeline.addStage(new IntermediateStateCalcStage());
    rebalancePipeline.addStage(new MessageGenerationPhase());
    rebalancePipeline.addStage(new MessageSelectionStage());
    rebalancePipeline.addStage(new MessageThrottleStage());
    rebalancePipeline.addStage(new TaskAssignmentStage());
    // round1: set node0 currentState to OFFLINE
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_0", "OFFLINE");
    runPipeline(event, dataRefresh);
    runPipeline(event, rebalancePipeline);
    MessageSelectionStageOutput msgSelOutput = event.getAttribute(AttributeName.MESSAGES_SELECTED.name());
    List<Message> messages = msgSelOutput.getMessages(resourceName, new Partition(resourceName + "_0"));
    Assert.assertEquals(messages.size(), 1, "Should output 1 message: OFFLINE-SLAVE for node0");
    Message message = messages.get(0);
    Assert.assertEquals(message.getFromState(), "OFFLINE");
    Assert.assertEquals(message.getToState(), "SLAVE");
    Assert.assertEquals(message.getTgtName(), "localhost_0");
    // round2: updates node0 currentState to SLAVE but keep the
    // message, make sure controller should not send S->M until removal is done
    setCurrentState(clusterName, "localhost_0", resourceName, resourceName + "_0", "session_1", "SLAVE");
    runPipeline(event, dataRefresh);
    refreshClusterConfig(clusterName, accessor);
    runPipeline(event, rebalancePipeline);
    msgSelOutput = event.getAttribute(AttributeName.MESSAGES_SELECTED.name());
    messages = msgSelOutput.getMessages(resourceName, new Partition(resourceName + "_0"));
    Assert.assertEquals(messages.size(), 0, "Should NOT output 1 message: SLAVE-MASTER for node1");
    System.out.println("END " + clusterName + " at " + new Date(System.currentTimeMillis()));
}

Example 29 with ZKHelixDataAccessor

use of org.apache.helix.manager.zk.ZKHelixDataAccessor in project helix by apache.

the class TestStateTransitionThrottle method testTransitionThrottleOnErrorPartition.

@Test
public void testTransitionThrottleOnErrorPartition() throws Exception {
    String clusterName = getShortClassName() + "testMaxErrorPartition";
    MockParticipantManager[] participants = new MockParticipantManager[participantCount];
    System.out.println("START " + clusterName + " at " + new Date(System.currentTimeMillis()));
    final ZKHelixDataAccessor accessor = new ZKHelixDataAccessor(clusterName, new ZkBaseDataAccessor<ZNRecord>(_gZkClient));
    setupCluster(clusterName, accessor);
    // Set throttle config to enable throttling
    PropertyKey.Builder keyBuilder = accessor.keyBuilder();
    ClusterConfig clusterConfig = accessor.getProperty(accessor.keyBuilder().clusterConfig());
    clusterConfig.setResourcePriorityField("Name");
    List<StateTransitionThrottleConfig> throttleConfigs = new ArrayList<>();
    throttleConfigs.add(new StateTransitionThrottleConfig(StateTransitionThrottleConfig.RebalanceType.LOAD_BALANCE, StateTransitionThrottleConfig.ThrottleScope.CLUSTER, 100));
    throttleConfigs.add(new StateTransitionThrottleConfig(StateTransitionThrottleConfig.RebalanceType.RECOVERY_BALANCE, StateTransitionThrottleConfig.ThrottleScope.CLUSTER, 100));
    clusterConfig.setStateTransitionThrottleConfigs(throttleConfigs);
    accessor.setProperty(keyBuilder.clusterConfig(), clusterConfig);
    // set one partition to be always Error, so load balance won't be triggered
    Map<String, Set<String>> errPartitions = new HashMap<>();
    errPartitions.put("OFFLINE-SLAVE", TestHelper.setOf(resourceName + "_0"));
    // start part of participants
    for (int i = 0; i < participantCount - 1; i++) {
        participants[i] = new MockParticipantManager(ZK_ADDR, clusterName, "localhost_" + (12918 + i));
        if (i == 0) {
            participants[i].setTransition(new ErrTransition(errPartitions));
        }
        participants[i].syncStart();
    }
    ClusterControllerManager controller = new ClusterControllerManager(ZK_ADDR, clusterName, "controller_0");
    controller.syncStart();
    BestPossibleExternalViewVerifier verifier = new BestPossibleExternalViewVerifier.Builder(clusterName).setZkClient(_gZkClient).build();
    Assert.assertTrue(verifier.verify(3000));
    // Adding one more participant.
    participants[participantCount - 1] = new MockParticipantManager(ZK_ADDR, clusterName, "localhost_" + (12918 + participantCount - 1));
    participants[participantCount - 1].syncStart();
    // Since error partition exists, no load balance transition will be done
    Assert.assertFalse(pollForPartitionAssignment(accessor, participants[participantCount - 1], resourceName, 5000));
    // Update cluster config to tolerate error partition, so load balance transition will be done
    clusterConfig = accessor.getProperty(accessor.keyBuilder().clusterConfig());
    clusterConfig.setErrorPartitionThresholdForLoadBalance(1);
    accessor.setProperty(keyBuilder.clusterConfig(), clusterConfig);
    _gSetupTool.rebalanceResource(clusterName, resourceName, 3);
    Assert.assertTrue(pollForPartitionAssignment(accessor, participants[participantCount - 1], resourceName, 3000));
    // clean up
    controller.syncStop();
    for (int i = 0; i < participantCount; i++) {
        participants[i].syncStop();
    }
    System.out.println("END " + clusterName + " at " + new Date(System.currentTimeMillis()));
}

