Examples with ApplicationAttemptStateData org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData used on opensource projects

Example 1 with ApplicationAttemptStateData

use of org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData in project hadoop by apache.

the class LeveldbRMStateStore method loadRMApp.

private int loadRMApp(RMState rmState, LeveldbIterator iter, String appIdStr, byte[] appData) throws IOException {
    ApplicationStateData appState = createApplicationState(appIdStr, appData);
    ApplicationId appId = appState.getApplicationSubmissionContext().getApplicationId();
    rmState.appState.put(appId, appState);
    String attemptNodePrefix = getApplicationNodeKey(appId) + SEPARATOR;
    while (iter.hasNext()) {
        Entry<byte[], byte[]> entry = iter.peekNext();
        String key = asString(entry.getKey());
        if (!key.startsWith(attemptNodePrefix)) {
            break;
        }
        String attemptId = key.substring(attemptNodePrefix.length());
        if (attemptId.startsWith(ApplicationAttemptId.appAttemptIdStrPrefix)) {
            ApplicationAttemptStateData attemptState = createAttemptState(attemptId, entry.getValue());
            appState.attempts.put(attemptState.getAttemptId(), attemptState);
        } else {
            LOG.warn("Ignoring unknown application key: " + key);
        }
        iter.next();
    }
    int numAttempts = appState.attempts.size();
    if (LOG.isDebugEnabled()) {
        LOG.debug("Loaded application " + appId + " with " + numAttempts + " attempts");
    }
    return numAttempts;
}

Example 2 with ApplicationAttemptStateData

use of org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData in project hadoop by apache.

the class RMAppAttemptImpl method rememberTargetTransitionsAndStoreState.

private void rememberTargetTransitionsAndStoreState(RMAppAttemptEvent event, Object transitionToDo, RMAppAttemptState targetFinalState, RMAppAttemptState stateToBeStored) {
    rememberTargetTransitions(event, transitionToDo, targetFinalState);
    stateBeforeFinalSaving = getState();
    // As of today, finalState, diagnostics, final-tracking-url and
    // finalAppStatus are the only things that we store into the StateStore
    // AFTER the initial saving on app-attempt-start
    // These fields can be visible from outside only after they are saved in
    // StateStore
    String diags = null;
    // don't leave the tracking URL pointing to a non-existent AM
    if (conf.getBoolean(YarnConfiguration.APPLICATION_HISTORY_ENABLED, YarnConfiguration.DEFAULT_APPLICATION_HISTORY_ENABLED)) {
        setTrackingUrlToAHSPage(stateToBeStored);
    } else {
        setTrackingUrlToRMAppPage(stateToBeStored);
    }
    String finalTrackingUrl = getOriginalTrackingUrl();
    FinalApplicationStatus finalStatus = null;
    int exitStatus = ContainerExitStatus.INVALID;
    switch(event.getType()) {
        case LAUNCH_FAILED:
            diags = event.getDiagnosticMsg();
            break;
        case REGISTERED:
            diags = getUnexpectedAMRegisteredDiagnostics();
            break;
        case UNREGISTERED:
            RMAppAttemptUnregistrationEvent unregisterEvent = (RMAppAttemptUnregistrationEvent) event;
            diags = unregisterEvent.getDiagnosticMsg();
            // reset finalTrackingUrl to url sent by am
            finalTrackingUrl = sanitizeTrackingUrl(unregisterEvent.getFinalTrackingUrl());
            finalStatus = unregisterEvent.getFinalApplicationStatus();
            break;
        case CONTAINER_FINISHED:
            RMAppAttemptContainerFinishedEvent finishEvent = (RMAppAttemptContainerFinishedEvent) event;
            diags = getAMContainerCrashedDiagnostics(finishEvent);
            exitStatus = finishEvent.getContainerStatus().getExitStatus();
            break;
        case KILL:
            break;
        case FAIL:
            diags = event.getDiagnosticMsg();
            break;
        case EXPIRE:
            diags = getAMExpiredDiagnostics(event);
            break;
        default:
            break;
    }
    AggregateAppResourceUsage resUsage = this.attemptMetrics.getAggregateAppResourceUsage();
    RMStateStore rmStore = rmContext.getStateStore();
    setFinishTime(System.currentTimeMillis());
    ApplicationAttemptStateData attemptState = ApplicationAttemptStateData.newInstance(applicationAttemptId, getMasterContainer(), rmStore.getCredentialsFromAppAttempt(this), startTime, stateToBeStored, finalTrackingUrl, diags, finalStatus, exitStatus, getFinishTime(), resUsage.getMemorySeconds(), resUsage.getVcoreSeconds(), this.attemptMetrics.getPreemptedMemory(), this.attemptMetrics.getPreemptedVcore());
    LOG.info("Updating application attempt " + applicationAttemptId + " with final state: " + targetedFinalState + ", and exit status: " + exitStatus);
    rmStore.updateApplicationAttemptState(attemptState);
}

