@Override
public void start() {
try {
/** 状态check时间间隔，较短，可以把任务及时分发到对应channel中 */
int sleepIntervalInMillSec =
this.configuration.getInt(CoreConstant.DATAX_CORE_CONTAINER_TASKGROUP_SLEEPINTERVAL, 100);
/** 状态汇报时间间隔，稍长，避免大量汇报 */
long reportIntervalInMillSec =
this.configuration.getLong(
CoreConstant.DATAX_CORE_CONTAINER_TASKGROUP_REPORTINTERVAL, 5000);
// 获取channel数目
int channelNumber =
this.configuration.getInt(CoreConstant.DATAX_CORE_CONTAINER_TASKGROUP_CHANNEL);
int taskMaxRetryTimes =
this.configuration.getInt(
CoreConstant.DATAX_CORE_CONTAINER_TASK_FAILOVER_MAXRETRYTIMES, 1);
long taskRetryIntervalInMsec =
this.configuration.getLong(
CoreConstant.DATAX_CORE_CONTAINER_TASK_FAILOVER_RETRYINTERVALINMSEC, 10000);
long taskMaxWaitInMsec =
this.configuration.getLong(
CoreConstant.DATAX_CORE_CONTAINER_TASK_FAILOVER_MAXWAITINMSEC, 60000);
List<Configuration> taskConfigs =
this.configuration.getListConfiguration(CoreConstant.DATAX_JOB_CONTENT);
if (LOG.isDebugEnabled()) {
LOG.debug(
"taskGroup[{}]'s task configs[{}]", this.taskGroupId, JSON.toJSONString(taskConfigs));
}
int taskCountInThisTaskGroup = taskConfigs.size();
LOG.info(
String.format(
"taskGroupId=[%d] start [%d] channels for [%d] tasks.",
this.taskGroupId, channelNumber, taskCountInThisTaskGroup));
this.containerCommunicator.registerCommunication(taskConfigs);
Map<Integer, Configuration> taskConfigMap = buildTaskConfigMap(taskConfigs); // taskId与task配置
List<Configuration> taskQueue = buildRemainTasks(taskConfigs); // 待运行task列表
Map<Integer, TaskExecutor> taskFailedExecutorMap =
new HashMap<Integer, TaskExecutor>(); // taskId与上次失败实例
List<TaskExecutor> runTasks = new ArrayList<TaskExecutor>(channelNumber); // 正在运行task
Map<Integer, Long> taskStartTimeMap = new HashMap<Integer, Long>(); // 任务开始时间
long lastReportTimeStamp = 0;
Communication lastTaskGroupContainerCommunication = new Communication();
while (true) {
// 1.判断task状态
boolean failedOrKilled = false;
Map<Integer, Communication> communicationMap = containerCommunicator.getCommunicationMap();
for (Map.Entry<Integer, Communication> entry : communicationMap.entrySet()) {
Integer taskId = entry.getKey();
Communication taskCommunication = entry.getValue();
if (!taskCommunication.isFinished()) {
continue;
}
TaskExecutor taskExecutor = removeTask(runTasks, taskId);
// 上面从runTasks里移除了，因此对应在monitor里移除
taskMonitor.removeTask(taskId);
// 失败，看task是否支持failover，重试次数未超过最大限制
if (taskCommunication.getState() == State.FAILED) {
taskFailedExecutorMap.put(taskId, taskExecutor);
if (taskExecutor.supportFailOver()
&& taskExecutor.getAttemptCount() < taskMaxRetryTimes) {
taskExecutor.shutdown(); // 关闭老的executor
containerCommunicator.resetCommunication(taskId); // 将task的状态重置
Configuration taskConfig = taskConfigMap.get(taskId);
taskQueue.add(taskConfig); // 重新加入任务列表
} else {
failedOrKilled = true;
break;
}
} else if (taskCommunication.getState() == State.KILLED) {
failedOrKilled = true;
break;
} else if (taskCommunication.getState() == State.SUCCEEDED) {
Long taskStartTime = taskStartTimeMap.get(taskId);
if (taskStartTime != null) {
Long usedTime = System.currentTimeMillis() - taskStartTime;
LOG.info(
"taskGroup[{}] taskId[{}] is successed, used[{}]ms",
