func fillRunningPodInfo(task *T, taskStatus *mesos.TaskStatus) {
if taskStatus.GetReason() == mesos.TaskStatus_REASON_RECONCILIATION && taskStatus.GetSource() == mesos.TaskStatus_SOURCE_MASTER {
// there is no data..
return
}
//TODO(jdef) determine the usefullness of this information (if any)
if result, err := ParsePodStatusResult(taskStatus); err != nil {
log.Errorf("invalid TaskStatus.Data for task '%v': %v", task.ID, err)
} else {
task.podStatus = result.Status
log.Infof("received pod status for task %v: %+v", task.ID, result.Status)
}
}
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 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()
}
// StatusUpdate method
func (c *Cluster) StatusUpdate(_ mesosscheduler.SchedulerDriver, taskStatus *mesosproto.TaskStatus) {
log.WithFields(log.Fields{"name": "mesos", "state": taskStatus.State.String()}).Debug("Status update")
taskID := taskStatus.TaskId.GetValue()
slaveID := taskStatus.SlaveId.GetValue()
s, ok := c.slaves[slaveID]
if !ok {
return
}
if task, ok := s.tasks[taskID]; ok {
task.sendStatus(taskStatus)
} else {
var reason = ""
if taskStatus.Reason != nil {
reason = taskStatus.GetReason().String()
}
log.WithFields(log.Fields{
"name": "mesos",
"state": taskStatus.State.String(),
"slaveId": taskStatus.SlaveId.GetValue(),
"reason": reason,
}).Warn("Status update received for unknown 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())
}
}
}
// 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())
}
}