Example 3 with ApplicationAttemptStateData

use of org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData in project hadoop by apache.

the class RMAppAttemptImpl method recover.

@Override
public void recover(RMState state) {
    ApplicationStateData appState = state.getApplicationState().get(getAppAttemptId().getApplicationId());
    ApplicationAttemptStateData attemptState = appState.getAttempt(getAppAttemptId());
    assert attemptState != null;
    if (attemptState.getState() == null) {
        LOG.info(String.format(RECOVERY_MESSAGE, getAppAttemptId(), "NONE"));
    } else if (LOG.isDebugEnabled()) {
        LOG.debug(String.format(RECOVERY_MESSAGE, getAppAttemptId(), attemptState.getState()));
    }
    this.diagnostics.append("Attempt recovered after RM restart");
    this.diagnostics.append(attemptState.getDiagnostics());
    this.amContainerExitStatus = attemptState.getAMContainerExitStatus();
    if (amContainerExitStatus == ContainerExitStatus.PREEMPTED) {
        this.attemptMetrics.setIsPreempted();
    }
    Credentials credentials = attemptState.getAppAttemptTokens();
    setMasterContainer(attemptState.getMasterContainer());
    recoverAppAttemptCredentials(credentials, attemptState.getState());
    this.recoveredFinalState = attemptState.getState();
    this.originalTrackingUrl = attemptState.getFinalTrackingUrl();
    this.finalStatus = attemptState.getFinalApplicationStatus();
    this.startTime = attemptState.getStartTime();
    this.finishTime = attemptState.getFinishTime();
    this.attemptMetrics.updateAggregateAppResourceUsage(attemptState.getMemorySeconds(), attemptState.getVcoreSeconds());
    this.attemptMetrics.updateAggregatePreemptedAppResourceUsage(attemptState.getPreemptedMemorySeconds(), attemptState.getPreemptedVcoreSeconds());
}

Example 4 with ApplicationAttemptStateData

use of org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData in project hadoop by apache.

the class TestRMRestart method testRMRestart.