this.taskGroupId,
taskId,
usedTime);
// usedTime*1000*1000 转换成PerfRecord记录的ns，这里主要是简单登记，进行最长任务的打印。因此增加特定静态方法
PerfRecord.addPerfRecord(
taskGroupId,
taskId,
PerfRecord.PHASE.TASK_TOTAL,
taskStartTime,
usedTime * 1000L * 1000L);
taskStartTimeMap.remove(taskId);
taskConfigMap.remove(taskId);
}
}
}
// 2.发现该taskGroup下taskExecutor的总状态失败则汇报错误
if (failedOrKilled) {
lastTaskGroupContainerCommunication =
reportTaskGroupCommunication(
lastTaskGroupContainerCommunication, taskCountInThisTaskGroup);
throw DataXException.asDataXException(
FrameworkErrorCode.PLUGIN_RUNTIME_ERROR,
lastTaskGroupContainerCommunication.getThrowable());
}
// 3.有任务未执行，且正在运行的任务数小于最大通道限制
Iterator<Configuration> iterator = taskQueue.iterator();
while (iterator.hasNext() && runTasks.size() < channelNumber) {
Configuration taskConfig = iterator.next();
Integer taskId = taskConfig.getInt(CoreConstant.TASK_ID);
int attemptCount = 1;
TaskExecutor lastExecutor = taskFailedExecutorMap.get(taskId);
if (lastExecutor != null) {
attemptCount = lastExecutor.getAttemptCount() + 1;
long now = System.currentTimeMillis();
long failedTime = lastExecutor.getTimeStamp();
if (now - failedTime < taskRetryIntervalInMsec) { // 未到等待时间，继续留在队列
continue;
}
if (!lastExecutor.isShutdown()) { // 上次失败的task仍未结束
if (now - failedTime > taskMaxWaitInMsec) {
markCommunicationFailed(taskId);
reportTaskGroupCommunication(
lastTaskGroupContainerCommunication, taskCountInThisTaskGroup);
throw DataXException.asDataXException(
CommonErrorCode.WAIT_TIME_EXCEED, "task failover等待超时");
} else {
lastExecutor.shutdown(); // 再次尝试关闭
continue;
}
} else {
LOG.info(
"taskGroup[{}] taskId[{}] attemptCount[{}] has already shutdown",
this.taskGroupId,
taskId,
lastExecutor.getAttemptCount());
}
}
Configuration taskConfigForRun = taskMaxRetryTimes > 1 ? taskConfig.clone() : taskConfig;
TaskExecutor taskExecutor = new TaskExecutor(taskConfigForRun, attemptCount);
taskStartTimeMap.put(taskId, System.currentTimeMillis());
taskExecutor.doStart();
iterator.remove();
runTasks.add(taskExecutor);
// 上面，增加task到runTasks列表，因此在monitor里注册。
taskMonitor.registerTask(taskId, this.containerCommunicator.getCommunication(taskId));
taskFailedExecutorMap.remove(taskId);
LOG.info(
"taskGroup[{}] taskId[{}] attemptCount[{}] is started",
this.taskGroupId,
taskId,
attemptCount);
}
// 4.任务列表为空，executor已结束, 搜集状态为success--->成功
if (taskQueue.isEmpty()
&& isAllTaskDone(runTasks)
&& containerCommunicator.collectState() == State.SUCCEEDED) {
// 成功的情况下，也需要汇报一次。否则在任务结束非常快的情况下，采集的信息将会不准确
lastTaskGroupContainerCommunication =
reportTaskGroupCommunication(
lastTaskGroupContainerCommunication, taskCountInThisTaskGroup);
LOG.info("taskGroup[{}] completed it's tasks.", this.taskGroupId);
break;
}
// 5.如果当前时间已经超出汇报时间的interval，那么我们需要马上汇报
long now = System.currentTimeMillis();
if (now - lastReportTimeStamp > reportIntervalInMillSec) {
lastTaskGroupContainerCommunication =
reportTaskGroupCommunication(
lastTaskGroupContainerCommunication, taskCountInThisTaskGroup);
lastReportTimeStamp = now;
// taskMonitor对于正在运行的task，每reportIntervalInMillSec进行检查
for (TaskExecutor taskExecutor : runTasks) {
taskMonitor.report(
taskExecutor.getTaskId(),
