func (k *inMemoryRegistry) UpdateStatus(status *mesos.TaskStatus) (*T, StateType) {
taskId := status.GetTaskId().GetValue()
k.rw.Lock()
defer k.rw.Unlock()
task, state := k._get(taskId)
switch status.GetState() {
case mesos.TaskState_TASK_STAGING:
k.handleTaskStaging(task, state, status)
case mesos.TaskState_TASK_STARTING:
k.handleTaskStarting(task, state, status)
case mesos.TaskState_TASK_RUNNING:
k.handleTaskRunning(task, state, status)
case mesos.TaskState_TASK_FINISHED:
k.handleTaskFinished(task, state, status)
case mesos.TaskState_TASK_FAILED:
k.handleTaskFailed(task, state, status)
case mesos.TaskState_TASK_ERROR:
k.handleTaskError(task, state, status)
case mesos.TaskState_TASK_KILLED:
k.handleTaskKilled(task, state, status)
case mesos.TaskState_TASK_LOST:
k.handleTaskLost(task, state, status)
default:
log.Warningf("unhandled status update for task: %v", taskId)
}
return task.Clone(), state
}
func (sched *NoneScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
taskId := status.GetTaskId().GetValue()
log.Infoln("Status update: task", taskId, "is in state", status.State.Enum().String())
c := sched.queue.GetCommandById(taskId)
if c == nil {
log.Errorln("Unable to find command for task", taskId)
driver.Abort()
}
if c.Status.GetState() == status.GetState() {
// ignore repeated status updates
return
}
c.Status = status
// send status update to CommandHandler
if status.GetState() == mesos.TaskState_TASK_RUNNING {
sched.handler.CommandRunning(c)
} else if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.handler.CommandEnded(c)
sched.handler.CommandFinished(c)
} else if status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED {
sched.handler.CommandEnded(c)
sched.handler.CommandFailed(c)
}
// stop if Commands channel was closed and all tasks are finished
if sched.queue.Closed() && !sched.handler.HasRunningTasks() {
log.Infoln("All tasks finished, stopping framework.")
sched.handler.FinishAllCommands()
driver.Stop(false)
}
}
func (sched *Scheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
log.Infoln("%v of %v tasks finished.", sched.tasksFinished, sched.totalTasks)
}
//TODO if a job is finished, failed, error, lost, killed
// figure out how this impacts dependent jobs and update job graph
/*
//never shut down framework!
if sched.tasksFinished >= sched.totalTasks {
log.Infoln("Total tasks completed, stopping framework.")
driver.Stop(false)
}
*/
/*
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED {
log.Infoln(
"Aborting because task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message", status.GetMessage(),
)
driver.Abort()
}
*/
}
// mesos.Scheduler interface method.
// Invoked when the status of a task has changed.
func (this *TransformScheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
fmt.Printf("Status update: task %s is in state %s\n", status.TaskId.GetValue(), status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_LOST || status.GetState() == mesos.TaskState_TASK_FAILED || status.GetState() == mesos.TaskState_TASK_FINISHED {
this.removeTask(status.GetTaskId())
this.decRunningInstances()
}
}
func (sched *Scheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
log.Infoln("%v of %v tasks finished.", sched.tasksFinished, sched.totalTasks)
}
}
// StatusUpdate is called when a status update message is sent to the scheduler.