@SuppressWarnings("rawtypes")
@Test(timeout = 180000)
public void testRMRestart() throws Exception {
    conf.setInt(YarnConfiguration.RM_AM_MAX_ATTEMPTS, YarnConfiguration.DEFAULT_RM_AM_MAX_ATTEMPTS);
    MemoryRMStateStore memStore = new MemoryRMStateStore();
    memStore.init(conf);
    RMState rmState = memStore.getState();
    Map<ApplicationId, ApplicationStateData> rmAppState = rmState.getApplicationState();
    // PHASE 1: create state in an RM
    // start RM
    MockRM rm1 = createMockRM(conf, memStore);
    // start like normal because state is empty
    rm1.start();
    MockNM nm1 = new MockNM("127.0.0.1:1234", 15120, rm1.getResourceTrackerService());
    MockNM nm2 = new MockNM("127.0.0.2:5678", 15120, rm1.getResourceTrackerService());
    nm1.registerNode();
    // nm2 will not heartbeat with RM1
    nm2.registerNode();
    // create app that will finish and the final state should be saved.
    RMApp app0 = rm1.submitApp(200);
    RMAppAttempt attempt0 = app0.getCurrentAppAttempt();
    // spot check that app is saved
    Assert.assertEquals(1, rmAppState.size());
    nm1.nodeHeartbeat(true);
    MockAM am0 = rm1.sendAMLaunched(attempt0.getAppAttemptId());
    am0.registerAppAttempt();
    finishApplicationMaster(app0, rm1, nm1, am0);
    // create app that gets launched and does allocate before RM restart
    RMApp app1 = rm1.submitApp(200);
    // assert app1 info is saved
    ApplicationStateData appState = rmAppState.get(app1.getApplicationId());
    Assert.assertNotNull(appState);
    Assert.assertEquals(0, appState.getAttemptCount());
    Assert.assertEquals(appState.getApplicationSubmissionContext().getApplicationId(), app1.getApplicationSubmissionContext().getApplicationId());
    //kick the scheduling to allocate AM container
    nm1.nodeHeartbeat(true);
    // assert app1 attempt is saved
    RMAppAttempt attempt1 = app1.getCurrentAppAttempt();
    ApplicationAttemptId attemptId1 = attempt1.getAppAttemptId();
    rm1.waitForState(attemptId1, RMAppAttemptState.ALLOCATED);
    Assert.assertEquals(1, appState.getAttemptCount());
    ApplicationAttemptStateData attemptState = appState.getAttempt(attemptId1);
    Assert.assertNotNull(attemptState);
    Assert.assertEquals(BuilderUtils.newContainerId(attemptId1, 1), attemptState.getMasterContainer().getId());
    // launch the AM
    MockAM am1 = rm1.sendAMLaunched(attempt1.getAppAttemptId());
    am1.registerAppAttempt();
    // AM request for containers
    am1.allocate("127.0.0.1", 1000, 1, new ArrayList<ContainerId>());
    // kick the scheduler
    nm1.nodeHeartbeat(true);
    List<Container> conts = am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers();
    while (conts.size() == 0) {
        nm1.nodeHeartbeat(true);
        conts.addAll(am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers());
        Thread.sleep(500);
    }
    // create app that does not get launched by RM before RM restart
    RMApp app2 = rm1.submitApp(200);
    // assert app2 info is saved
    appState = rmAppState.get(app2.getApplicationId());
    Assert.assertNotNull(appState);
    Assert.assertEquals(0, appState.getAttemptCount());
    Assert.assertEquals(appState.getApplicationSubmissionContext().getApplicationId(), app2.getApplicationSubmissionContext().getApplicationId());
    // create unmanaged app
    RMApp appUnmanaged = rm1.submitApp(200, "someApp", "someUser", null, true, null, conf.getInt(YarnConfiguration.RM_AM_MAX_ATTEMPTS, YarnConfiguration.DEFAULT_RM_AM_MAX_ATTEMPTS), null);
    ApplicationAttemptId unmanagedAttemptId = appUnmanaged.getCurrentAppAttempt().getAppAttemptId();
    // assert appUnmanaged info is saved
    ApplicationId unmanagedAppId = appUnmanaged.getApplicationId();
    appState = rmAppState.get(unmanagedAppId);
    Assert.assertNotNull(appState);
    // wait for attempt to reach LAUNCHED state 
    rm1.waitForState(unmanagedAttemptId, RMAppAttemptState.LAUNCHED);
    rm1.waitForState(unmanagedAppId, RMAppState.ACCEPTED);
    // assert unmanaged attempt info is saved
    Assert.assertEquals(1, appState.getAttemptCount());
    Assert.assertEquals(appState.getApplicationSubmissionContext().getApplicationId(), appUnmanaged.getApplicationSubmissionContext().getApplicationId());
    // PHASE 2: create new RM and start from old state
    // create new RM to represent restart and recover state
    MockRM rm2 = createMockRM(conf, memStore);
    // start new RM
    rm2.start();
    // change NM to point to new RM
    nm1.setResourceTrackerService(rm2.getResourceTrackerService());
    nm2.setResourceTrackerService(rm2.getResourceTrackerService());
    // verify load of old state
    // 4 apps are loaded.
    // FINISHED app and attempt is also loaded back.
    // Unmanaged app state is still loaded back but it cannot be restarted by
    // the RM. this will change with work preserving RM restart in which AMs/NMs
    // are not rebooted.
    Assert.assertEquals(4, rm2.getRMContext().getRMApps().size());
    // check that earlier finished app and attempt is also loaded back and move
    // to finished state.
    rm2.waitForState(app0.getApplicationId(), RMAppState.FINISHED);
    rm2.waitForState(am0.getApplicationAttemptId(), RMAppAttemptState.FINISHED);
    // verify correct number of attempts and other data
    RMApp loadedApp1 = rm2.getRMContext().getRMApps().get(app1.getApplicationId());