this.containerCommunicator.getCommunication(taskExecutor.getTaskId()));
}
}
Thread.sleep(sleepIntervalInMillSec);
}
// 6.最后还要汇报一次
reportTaskGroupCommunication(lastTaskGroupContainerCommunication, taskCountInThisTaskGroup);
} catch (Throwable e) {
Communication nowTaskGroupContainerCommunication = this.containerCommunicator.collect();
if (nowTaskGroupContainerCommunication.getThrowable() == null) {
nowTaskGroupContainerCommunication.setThrowable(e);
}
nowTaskGroupContainerCommunication.setState(State.FAILED);
this.containerCommunicator.report(nowTaskGroupContainerCommunication);
throw DataXException.asDataXException(FrameworkErrorCode.RUNTIME_ERROR, e);
} finally {
if (!PerfTrace.getInstance().isJob()) {
// 最后打印cpu的平均消耗，GC的统计
VMInfo vmInfo = VMInfo.getVmInfo();
if (vmInfo != null) {
vmInfo.getDelta(false);
LOG.info(vmInfo.totalString());
}
LOG.info(PerfTrace.getInstance().summarizeNoException());
}
}
}
@Test
public void testNormal() throws Exception {
// register task
long ttl = System.currentTimeMillis();
Communication communication1 = new Communication();
taskMonitor.registerTask(1, communication1);
TaskMonitor.TaskCommunication taskCommunication1 = taskMonitor.getTaskCommunication(1);
Assert.assertEquals(taskCommunication1.getLastAllReadRecords(), 0L);
Assert.assertEquals(this.tasks.size(), 1);
Assert.assertTrue(taskCommunication1.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication1.getTtl() >= ttl);
// report 没有任何变化的communication
long oldTS = taskCommunication1.getLastUpdateComunicationTS();
long oldTTL = taskCommunication1.getTtl();
Thread.sleep(1000);
taskMonitor.report(1, communication1);
TaskMonitor.TaskCommunication taskCommunication1_1 = taskMonitor.getTaskCommunication(1);
Assert.assertEquals(taskCommunication1_1.getLastAllReadRecords(), 0L);
Assert.assertEquals(taskCommunication1_1.getLastUpdateComunicationTS(), oldTS);
Assert.assertTrue(taskCommunication1_1.getTtl() > oldTTL);
// report 已经finish的communication
Communication communication2 = new Communication();
communication2.setState(State.KILLED);
taskMonitor.registerTask(2, communication2);
Assert.assertEquals(this.tasks.size(), 1);
// report 另一个communication
Communication communication3 = new Communication();
taskMonitor.registerTask(3, communication3);
Assert.assertEquals(this.tasks.size(), 2);
System.out.println(this.tasks);
// report communication
ttl = System.currentTimeMillis();
communication1.setLongCounter(CommunicationTool.READ_SUCCEED_RECORDS, 100);
communication3.setLongCounter(CommunicationTool.READ_FAILED_RECORDS, 10);
taskMonitor.report(1, communication1);
taskMonitor.report(3, communication3);
taskCommunication1 = taskMonitor.getTaskCommunication(1);
Assert.assertEquals(taskCommunication1.getLastAllReadRecords(), 100L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(taskCommunication1.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication1.getTtl() >= ttl);
TaskMonitor.TaskCommunication taskCommunication3 = taskMonitor.getTaskCommunication(3);
Assert.assertEquals(taskCommunication3.getLastAllReadRecords(), 10L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(taskCommunication3.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication3.getTtl() >= ttl);