func (k *KubernetesScheduler) StatusUpdate(driver bindings.SchedulerDriver, taskStatus *mesos.TaskStatus) {
source, reason := "none", "none"
if taskStatus.Source != nil {
source = (*taskStatus.Source).String()
}
if taskStatus.Reason != nil {
reason = (*taskStatus.Reason).String()
}
taskState := taskStatus.GetState()
metrics.StatusUpdates.WithLabelValues(source, reason, taskState.String()).Inc()
log.Infof(
"task status update %q from %q for task %q on slave %q executor %q for reason %q",
taskState.String(),
source,
taskStatus.TaskId.GetValue(),
taskStatus.SlaveId.GetValue(),
taskStatus.ExecutorId.GetValue(),
reason)
switch taskState {
case mesos.TaskState_TASK_RUNNING, mesos.TaskState_TASK_FINISHED, mesos.TaskState_TASK_STARTING, mesos.TaskState_TASK_STAGING:
if _, state := k.taskRegistry.UpdateStatus(taskStatus); state == podtask.StateUnknown {
if taskState != mesos.TaskState_TASK_FINISHED {
//TODO(jdef) what if I receive this after a TASK_LOST or TASK_KILLED?
//I don't want to reincarnate then.. TASK_LOST is a special case because
//the master is stateless and there are scenarios where I may get TASK_LOST
//followed by TASK_RUNNING.
//TODO(jdef) consider running this asynchronously since there are API server
//calls that may be made
k.reconcileNonTerminalTask(driver, taskStatus)
} // else, we don't really care about FINISHED tasks that aren't registered
return
}
if _, exists := k.slaves.getSlave(taskStatus.GetSlaveId().GetValue()); !exists {
// a registered task has an update reported by a slave that we don't recognize.
// this should never happen! So we don't reconcile it.
log.Errorf("Ignore status %+v because the slave does not exist", taskStatus)
return
}
case mesos.TaskState_TASK_FAILED:
if task, _ := k.taskRegistry.UpdateStatus(taskStatus); task != nil {
if task.Has(podtask.Launched) && !task.Has(podtask.Bound) {
go k.plugin.reconcilePod(task.Pod)
return
}
} else {
// unknown task failed, not much we can do about it
return
}
// last-ditch effort to reconcile our records
fallthrough
case mesos.TaskState_TASK_LOST, mesos.TaskState_TASK_KILLED:
k.reconcileTerminalTask(driver, taskStatus)
}
}
func (ctx *RunOnceApplicationContext) updateTaskState(status *mesos.TaskStatus) {
for _, task := range ctx.tasks {
if task.TaskID == status.GetTaskId().GetValue() {
task.State = status.GetState()
return
}
}
framework.Logger.Warn("Got unexpected status update for unknown task with ID %s", status.GetTaskId().GetValue())
}
func (sched *Scheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
if glog.V(1) {
glog.Infof("status update from task %s in state %s under executor %s on slave %s: %s",
status.GetTaskId().GetValue(),
status.GetState(),
status.GetExecutorId().GetValue(),
status.GetSlaveId().GetValue(),
status.GetMessage(),
)
}
}
// StatusUpdate handles status updates messages received from Mesos master
//
// Currently this method only logs status updates. This might change in the future
func (bw *BasicWorker) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
taskId := status.TaskId.GetValue()
taskStatus := status.GetState()
log.Println("Task", taskId, "is in state", taskStatus.String())
switch taskStatus {
case mesos.TaskState_TASK_RUNNING:
log.Printf("Marking task %s as %s", taskId, taurus.RUNNING)
case mesos.TaskState_TASK_KILLED, mesos.TaskState_TASK_FINISHED,
mesos.TaskState_TASK_FAILED, mesos.TaskState_TASK_LOST:
log.Printf("Marking task %s as %s", taskId, taurus.STOPPED)
}
}
func statusString(status *mesos.TaskStatus) string {
s := fmt.Sprintf("%s %s slave: %s", status.GetTaskId().GetValue(), status.GetState().String(), idString(status.GetSlaveId().GetValue()))
if status.GetState() != mesos.TaskState_TASK_RUNNING {
s += " reason: " + status.GetReason().String()
}
if status.GetMessage() != "" {
s += " message: " + status.GetMessage()
}
return s
}
func (sched *SdcScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
// KillTaskを実行するとTASK_LOSTが検知され、フレームワークが止まる
// driver.KillTask(status.TaskId)
// log.Infoln("!! Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
// return
}
if sched.tasksFinished >= sched.totalTasks {
// log.Infoln("Total tasks completed, stopping framework.")
