Search in sources :

Example 1 with JobManagerConfiguration

use of com.netflix.titus.master.jobmanager.service.JobManagerConfiguration in project titus-control-plane by Netflix.

the class DifferenceResolverUtils method findTaskStateTimeouts.

/**
 * Find all tasks that are stuck in a specific state. The number of {@link ChangeAction changes} will be limited
 * by the {@link TokenBucket stuckInStateRateLimiter}
 */
public static List<ChangeAction> findTaskStateTimeouts(ReconciliationEngine<JobManagerReconcilerEvent> engine, JobView runningJobView, JobManagerConfiguration configuration, JobServiceRuntime runtime, JobStore jobStore, VersionSupplier versionSupplier, TokenBucket stuckInStateRateLimiter, TitusRuntime titusRuntime) {
    Clock clock = titusRuntime.getClock();
    List<ChangeAction> actions = new ArrayList<>();
    runningJobView.getJobHolder().getChildren().forEach(taskHolder -> {
        Task task = taskHolder.getEntity();
        TaskState taskState = task.getStatus().getState();
        if (JobFunctions.isBatchJob(runningJobView.getJob()) && taskState == TaskState.Started) {
            Job<BatchJobExt> batchJob = runningJobView.getJob();
            // We expect runtime limit to be always set, so this is just extra safety measure.
            long runtimeLimitMs = Math.max(BatchJobExt.RUNTIME_LIMIT_MIN, batchJob.getJobDescriptor().getExtensions().getRuntimeLimitMs());
            long deadline = task.getStatus().getTimestamp() + runtimeLimitMs;
            if (deadline < clock.wallTime()) {
                actions.add(KillInitiatedActions.reconcilerInitiatedTaskKillInitiated(engine, task, runtime, jobStore, versionSupplier, TaskStatus.REASON_RUNTIME_LIMIT_EXCEEDED, "Task running too long (runtimeLimit=" + runtimeLimitMs + "ms)", titusRuntime));
            }
            return;
        }
        TaskTimeoutChangeActions.TimeoutStatus timeoutStatus = TaskTimeoutChangeActions.getTimeoutStatus(taskHolder, clock);
        switch(timeoutStatus) {
            case Ignore:
            case Pending:
                break;
            case NotSet:
                long timeoutMs = -1;
                switch(taskState) {
                    case Launched:
                        timeoutMs = configuration.getTaskInLaunchedStateTimeoutMs();
                        break;
                    case StartInitiated:
                        timeoutMs = isBatch(runningJobView.getJob()) ? configuration.getBatchTaskInStartInitiatedStateTimeoutMs() : configuration.getServiceTaskInStartInitiatedStateTimeoutMs();
                        break;
                    case KillInitiated:
                        timeoutMs = configuration.getTaskInKillInitiatedStateTimeoutMs();
                        break;
                }
                if (timeoutMs > 0) {
                    actions.add(TaskTimeoutChangeActions.setTimeout(taskHolder.getId(), task.getStatus().getState(), timeoutMs, clock));
                }
                break;
            case TimedOut:
                if (!stuckInStateRateLimiter.tryTake()) {
                    break;
                }
                if (task.getStatus().getState() == TaskState.KillInitiated) {
                    int attempts = TaskTimeoutChangeActions.getKillInitiatedAttempts(taskHolder) + 1;
                    if (attempts >= configuration.getTaskKillAttempts()) {
                        actions.add(BasicTaskActions.updateTaskInRunningModel(task.getId(), V3JobOperations.Trigger.Reconciler, configuration, engine, taskParam -> Optional.of(taskParam.toBuilder().withStatus(taskParam.getStatus().toBuilder().withState(TaskState.Finished).withReasonCode(TaskStatus.REASON_STUCK_IN_KILLING_STATE).withReasonMessage("stuck in " + taskState + "state").build()).build()), "TimedOut in KillInitiated state", versionSupplier, titusRuntime, JobManagerConstants.RECONCILER_CALLMETADATA.toBuilder().withCallReason("Kill initiated").build()));
                    } else {
                        actions.add(TaskTimeoutChangeActions.incrementTaskKillAttempt(task.getId(), configuration.getTaskInKillInitiatedStateTimeoutMs(), clock));
                        actions.add(KillInitiatedActions.reconcilerInitiatedTaskKillInitiated(engine, task, runtime, jobStore, versionSupplier, TaskStatus.REASON_STUCK_IN_KILLING_STATE, "Another kill attempt (" + (attempts + 1) + ')', titusRuntime));
                    }
                } else {
                    actions.add(KillInitiatedActions.reconcilerInitiatedTaskKillInitiated(engine, task, runtime, jobStore, versionSupplier, TaskStatus.REASON_STUCK_IN_STATE, "Task stuck in " + taskState + " state", titusRuntime));
                }
                break;
        }
    });
    return actions;
}
Also used : JobManagerConstants(com.netflix.titus.api.jobmanager.service.JobManagerConstants) JobServiceRuntime(com.netflix.titus.master.jobmanager.service.JobServiceRuntime) Task(com.netflix.titus.api.jobmanager.model.job.Task) HashMap(java.util.HashMap) Function(java.util.function.Function) TaskTimeoutChangeActions(com.netflix.titus.master.jobmanager.service.common.action.task.TaskTimeoutChangeActions) ArrayList(java.util.ArrayList) EbsVolume(com.netflix.titus.api.jobmanager.model.job.ebs.EbsVolume) TASK_ATTRIBUTES_EBS_VOLUME_ID(com.netflix.titus.api.jobmanager.TaskAttributes.TASK_ATTRIBUTES_EBS_VOLUME_ID) HashSet(java.util.HashSet) Map(java.util.Map) JobState(com.netflix.titus.api.jobmanager.model.job.JobState) BatchJobExt(com.netflix.titus.api.jobmanager.model.job.ext.BatchJobExt) ChangeAction(com.netflix.titus.common.framework.reconciler.ChangeAction) JobManagerConfiguration(com.netflix.titus.master.jobmanager.service.JobManagerConfiguration) JobStore(com.netflix.titus.api.jobmanager.store.JobStore) JobDescriptor(com.netflix.titus.api.jobmanager.model.job.JobDescriptor) Job(com.netflix.titus.api.jobmanager.model.job.Job) ServiceJobExt(com.netflix.titus.api.jobmanager.model.job.ext.ServiceJobExt) TaskStatus(com.netflix.titus.api.jobmanager.model.job.TaskStatus) Set(java.util.Set) JobFunctions(com.netflix.titus.api.jobmanager.model.job.JobFunctions) Collectors(java.util.stream.Collectors) TaskState(com.netflix.titus.api.jobmanager.model.job.TaskState) EntityHolder(com.netflix.titus.common.framework.reconciler.EntityHolder) Consumer(java.util.function.Consumer) List(java.util.List) ExecutableStatus(com.netflix.titus.api.jobmanager.model.job.ExecutableStatus) V3JobOperations(com.netflix.titus.api.jobmanager.service.V3JobOperations) VersionSupplier(com.netflix.titus.master.jobmanager.service.VersionSupplier) ReconciliationEngine(com.netflix.titus.common.framework.reconciler.ReconciliationEngine) Optional(java.util.Optional) BasicTaskActions(com.netflix.titus.master.jobmanager.service.common.action.task.BasicTaskActions) JobManagerReconcilerEvent(com.netflix.titus.master.jobmanager.service.event.JobManagerReconcilerEvent) TitusRuntime(com.netflix.titus.common.runtime.TitusRuntime) TokenBucket(com.netflix.titus.common.util.limiter.tokenbucket.TokenBucket) Clock(com.netflix.titus.common.util.time.Clock) KillInitiatedActions(com.netflix.titus.master.jobmanager.service.common.action.task.KillInitiatedActions) TASK_ATTRIBUTES_IP_ALLOCATION_ID(com.netflix.titus.api.jobmanager.TaskAttributes.TASK_ATTRIBUTES_IP_ALLOCATION_ID) Task(com.netflix.titus.api.jobmanager.model.job.Task) ChangeAction(com.netflix.titus.common.framework.reconciler.ChangeAction) BatchJobExt(com.netflix.titus.api.jobmanager.model.job.ext.BatchJobExt) ArrayList(java.util.ArrayList) TaskTimeoutChangeActions(com.netflix.titus.master.jobmanager.service.common.action.task.TaskTimeoutChangeActions) Clock(com.netflix.titus.common.util.time.Clock) TaskState(com.netflix.titus.api.jobmanager.model.job.TaskState)

