func (ks *KubernetesScheduler) recoverTasks() error {
podList, err := ks.client.Pods(api.NamespaceAll).List(labels.Everything(), fields.Everything())
if err != nil {
log.V(1).Infof("failed to recover pod registry, madness may ensue: %v", err)
return err
}
recoverSlave := func(t *podtask.T) {
slaveId := t.Spec.SlaveID
ks.slaveHostNames.Register(slaveId, t.Offer.Host())
}
for _, pod := range podList.Items {
if _, isMirrorPod := pod.Annotations[kubelet.ConfigMirrorAnnotationKey]; isMirrorPod {
// mirrored pods are never reconciled because the scheduler isn't responsible for
// scheduling them; they're started by the executor/kubelet upon instantiation and
// reflected in the apiserver afterward. the scheduler has no knowledge of them.
continue
}
if t, ok, err := podtask.RecoverFrom(pod); err != nil {
log.Errorf("failed to recover task from pod, will attempt to delete '%v/%v': %v", pod.Namespace, pod.Name, err)
err := ks.client.Pods(pod.Namespace).Delete(pod.Name, nil)
//TODO(jdef) check for temporary or not-found errors
if err != nil {
log.Errorf("failed to delete pod '%v/%v': %v", pod.Namespace, pod.Name, err)
}
} else if ok {
ks.taskRegistry.Register(t, nil)
recoverSlave(t)
log.Infof("recovered task %v from pod %v/%v", t.ID, pod.Namespace, pod.Name)
}
}
return nil
}
// 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)
}
}