log.Infoln("Total tasks completed.")
sched.tasksFinished = 0
sched.totalTasks = 0
sched.tasksLaunched = 0
// driver.Stop(false)
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR {
log.Infoln(
"Aborting because task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message", status.GetMessage(),
)
driver.Abort()
}
}
func (sched *MesosRunonceScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.V(1).Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
eventCh <- status
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
}
if sched.tasksFinished >= sched.totalTasks {
log.V(1).Infoln("Total tasks completed, stopping framework.")
driver.Stop(false)
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR {
exitStatus = 1
log.Warningf("mesos TaskStatus: %v", status)
driver.Stop(false)
log.Errorln(
"Aborting because task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message.", status.GetMessage(),
)
}
}
func testErrorStatus(t *testing.T, ctx *RunOnceApplicationContext, driver *mesostest.MockSchedulerDriver, status *mesos.TaskStatus) {
appDone := ctx.StatusUpdate(driver, status)
So(appDone, ShouldBeTrue)
select {
case sts := <-ctx.StatusChan:
So(sts.Error, ShouldNotBeNil)
So(sts.Error.Error(), ShouldContainSubstring, "failed to run on host")
default:
t.Fail()
}
So(ctx.tasks[0].State, ShouldEqual, status.GetState())
}
func (s *StackDeployScheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
Logger.Info("[StatusUpdate] %s", pretty.Status(status))
if status.GetState() == mesos.TaskState_TASK_FINISHED {
driver.ReviveOffers()
}
for _, runner := range MesosTaskRunners {
if runner.StatusUpdate(driver, status) {
return
}
}
Logger.Warn("Received status update that was not handled by any Mesos Task Runner: %s", pretty.Status(status))
}
func (s *Scheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
Logger.Infof("[StatusUpdate] %s", pretty.Status(status))
id := s.idFromTaskId(status.GetTaskId().GetValue())
switch status.GetState() {
case mesos.TaskState_TASK_RUNNING:
s.onTaskStarted(id, status)
case mesos.TaskState_TASK_LOST, mesos.TaskState_TASK_FAILED, mesos.TaskState_TASK_ERROR:
s.onTaskFailed(id, status)
case mesos.TaskState_TASK_FINISHED, mesos.TaskState_TASK_KILLED:
s.onTaskFinished(id, status)
default:
Logger.Warnf("Got unexpected task state %s for task %s", pretty.Status(status), id)
}
s.cluster.Save()
}
func (s *Scheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
Logger.Infof("[StatusUpdate] %s", statusString(status))
slave := s.slaveFromTaskId(status.GetTaskId().GetValue())
if status.GetState() == mesos.TaskState_TASK_FAILED || status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_LOST || status.GetState() == mesos.TaskState_TASK_ERROR ||
status.GetState() == mesos.TaskState_TASK_FINISHED {
s.cluster.Remove(slave)
}
}
func (k *framework) reconcileTerminalTask(driver bindings.SchedulerDriver, taskStatus *mesos.TaskStatus) {
task, state := k.sched.Tasks().UpdateStatus(taskStatus)
if (state == podtask.StateRunning || state == podtask.StatePending) &&
((taskStatus.GetSource() == mesos.TaskStatus_SOURCE_MASTER && taskStatus.GetReason() == mesos.TaskStatus_REASON_RECONCILIATION) ||
(taskStatus.GetSource() == mesos.TaskStatus_SOURCE_SLAVE && taskStatus.GetReason() == mesos.TaskStatus_REASON_EXECUTOR_TERMINATED) ||
(taskStatus.GetSource() == mesos.TaskStatus_SOURCE_SLAVE && taskStatus.GetReason() == mesos.TaskStatus_REASON_EXECUTOR_UNREGISTERED) ||
(taskStatus.GetSource() == mesos.TaskStatus_SOURCE_EXECUTOR && taskStatus.GetMessage() == messages.ContainersDisappeared) ||
(taskStatus.GetSource() == mesos.TaskStatus_SOURCE_EXECUTOR && taskStatus.GetMessage() == messages.KubeletPodLaunchFailed) ||
(taskStatus.GetSource() == mesos.TaskStatus_SOURCE_EXECUTOR && taskStatus.GetMessage() == messages.TaskKilled && !task.Has(podtask.Deleted))) {
//--
// pod-task has metadata that refers to:
// (1) a task that Mesos no longer knows about, or else
// (2) a pod that the Kubelet will never report as "failed"
// (3) a pod that the kubeletExecutor reported as lost (likely due to docker daemon crash/restart)
// (4) a pod that the kubeletExecutor reported as lost because the kubelet didn't manage to launch it (in time)
// (5) a pod that the kubeletExecutor killed, but the scheduler didn't ask for that (maybe killed by the master)
// For now, destroy the pod and hope that there's a replication controller backing it up.
