Examples with Clock com.netflix.titus.common.util.time.Clock used on opensource projects

Example 1 with Clock

use of com.netflix.titus.common.util.time.Clock in project titus-control-plane by Netflix.

the class BatchDifferenceResolver method createNewTaskAction.

private Optional<TitusChangeAction> createNewTaskAction(BatchJobView refJobView, int taskIndex, Optional<EntityHolder> previousTask, List<String> unassignedIpAllocations, List<String> ebsVolumeIds) {
    // Safety check
    long numberOfNotFinishedTasks = refJobView.getJobHolder().getChildren().stream().filter(holder -> TaskState.isRunning(((Task) holder.getEntity()).getStatus().getState())).count();
    if (numberOfNotFinishedTasks >= refJobView.getRequiredSize()) {
        titusRuntime.getCodeInvariants().inconsistent("Batch job reconciler attempts to create too many tasks: jobId=%s, requiredSize=%s, current=%s", refJobView.getJob().getId(), refJobView.getRequiredSize(), numberOfNotFinishedTasks);
        return Optional.empty();
    }
    Map<String, String> taskContext = getTaskContext(previousTask, unassignedIpAllocations, ebsVolumeIds);
    JobDescriptor jobDescriptor = refJobView.getJob().getJobDescriptor();
    ApplicationSLA capacityGroupDescriptor = JobManagerUtil.getCapacityGroupDescriptor(jobDescriptor, capacityGroupService);
    String resourcePool = capacityGroupDescriptor.getResourcePool();
    taskContext = CollectionsExt.copyAndAdd(taskContext, ImmutableMap.of(TaskAttributes.TASK_ATTRIBUTES_RESOURCE_POOL, resourcePool, TaskAttributes.TASK_ATTRIBUTES_TIER, capacityGroupDescriptor.getTier().name()));
    TitusChangeAction storeAction = storeWriteRetryInterceptor.apply(createOrReplaceTaskAction(runtime, jobStore, refJobView.getJobHolder(), taskIndex, versionSupplier, clock, taskContext));
    return Optional.of(storeAction);
}

Example 2 with Clock

use of com.netflix.titus.common.util.time.Clock in project titus-control-plane by Netflix.

the class TaskTimeoutChangeActions method setTimeout.

public static TitusChangeAction setTimeout(String taskId, TaskState taskState, long timeoutMs, Clock clock) {
    String tagName = STATE_TAGS.get(taskState);
    Preconditions.checkArgument(tagName != null, "Timeout not tracked for state %s", taskState);
    return TitusChangeAction.newAction("setTimeout").id(taskId).trigger(Trigger.Reconciler).summary("Setting timeout for task in state %s: %s", taskState, DateTimeExt.toTimeUnitString(timeoutMs)).callMetadata(JobManagerConstants.RECONCILER_CALLMETADATA.toBuilder().withCallReason("configure timeout").build()).applyModelUpdate(self -> {
        TitusModelAction modelAction = TitusModelAction.newModelUpdate(self).taskMaybeUpdate(jobHolder -> jobHolder.findById(taskId).map(taskHolder -> {
            EntityHolder newTaskHolder = taskHolder.addTag(tagName, clock.wallTime() + timeoutMs);
            if (taskState == TaskState.KillInitiated) {
                newTaskHolder = newTaskHolder.addTag(KILL_INITIATED_ATTEMPT_TAG, 0);
            }
            return Pair.of(jobHolder.addChild(newTaskHolder), newTaskHolder);
        }));
        return ModelActionHolder.running(modelAction);
    });
}

Example 3 with Clock

use of com.netflix.titus.common.util.time.Clock in project titus-control-plane by Netflix.

the class DefaultV3JobOperations method enterActiveMode.

@Activator
public void enterActiveMode() {
    this.reconciliationFramework = jobReconciliationFrameworkFactory.newInstance();
    // BUG: event stream breaks permanently, and cannot be retried.
    // As we cannot fix the underlying issue yet, we have to be able to discover when it happens.
    AtomicLong eventStreamLastError = new AtomicLong();
    Clock clock = titusRuntime.getClock();
    this.transactionLoggerSubscription = JobTransactionLogger.logEvents(reconciliationFramework, eventStreamLastError, clock);
    PolledMeter.using(titusRuntime.getRegistry()).withName(METRIC_EVENT_STREAM_LAST_ERROR).monitorValue(eventStreamLastError, value -> value.get() <= 0 ? 0 : clock.wallTime() - value.get());
    // Remove finished jobs from the reconciliation framework.
    Observable<JobManagerReconcilerEvent> reconciliationEventsObservable = reconciliationFramework.events().onBackpressureBuffer(OBSERVE_JOBS_BACKPRESSURE_BUFFER_SIZE, () -> logger.warn("Overflowed the buffer size: " + OBSERVE_JOBS_BACKPRESSURE_BUFFER_SIZE), BackpressureOverflow.ON_OVERFLOW_ERROR).doOnSubscribe(() -> {
        List<EntityHolder> entityHolders = reconciliationFramework.orderedView(IndexKind.StatusCreationTime);
        for (EntityHolder entityHolder : entityHolders) {
            handleJobCompletedEvent(entityHolder);
        }
    });
    this.reconcilerEventSubscription = titusRuntime.persistentStream(reconciliationEventsObservable).subscribe(event -> {
        if (event instanceof JobModelUpdateReconcilerEvent) {
            JobModelUpdateReconcilerEvent jobUpdateEvent = (JobModelUpdateReconcilerEvent) event;
            handleJobCompletedEvent(jobUpdateEvent.getChangedEntityHolder());
        }
    }, e -> logger.error("Event stream terminated with an error", e), () -> logger.info("Event stream completed"));
    reconciliationFramework.start();
}

Example 4 with Clock

use of com.netflix.titus.common.util.time.Clock 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;
}