Aggregations

TASK_ATTRIBUTES_EBS_VOLUME_ID (com.netflix.titus.api.jobmanager.TaskAttributes.TASK_ATTRIBUTES_EBS_VOLUME_ID)1 TASK_ATTRIBUTES_IP_ALLOCATION_ID (com.netflix.titus.api.jobmanager.TaskAttributes.TASK_ATTRIBUTES_IP_ALLOCATION_ID)1 ExecutableStatus (com.netflix.titus.api.jobmanager.model.job.ExecutableStatus)1 Job (com.netflix.titus.api.jobmanager.model.job.Job)1 JobDescriptor (com.netflix.titus.api.jobmanager.model.job.JobDescriptor)1 JobFunctions (com.netflix.titus.api.jobmanager.model.job.JobFunctions)1 JobState (com.netflix.titus.api.jobmanager.model.job.JobState)1 Task (com.netflix.titus.api.jobmanager.model.job.Task)1 TaskState (com.netflix.titus.api.jobmanager.model.job.TaskState)1 TaskStatus (com.netflix.titus.api.jobmanager.model.job.TaskStatus)1 EbsVolume (com.netflix.titus.api.jobmanager.model.job.ebs.EbsVolume)1 BatchJobExt (com.netflix.titus.api.jobmanager.model.job.ext.BatchJobExt)1 ServiceJobExt (com.netflix.titus.api.jobmanager.model.job.ext.ServiceJobExt)1 JobManagerConstants (com.netflix.titus.api.jobmanager.service.JobManagerConstants)1 V3JobOperations (com.netflix.titus.api.jobmanager.service.V3JobOperations)1 JobStore (com.netflix.titus.api.jobmanager.store.JobStore)1 ChangeAction (com.netflix.titus.common.framework.reconciler.ChangeAction)1 EntityHolder (com.netflix.titus.common.framework.reconciler.EntityHolder)1 ReconciliationEngine (com.netflix.titus.common.framework.reconciler.ReconciliationEngine)1 TitusRuntime (com.netflix.titus.common.runtime.TitusRuntime)1