// TODO(jdef) for case #2 don't delete the pod, just update it's status to Failed
pod := &task.Pod
log.Warningf("deleting rogue pod %v/%v for lost task %v", pod.Namespace, pod.Name, task.ID)
if err := k.client.Core().Pods(pod.Namespace).Delete(pod.Name, api.NewDeleteOptions(0)); err != nil && !errors.IsNotFound(err) {
log.Errorf("failed to delete pod %v/%v for terminal task %v: %v", pod.Namespace, pod.Name, task.ID, err)
}
} else if taskStatus.GetReason() == mesos.TaskStatus_REASON_EXECUTOR_TERMINATED || taskStatus.GetReason() == mesos.TaskStatus_REASON_EXECUTOR_UNREGISTERED {
// attempt to prevent dangling pods in the pod and task registries
log.V(1).Infof("request explicit reconciliation to clean up for task %v after executor reported (terminated/unregistered)", taskStatus.TaskId.GetValue())
k.tasksReconciler.RequestExplicit()
} else if taskStatus.GetState() == mesos.TaskState_TASK_LOST && state == podtask.StateRunning && taskStatus.ExecutorId != nil && taskStatus.SlaveId != nil {
//TODO(jdef) this may not be meaningful once we have proper checkpointing and master detection
//If we're reconciling and receive this then the executor may be
//running a task that we need it to kill. It's possible that the framework
//is unrecognized by the master at this point, so KillTask is not guaranteed
//to do anything. The underlying driver transport may be able to send a
//FrameworkMessage directly to the slave to terminate the task.
log.V(2).Info("forwarding TASK_LOST message to executor %v on slave %v", taskStatus.ExecutorId, taskStatus.SlaveId)
data := fmt.Sprintf("%s:%s", messages.TaskLost, task.ID) //TODO(jdef) use a real message type
if _, err := driver.SendFrameworkMessage(taskStatus.ExecutorId, taskStatus.SlaveId, data); err != nil {
log.Error(err.Error())
}
}
}
func Status(status *mesos.TaskStatus) string {
var buffer bytes.Buffer
buffer.WriteString(fmt.Sprintf("%s %s", status.GetTaskId().GetValue(), status.GetState().String()))
if status.GetSlaveId() != nil && status.GetSlaveId().GetValue() != "" {
buffer.WriteString(" slave: ")
buffer.WriteString(ID(status.GetSlaveId().GetValue()))
}
if status.GetState() != mesos.TaskState_TASK_RUNNING {
buffer.WriteString(" reason: ")
buffer.WriteString(status.GetReason().String())
}
if status.GetMessage() != "" {
buffer.WriteString(" message: ")
buffer.WriteString(status.GetMessage())
}
return buffer.String()
}
// SendStatusUpdate sends status updates to the slave.