// 继续report
ttl = System.currentTimeMillis();
communication1.setLongCounter(CommunicationTool.READ_SUCCEED_RECORDS, 1001);
communication3.setLongCounter(CommunicationTool.READ_FAILED_RECORDS, 101);
taskMonitor.report(1, communication1);
taskMonitor.report(3, communication3);
taskCommunication1 = taskMonitor.getTaskCommunication(1);
Assert.assertEquals(taskCommunication1.getLastAllReadRecords(), 1001L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(taskCommunication1.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication1.getTtl() >= ttl);
taskCommunication3 = taskMonitor.getTaskCommunication(3);
Assert.assertEquals(taskCommunication3.getLastAllReadRecords(), 101L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(taskCommunication3.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication3.getTtl() >= ttl);
// 设置EXPIRED_TIME
Field EXPIRED_TIME = taskMonitor.getClass().getDeclaredField("EXPIRED_TIME");
EXPIRED_TIME.setAccessible(true);
EXPIRED_TIME.set(null, 1000);
Thread.sleep(2000);
// 超时没有变更
taskMonitor.report(1, communication1);
System.out.println(communication1.getCounter());
System.out.println(communication1.getThrowable());
System.out.println(communication1.getThrowableMessage());
System.out.println(communication1.getState());
Assert.assertTrue(communication1.getThrowableMessage().contains("任务hung住，Expired"));
Assert.assertEquals(communication1.getState(), State.FAILED);
// communicatio1 已经fail， communication3 在超时后进行变更，update正常
ttl = System.currentTimeMillis();
communication1.setLongCounter(CommunicationTool.READ_SUCCEED_RECORDS, 2001);
communication3.setLongCounter(CommunicationTool.READ_FAILED_RECORDS, 201);
taskMonitor.report(1, communication1);
taskMonitor.report(3, communication3);
taskCommunication1 = taskMonitor.getTaskCommunication(1);
Assert.assertEquals(taskCommunication1.getLastAllReadRecords(), 1001L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(communication1.getThrowableMessage().contains("任务hung住，Expired"));
Assert.assertEquals(communication1.getState(), State.FAILED);
taskCommunication3 = taskMonitor.getTaskCommunication(3);
Assert.assertEquals(taskCommunication3.getLastAllReadRecords(), 201L);
Assert.assertEquals(this.tasks.size(), 2);
Assert.assertTrue(taskCommunication3.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication3.getTtl() >= ttl);
// remove 1
taskMonitor.removeTask(1);
Assert.assertEquals(this.tasks.size(), 1);
// remove 3
taskMonitor.removeTask(3);
Assert.assertEquals(this.tasks.size(), 0);
// 没有register communication3 直接report
ttl = System.currentTimeMillis();
communication3.setLongCounter(CommunicationTool.READ_FAILED_RECORDS, 301);
taskMonitor.report(3, communication3);
taskCommunication3 = taskMonitor.getTaskCommunication(3);
Assert.assertEquals(taskCommunication3.getLastAllReadRecords(), 301L);
Assert.assertEquals(this.tasks.size(), 1);
Assert.assertTrue(taskCommunication3.getLastUpdateComunicationTS() >= ttl);
Assert.assertTrue(taskCommunication3.getTtl() >= ttl);
}
private void markCommunicationFailed(Integer taskId) {
Communication communication = containerCommunicator.getCommunication(taskId);
communication.setState(State.FAILED);
}