    Assert.assertNotNull(loadedApp1);
    Assert.assertEquals(1, loadedApp1.getAppAttempts().size());
    Assert.assertEquals(app1.getApplicationSubmissionContext().getApplicationId(), loadedApp1.getApplicationSubmissionContext().getApplicationId());
    RMApp loadedApp2 = rm2.getRMContext().getRMApps().get(app2.getApplicationId());
    Assert.assertNotNull(loadedApp2);
    //Assert.assertEquals(0, loadedApp2.getAppAttempts().size());
    Assert.assertEquals(app2.getApplicationSubmissionContext().getApplicationId(), loadedApp2.getApplicationSubmissionContext().getApplicationId());
    // verify state machine kicked into expected states
    rm2.waitForState(loadedApp1.getApplicationId(), RMAppState.ACCEPTED);
    rm2.waitForState(loadedApp2.getApplicationId(), RMAppState.ACCEPTED);
    // verify attempts for apps
    // The app for which AM was started will wait for previous am
    // container finish event to arrive. However for an application for which
    // no am container was running will start new application attempt.
    Assert.assertEquals(1, loadedApp1.getAppAttempts().size());
    Assert.assertEquals(1, loadedApp2.getAppAttempts().size());
    // verify old AM is not accepted
    // change running AM to talk to new RM
    am1.setAMRMProtocol(rm2.getApplicationMasterService(), rm2.getRMContext());
    try {
        am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>());
        Assert.fail();
    } catch (ApplicationAttemptNotFoundException e) {
        Assert.assertTrue(e instanceof ApplicationAttemptNotFoundException);
    }
    // NM should be rebooted on heartbeat, even first heartbeat for nm2
    NodeHeartbeatResponse hbResponse = nm1.nodeHeartbeat(true);
    Assert.assertEquals(NodeAction.RESYNC, hbResponse.getNodeAction());
    hbResponse = nm2.nodeHeartbeat(true);
    Assert.assertEquals(NodeAction.RESYNC, hbResponse.getNodeAction());
    // new NM to represent NM re-register
    nm1 = new MockNM("127.0.0.1:1234", 15120, rm2.getResourceTrackerService());
    nm2 = new MockNM("127.0.0.2:5678", 15120, rm2.getResourceTrackerService());
    NMContainerStatus status = TestRMRestart.createNMContainerStatus(loadedApp1.getCurrentAppAttempt().getAppAttemptId(), 1, ContainerState.COMPLETE);
    nm1.registerNode(Arrays.asList(status), null);
    nm2.registerNode();
    rm2.waitForState(loadedApp1.getApplicationId(), RMAppState.ACCEPTED);
    // wait for the 2nd attempt to be started.
    int timeoutSecs = 0;
    while (loadedApp1.getAppAttempts().size() != 2 && timeoutSecs++ < 40) {
        ;
        Thread.sleep(200);
    }
    // verify no more reboot response sent
    hbResponse = nm1.nodeHeartbeat(true);
    Assert.assertTrue(NodeAction.RESYNC != hbResponse.getNodeAction());
    hbResponse = nm2.nodeHeartbeat(true);
    Assert.assertTrue(NodeAction.RESYNC != hbResponse.getNodeAction());
    // assert app1 attempt is saved
    attempt1 = loadedApp1.getCurrentAppAttempt();
    attemptId1 = attempt1.getAppAttemptId();
    rm2.waitForState(attemptId1, RMAppAttemptState.ALLOCATED);
    appState = rmAppState.get(loadedApp1.getApplicationId());
    attemptState = appState.getAttempt(attemptId1);
    Assert.assertNotNull(attemptState);
    Assert.assertEquals(BuilderUtils.newContainerId(attemptId1, 1), attemptState.getMasterContainer().getId());
    // Nodes on which the AM's run 
    MockNM am1Node = nm1;
    if (attemptState.getMasterContainer().getNodeId().toString().contains("127.0.0.2")) {
        am1Node = nm2;
    }
    // assert app2 attempt is saved
    RMAppAttempt attempt2 = loadedApp2.getCurrentAppAttempt();
    ApplicationAttemptId attemptId2 = attempt2.getAppAttemptId();
    rm2.waitForState(attemptId2, RMAppAttemptState.ALLOCATED);
    appState = rmAppState.get(loadedApp2.getApplicationId());
    attemptState = appState.getAttempt(attemptId2);
    Assert.assertNotNull(attemptState);
    Assert.assertEquals(BuilderUtils.newContainerId(attemptId2, 1), attemptState.getMasterContainer().getId());
    MockNM am2Node = nm1;
    if (attemptState.getMasterContainer().getNodeId().toString().contains("127.0.0.2")) {
        am2Node = nm2;
    }
    // start the AM's
    am1 = rm2.sendAMLaunched(attempt1.getAppAttemptId());
    am1.registerAppAttempt();
    MockAM am2 = rm2.sendAMLaunched(attempt2.getAppAttemptId());
    am2.registerAppAttempt();
    //request for containers
    am1.allocate("127.0.0.1", 1000, 3, new ArrayList<ContainerId>());
    am2.allocate("127.0.0.2", 1000, 1, new ArrayList<ContainerId>());
    // verify container allocate continues to work
    nm1.nodeHeartbeat(true);
    nm2.nodeHeartbeat(true);
    conts = am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers();
    while (conts.size() == 0) {
        nm1.nodeHeartbeat(true);
        nm2.nodeHeartbeat(true);
        conts.addAll(am1.allocate(new ArrayList<ResourceRequest>(), new ArrayList<ContainerId>()).getAllocatedContainers());
        Thread.sleep(500);
    }
    // finish the AMs
    finishApplicationMaster(loadedApp1, rm2, am1Node, am1);
    finishApplicationMaster(loadedApp2, rm2, am2Node, am2);
    // stop RM's
    rm2.stop();
    rm1.stop();
    // completed apps are not removed immediately after app finish
    // And finished app is also loaded back.
    Assert.assertEquals(4, rmAppState.size());
}