func (driver *MesosExecutorDriver) SendStatusUpdate(taskStatus *mesosproto.TaskStatus) (mesosproto.Status, error) {
log.V(3).Infoln("Sending task status update: ", taskStatus.String())
if stat := driver.Status(); stat != mesosproto.Status_DRIVER_RUNNING {
return stat, fmt.Errorf("Unable to SendStatusUpdate, expecting driver.status %s, but got %s", mesosproto.Status_DRIVER_RUNNING, stat)
}
if taskStatus.GetState() == mesosproto.TaskState_TASK_STAGING {
err := fmt.Errorf("Executor is not allowed to send TASK_STAGING status update. Aborting!")
log.Errorln(err)
if err0 := driver.stop(mesosproto.Status_DRIVER_ABORTED); err0 != nil {
log.Errorln("Error while stopping the driver", err0)
}
return driver.Status(), err
}
// Set up status update.
update := driver.makeStatusUpdate(taskStatus)
log.Infof("Executor sending status update %v\n", update.String())
// Capture the status update.
driver.lock.Lock()
driver.updates[uuid.UUID(update.GetUuid()).String()] = update
driver.lock.Unlock()
// Put the status update in the message.
message := &mesosproto.StatusUpdateMessage{
Update: update,
Pid: proto.String(driver.self.String()),
}
// Send the message.
if err := driver.send(driver.slaveUPID, message); err != nil {
log.Errorf("Failed to send %v: %v\n", message, err)
return driver.status, err
}
return driver.Status(), nil
}
// mesos.Scheduler interface method.
// Invoked when the status of a task has changed.
func (this *ElodinaTransportScheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
if *status.GetState().Enum() == mesos.TaskState_TASK_RUNNING {
this.taskIdToTaskState[*status.TaskId.Value].pending = true
} else if status.GetState() == mesos.TaskState_TASK_LOST || status.GetState() == mesos.TaskState_TASK_FAILED || status.GetState() == mesos.TaskState_TASK_FINISHED {
this.TakenTopicPartitions.RemoveAll(this.taskIdToTaskState[*status.TaskId.Value].GetAssignment())
delete(this.taskIdToTaskState, *status.TaskId.Value)
}
}
// mesos.Scheduler interface method.
// Invoked when the status of a task has changed.
func (this *ElodinaTransportScheduler) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) {
log.Logger.Info("Received status %s for task %s", status.GetState().Enum(), status.TaskId.GetValue())
if status.GetState() == mesos.TaskState_TASK_RUNNING {
this.taskIdToTaskState[status.TaskId.GetValue()].pending = true
} else if isTerminated(status.GetState()) {
this.TakenTopicPartitions.RemoveAll(this.taskIdToTaskState[status.TaskId.GetValue()].GetAssignment())
delete(this.taskIdToTaskState, status.TaskId.GetValue())
}
}
// Will obtain the log file you desire from the mesos-agent and react to the status of the message accordingly.
func FetchLogs(status *mesos.TaskStatus, offset int, file string, frameworkId string) ([]byte, error) {
var (
dir string
hostname string
err error
)
switch status.GetState() {
case mesos.TaskState_TASK_FAILED,
mesos.TaskState_TASK_KILLED:
hostDir, err := hostDirFromState(status, frameworkId)
if err != nil {
return nil, err
}
hostname, dir = hostDir.Host, hostDir.Dir
default:
hostDir, err := hostDirFromTaskStatus(status)
if err != nil {
return nil, err
}
hostname, dir = hostDir.Host, hostDir.Dir
}
url := fmt.Sprintf("http://%s:5051/files/read.json?path=%s/%s&offset=%d",
hostname, dir, file, offset)
bodyData, err := fetchUrl(url)
if err != nil {
return nil, err
}
var logData LogData
err = json.Unmarshal(bodyData, &logData)
if err != nil {
return nil, err
}
return []byte(logData.Data), nil
}
func (ctx *RunOnceApplicationContext) StatusUpdate(driver scheduler.SchedulerDriver, status *mesos.TaskStatus) bool {
ctx.lock.Lock()
defer ctx.lock.Unlock()
hostname := hostnameFromTaskID(status.GetTaskId().GetValue())
ctx.updateTaskState(status)
switch status.GetState() {
case mesos.TaskState_TASK_RUNNING:
log.Infof("Task %s received status update in state %s", status.GetTaskId().GetValue(), status.GetState().String())
case mesos.TaskState_TASK_LOST, mesos.TaskState_TASK_FAILED, mesos.TaskState_TASK_ERROR:
//TODO also kill all other running tasks sometime?