Example 30 with ZKHelixDataAccessor

use of org.apache.helix.manager.zk.ZKHelixDataAccessor in project helix by apache.

the class TestAddStateModelFactoryAfterConnect method testBasic.

@Test
public void testBasic() throws Exception {
    // Logger.getRootLogger().setLevel(Level.INFO);
    String className = TestHelper.getTestClassName();
    String methodName = TestHelper.getTestMethodName();
    String clusterName = className + "_" + methodName;
    final int n = 5;
    System.out.println("START " + clusterName + " at " + new Date(System.currentTimeMillis()));
    MockParticipantManager[] participants = new MockParticipantManager[n];
    // participant port
    TestHelper.setupCluster(// participant port
    clusterName, // participant port
    ZK_ADDR, // participant port
    12918, // participant name prefix
    "localhost", // resource name prefix
    "TestDB", // resources
    1, // partitions per resource
    10, // number of nodes
    n, // replicas
    3, "MasterSlave", // do rebalance
    true);
    ClusterControllerManager controller = new ClusterControllerManager(ZK_ADDR, clusterName, "controller_0");
    controller.syncStart();
    // start participants
    for (int i = 0; i < n; i++) {
        String instanceName = "localhost_" + (12918 + i);
        participants[i] = new MockParticipantManager(ZK_ADDR, clusterName, instanceName);
        participants[i].syncStart();
    }
    boolean result = ClusterStateVerifier.verifyByZkCallback(new BestPossAndExtViewZkVerifier(ZK_ADDR, clusterName));
    Assert.assertTrue(result);
    // add a new idealState without registering message handling factory
    ClusterSetup setupTool = new ClusterSetup(ZK_ADDR);
    setupTool.addResourceToCluster(clusterName, "TestDB1", 16, "MasterSlave");
    ZkBaseDataAccessor<ZNRecord> baseAccessor = new ZkBaseDataAccessor<ZNRecord>(_gZkClient);
    ZKHelixDataAccessor accessor = new ZKHelixDataAccessor(clusterName, baseAccessor);
    Builder keyBuilder = accessor.keyBuilder();
    IdealState idealState = accessor.getProperty(keyBuilder.idealStates("TestDB1"));
    idealState.setStateModelFactoryName("TestDB1_Factory");
    accessor.setProperty(keyBuilder.idealStates("TestDB1"), idealState);
    setupTool.rebalanceStorageCluster(clusterName, "TestDB1", 3);
    // assert that we have received OFFLINE->SLAVE messages for all partitions
    int totalMsgs = 0;
    for (int retry = 0; retry < 5; retry++) {
        Thread.sleep(100);
        totalMsgs = 0;
        for (int i = 0; i < n; i++) {
            List<Message> msgs = accessor.getChildValues(keyBuilder.messages(participants[i].getInstanceName()));
            totalMsgs += msgs.size();
        }
        if (// partition# x replicas
        totalMsgs == 48)
            break;
    }
    Assert.assertEquals(totalMsgs, 48, "Should accumulated 48 unprocessed messages (1 O->S per partition per replica) because TestDB1 is added without state-model-factory but was " + totalMsgs);
    // Logger.getRootLogger().setLevel(Level.INFO);
    for (int i = 0; i < n; i++) {
        participants[i].getStateMachineEngine().registerStateModelFactory("MasterSlave", new MockMSModelFactory(), "TestDB1_Factory");
    }
    result = ClusterStateVerifier.verifyByZkCallback(new BestPossAndExtViewZkVerifier(ZK_ADDR, clusterName));
    Assert.assertTrue(result);
    // clean up
    // wait for all zk callbacks done
    controller.syncStop();
    for (int i = 0; i < 5; i++) {
        participants[i].syncStop();
    }
    System.out.println("END " + clusterName + " at " + new Date(System.currentTimeMillis()));
}