Example 5 with ApplicationAttemptStateData

use of org.apache.hadoop.yarn.server.resourcemanager.recovery.records.ApplicationAttemptStateData in project hadoop by apache.

the class TestRMRestart method testRMRestartOnMissingAttempts.

@Test(timeout = 60000)
public void testRMRestartOnMissingAttempts() throws Exception {
    conf.setInt(YarnConfiguration.RM_AM_MAX_ATTEMPTS, 5);
    MemoryRMStateStore memStore = new MemoryRMStateStore();
    memStore.init(conf);
    // start RM
    MockRM rm1 = createMockRM(conf, memStore);
    rm1.start();
    MockNM nm1 = new MockNM("127.0.0.1:1234", 15120, rm1.getResourceTrackerService());
    nm1.registerNode();
    // create an app and finish the app.
    RMApp app0 = rm1.submitApp(200);
    ApplicationStateData app0State = memStore.getState().getApplicationState().get(app0.getApplicationId());
    MockAM am0 = launchAndFailAM(app0, rm1, nm1);
    MockAM am1 = launchAndFailAM(app0, rm1, nm1);
    MockAM am2 = launchAndFailAM(app0, rm1, nm1);
    MockAM am3 = launchAM(app0, rm1, nm1);
    // am1 is missed from MemoryRMStateStore
    memStore.removeApplicationAttemptInternal(am1.getApplicationAttemptId());
    ApplicationAttemptStateData am2State = app0State.getAttempt(am2.getApplicationAttemptId());
    // am2's state is not consistent: MemoryRMStateStore just saved its initial
    // state and failed to store its final state
    am2State.setState(null);
    // restart rm
    MockRM rm2 = createMockRM(conf, memStore);
    rm2.start();
    Assert.assertEquals(1, rm2.getRMContext().getRMApps().size());
    RMApp recoveredApp0 = rm2.getRMContext().getRMApps().values().iterator().next();
    Map<ApplicationAttemptId, RMAppAttempt> recoveredAppAttempts = recoveredApp0.getAppAttempts();
    Assert.assertEquals(3, recoveredAppAttempts.size());
    Assert.assertEquals(RMAppAttemptState.FAILED, recoveredAppAttempts.get(am0.getApplicationAttemptId()).getAppAttemptState());
    Assert.assertEquals(RMAppAttemptState.FAILED, recoveredAppAttempts.get(am2.getApplicationAttemptId()).getAppAttemptState());
    Assert.assertEquals(RMAppAttemptState.LAUNCHED, recoveredAppAttempts.get(am3.getApplicationAttemptId()).getAppAttemptState());
    Assert.assertEquals(5, ((RMAppImpl) app0).getNextAttemptId());
}