ctx.StatusChan <- framework.NewApplicationRunStatus(ctx.Application, fmt.Errorf("Application %s failed to run on host %s with status %s: %s", ctx.Application.ID, hostname, status.GetState().String(), status.GetMessage()))
return true
case mesos.TaskState_TASK_FINISHED, mesos.TaskState_TASK_KILLED:
if ctx.allTasksFinished() {
ctx.StatusChan <- framework.NewApplicationRunStatus(ctx.Application, nil)
return true
}
default:
log.Warningf("Got unexpected task state %s", pretty.Status(status))
}
return false
}
func (sched *ExampleScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
driver.ReviveOffers() // TODO(jdef) rate-limit this
}
if sched.tasksFinished >= sched.totalTasks {
log.Infoln("Total tasks completed, stopping framework.")
driver.Stop(false)
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR {
sched.tasksErrored++
}
}
func (sched *ExampleScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
}
if sched.tasksFinished >= sched.totalTasks {
log.Infoln("Total tasks completed, stopping framework.")
driver.Stop(false)
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR {
log.Infoln(
"Aborting because task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message", status.GetMessage(),
)
driver.Abort()
}
}
func (sched *ScraperScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
if status.GetState() == mesos.TaskState_TASK_FINISHED {
sched.tasksFinished++
log.Infof("%v of %v tasks finished.", sched.tasksFinished, sched.totalTasks)
uri := string(status.Data)
db, err := sql.Open("sqlite3", "./database.db")
if err != nil {
log.Fatal("Failed to connect to database: %v\n", err)
}
defer db.Close()
tx, err := db.Begin()
if err != nil {
log.Fatal(err)
}
stmt, err := tx.Prepare("insert into runs(uri, storage_path, last_scrape_time) values(?, ?, ?)")
if err != nil {
log.Fatal(err)
}
path := base64.StdEncoding.EncodeToString([]byte(uri))
_, err = stmt.Exec(uri, path, time.Now().Unix())
if err != nil {
log.Fatal(err)
}
defer stmt.Close()
tx.Commit()
}
if sched.tasksFinished >= sched.totalTasks {
log.Infoln("Tasks that we know about are done!")
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED {
log.Infoln(
"Aborting because task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message", status.GetMessage(),
)
driver.Abort()
}
}
// reconcile an unknown (from the perspective of our registry) non-terminal task
func (k *KubernetesScheduler) reconcileNonTerminalTask(driver bindings.SchedulerDriver, taskStatus *mesos.TaskStatus) {
// attempt to recover task from pod info:
// - task data may contain an api.PodStatusResult; if status.reason == REASON_RECONCILIATION then status.data == nil
// - the Name can be parsed by container.ParseFullName() to yield a pod Name and Namespace
// - pull the pod metadata down from the api server
// - perform task recovery based on pod metadata
taskId := taskStatus.TaskId.GetValue()
if taskStatus.GetReason() == mesos.TaskStatus_REASON_RECONCILIATION && taskStatus.GetSource() == mesos.TaskStatus_SOURCE_MASTER {
// there will be no data in the task status that we can use to determine the associated pod
switch taskStatus.GetState() {
case mesos.TaskState_TASK_STAGING:
// there is still hope for this task, don't kill it just yet
//TODO(jdef) there should probably be a limit for how long we tolerate tasks stuck in this state
return
default:
// for TASK_{STARTING,RUNNING} we should have already attempted to recoverTasks() for.
// if the scheduler failed over before the executor fired TASK_STARTING, then we should *not*
// be processing this reconciliation update before we process the one from the executor.
// point: we don't know what this task is (perhaps there was unrecoverable metadata in the pod),
// so it gets killed.
log.Errorf("killing non-terminal, unrecoverable task %v", taskId)
}
} else if podStatus, err := podtask.ParsePodStatusResult(taskStatus); err != nil {
// possible rogue pod exists at this point because we can't identify it; should kill the task
log.Errorf("possible rogue pod; illegal task status data for task %v, expected an api.PodStatusResult: %v", taskId, err)
} else if name, namespace, err := container.ParsePodFullName(podStatus.Name); err != nil {
// possible rogue pod exists at this point because we can't identify it; should kill the task
log.Errorf("possible rogue pod; illegal api.PodStatusResult, unable to parse full pod name from: '%v' for task %v: %v",
podStatus.Name, taskId, err)
} else if pod, err := k.client.Pods(namespace).Get(name); err == nil {
if t, ok, err := podtask.RecoverFrom(*pod); ok {
log.Infof("recovered task %v from metadata in pod %v/%v", taskId, namespace, name)
_, err := k.taskRegistry.Register(t, nil)
if err != nil {
// someone beat us to it?!
log.Warningf("failed to register recovered task: %v", err)
return
} else {
k.taskRegistry.UpdateStatus(taskStatus)
}
return
} else if err != nil {
//should kill the pod and the task
log.Errorf("killing pod, failed to recover task from pod %v/%v: %v", namespace, name, err)
if err := k.client.Pods(namespace).Delete(name, nil); err != nil {
log.Errorf("failed to delete pod %v/%v: %v", namespace, name, err)
}
} else {
//this is pretty unexpected: we received a TASK_{STARTING,RUNNING} message, but the apiserver's pod
//metadata is not appropriate for task reconstruction -- which should almost certainly never
//be the case unless someone swapped out the pod on us (and kept the same namespace/name) while
//we were failed over.
//kill this task, allow the newly launched scheduler to schedule the new pod
log.Warningf("unexpected pod metadata for task %v in apiserver, assuming new unscheduled pod spec: %+v", taskId, pod)
}
} else if errors.IsNotFound(err) {
// pod lookup failed, should delete the task since the pod is no longer valid; may be redundant, that's ok
log.Infof("killing task %v since pod %v/%v no longer exists", taskId, namespace, name)
} else if errors.IsServerTimeout(err) {
log.V(2).Infof("failed to reconcile task due to API server timeout: %v", err)
return
} else {
log.Errorf("unexpected API server error, aborting reconcile for task %v: %v", taskId, err)
return
}
if _, err := driver.KillTask(taskStatus.TaskId); err != nil {
log.Errorf("failed to kill task %v: %v", taskId, err)
}
}
func (s *EtcdScheduler) StatusUpdate(
driver scheduler.SchedulerDriver,
status *mesos.TaskStatus,
) {
s.mut.Lock()
defer s.mut.Unlock()
log.Infoln(
"Status update: task",
status.TaskId.GetValue(),
" is in state ",
status.State.Enum().String(),
)
node, err := config.Parse(status.GetTaskId().GetValue())
if err != nil {
log.Errorf("scheduler: failed to unmarshal config.Node from TaskId: %s", err)
return
}
node.SlaveID = status.SlaveId.GetValue()
// record that we've heard about this task
s.heardFrom[status.GetTaskId().GetValue()] = struct{}{}
switch status.GetState() {
case mesos.TaskState_TASK_LOST,
mesos.TaskState_TASK_FINISHED,
mesos.TaskState_TASK_KILLED,
mesos.TaskState_TASK_ERROR,
mesos.TaskState_TASK_FAILED:
log.Errorf("Task contraction: %+v", status.GetState())
log.Errorf("message: %s", status.GetMessage())
log.Errorf("reason: %+v", status.GetReason())
atomic.AddUint32(&s.Stats.FailedServers, 1)
// TODO(tyler) kill this
// Pump the brakes so that we have time to deconfigure the lost node
// before adding a new one. If we don't deconfigure first, we risk
// split brain.
s.PumpTheBrakes()
// now we know this task is dead
delete(s.pending, node.Name)
delete(s.running, node.Name)
delete(s.tasks, node.Name)
// We don't have to clean up the state in ZK for this
// as it is fine to eventually just persist when we
// receive a new TASK_RUNNING below.
delete(s.reconciliationInfo, status.TaskId.GetValue())
s.QueueLaunchAttempt()
// TODO(tyler) do we want to lock if the first task fails?
// TODO(tyler) can we handle a total loss at reconciliation time,
// when s.state == Immutable?
if len(s.running) == 0 && s.state == Mutable {
log.Error("TOTAL CLUSTER LOSS! LOCKING SCHEDULER, " +
"FOLLOW RESTORATION GUIDE AT " +
"https://github.com/mesosphere/" +
"etcd-mesos/blob/master/docs/response.md")
s.state = Immutable
}
case mesos.TaskState_TASK_STARTING:
case mesos.TaskState_TASK_RUNNING:
// We update data to ZK synchronously because it must happen
// in-order. If we spun off a goroutine this would possibly retry
// and succeed in the wrong order, and older data would win.
// We keep this simple here, as if ZK is healthy this won't take long.
// If this takes long, we're probably about to die anyway, as ZK is
// displeased and mesos-go will panic when it loses contact.
s.reconciliationInfo[status.TaskId.GetValue()] = status.SlaveId.GetValue()
err = s.updateReconciliationInfoFunc(
s.reconciliationInfo,
s.ZkServers,
s.ZkChroot,
s.FrameworkName,
)
if err != nil {
log.Errorf("Failed to persist reconciliation info: %+v", err)
}
delete(s.pending, node.Name)
_, present := s.running[node.Name]
if !present {
s.running[node.Name] = node
s.tasks[node.Name] = status.TaskId
}
// During reconcilliation, we may find nodes with higher ID's due to ntp drift
etcdIndexParts := strings.Split(node.Name, "-")
if len(etcdIndexParts) != 2 {
log.Warning("Task has a Name that does not follow the form etcd-<index>")
} else {
etcdIndex, err := strconv.ParseInt(etcdIndexParts[1], 10, 64)
if err != nil {
log.Warning("Task has a Name that does not follow the form etcd-<index>")
} else {
if etcdIndex > s.highestInstanceID {
s.highestInstanceID = etcdIndex + 1
}
}
}
default:
log.Warningf("Received unhandled task state: %+v", status.GetState())
}
}
func (sched *ExampleScheduler) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
log.Infoln("Status update: task", status.TaskId.GetValue(), " is in state ", status.State.Enum().String())
for _, task := range sched.tasks {
if task.name == status.TaskId.GetValue() &&
(status.GetState() == mesos.TaskState_TASK_FINISHED ||
status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR) {
// No matter what the outcome was, move to finished state so that we can unreserve resources
task.state = FinishedState
}
}
if status.GetState() == mesos.TaskState_TASK_LOST ||
status.GetState() == mesos.TaskState_TASK_KILLED ||
status.GetState() == mesos.TaskState_TASK_FAILED ||
status.GetState() == mesos.TaskState_TASK_ERROR {
log.Infoln(
"Task", status.TaskId.GetValue(),
"is in unexpected state", status.State.String(),
"with message", status.GetMessage(),
". Unreserving resources",
)
}
}
