// this pod may be out of sync with respect to the API server registry:
// this pod | apiserver registry
// -------------|----------------------
// host=.* | 404 ; pod was deleted
// host=.* | 5xx ; failed to sync, try again later?
// host="" | host="" ; perhaps no updates to process?
// host="" | host="..." ; pod has been scheduled and assigned, is there a task assigned? (check TaskIdKey in binding?)
// host="..." | host="" ; pod is no longer scheduled, does it need to be re-queued?
// host="..." | host="..." ; perhaps no updates to process?
//
// TODO(jdef) this needs an integration test
func (s *schedulingPlugin) reconcileTask(t *podtask.T) {
log.V(1).Infof("reconcile pod %v, assigned to slave %q", t.Pod.Name, t.Spec.AssignedSlave)
ctx := api.WithNamespace(api.NewDefaultContext(), t.Pod.Namespace)
pod, err := s.client.Pods(api.NamespaceValue(ctx)).Get(t.Pod.Name)
if err != nil {
if errors.IsNotFound(err) {
// attempt to delete
if err = s.deleter.deleteOne(&Pod{Pod: &t.Pod}); err != nil && err != noSuchPodErr && err != noSuchTaskErr {
log.Errorf("failed to delete pod: %v: %v", t.Pod.Name, err)
}
} else {
//TODO(jdef) other errors should probably trigger a retry (w/ backoff).
//For now, drop the pod on the floor
log.Warning("aborting reconciliation for pod %v: %v", t.Pod.Name, err)
}
return
}
log.Infof("pod %v scheduled on %q according to apiserver", pod.Name, pod.Spec.NodeName)
if t.Spec.AssignedSlave != pod.Spec.NodeName {
if pod.Spec.NodeName == "" {
// pod is unscheduled.
// it's possible that we dropped the pod in the scheduler error handler
// because of task misalignment with the pod (task.Has(podtask.Launched) == true)
podKey, err := podtask.MakePodKey(ctx, pod.Name)
if err != nil {
log.Error(err)
return
}
s.api.Lock()
defer s.api.Unlock()
if _, state := s.api.tasks().ForPod(podKey); state != podtask.StateUnknown {
//TODO(jdef) reconcile the task
log.Errorf("task already registered for pod %v", pod.Name)
return
}
now := time.Now()
log.V(3).Infof("reoffering pod %v", podKey)
s.qr.reoffer(&Pod{
Pod: pod,
deadline: &now,
})
} else {
// pod is scheduled.
// not sure how this happened behind our backs. attempt to reconstruct
// at least a partial podtask.T record.
//TODO(jdef) reconcile the task
log.Errorf("pod already scheduled: %v", pod.Name)
}
} else {
//TODO(jdef) for now, ignore the fact that the rest of the spec may be different
//and assume that our knowledge of the pod aligns with that of the apiserver
log.Error("pod reconciliation does not support updates; not yet implemented")
}
}
func (k *deleter) deleteOne(pod *Pod) error {
ctx := api.WithNamespace(api.NewDefaultContext(), pod.Namespace)
podKey, err := podtask.MakePodKey(ctx, pod.Name)
if err != nil {
return err
}
log.V(2).Infof("pod deleted: %v", podKey)
// order is important here: we want to make sure we have the lock before
// removing the pod from the scheduling queue. this makes the concurrent
// execution of scheduler-error-handling and delete-handling easier to
// reason about.
k.api.Lock()
defer k.api.Unlock()
// prevent the scheduler from attempting to pop this; it's also possible that
// it's concurrently being scheduled (somewhere between pod scheduling and
// binding) - if so, then we'll end up removing it from taskRegistry which
// will abort Bind()ing
k.qr.dequeue(pod.GetUID())
switch task, state := k.api.tasks().ForPod(podKey); state {
case podtask.StateUnknown:
log.V(2).Infof("Could not resolve pod '%s' to task id", podKey)
return noSuchPodErr
// determine if the task has already been launched to mesos, if not then
// cleanup is easier (unregister) since there's no state to sync
case podtask.StatePending:
if !task.Has(podtask.Launched) {
// we've been invoked in between Schedule() and Bind()
if task.HasAcceptedOffer() {
task.Offer.Release()
task.Reset()
task.Set(podtask.Deleted)
//TODO(jdef) probably want better handling here
if err := k.api.tasks().Update(task); err != nil {
return err
}
}
k.api.tasks().Unregister(task)
return nil
}
fallthrough
case podtask.StateRunning:
// signal to watchers that the related pod is going down
task.Set(podtask.Deleted)
if err := k.api.tasks().Update(task); err != nil {
log.Errorf("failed to update task w/ Deleted status: %v", err)
}
return k.api.killTask(task.ID)
default:
log.Infof("cannot kill pod '%s': non-terminal task not found %v", podKey, task.ID)
return noSuchTaskErr
}
}
// implementation of scheduling plugin's Error func; see plugin/pkg/scheduler
func (k *errorHandler) handleSchedulingError(pod *api.Pod, schedulingErr error) {
if schedulingErr == noSuchPodErr {
log.V(2).Infof("Not rescheduling non-existent pod %v", pod.Name)
return
}
log.Infof("Error scheduling %v: %v; retrying", pod.Name, schedulingErr)
defer util.HandleCrash()
// default upstream scheduler passes pod.Name as binding.PodID
ctx := api.WithNamespace(api.NewDefaultContext(), pod.Namespace)
podKey, err := podtask.MakePodKey(ctx, pod.Name)
if err != nil {
log.Errorf("Failed to construct pod key, aborting scheduling for pod %v: %v", pod.Name, err)
return
}
k.backoff.GC()
k.api.Lock()
defer k.api.Unlock()
switch task, state := k.api.tasks().ForPod(podKey); state {
case podtask.StateUnknown:
// if we don't have a mapping here any more then someone deleted the pod
log.V(2).Infof("Could not resolve pod to task, aborting pod reschdule: %s", podKey)
return
case podtask.StatePending:
if task.Has(podtask.Launched) {
log.V(2).Infof("Skipping re-scheduling for already-launched pod %v", podKey)
return
}
breakoutEarly := queue.BreakChan(nil)
if schedulingErr == noSuitableOffersErr {
log.V(3).Infof("adding backoff breakout handler for pod %v", podKey)
breakoutEarly = queue.BreakChan(k.api.offers().Listen(podKey, func(offer *mesos.Offer) bool {
k.api.Lock()
defer k.api.Unlock()
switch task, state := k.api.tasks().Get(task.ID); state {
case podtask.StatePending:
return !task.Has(podtask.Launched) && k.api.algorithm().FitPredicate()(task, offer)
default:
// no point in continuing to check for matching offers
return true
}
}))
}
delay := k.backoff.Get(podKey)
log.V(3).Infof("requeuing pod %v with delay %v", podKey, delay)
k.qr.requeue(&Pod{Pod: pod, delay: &delay, notify: breakoutEarly})
default:
log.V(2).Infof("Task is no longer pending, aborting reschedule for pod %v", podKey)
}
}
// Schedule implements the Scheduler interface of Kubernetes.
// It returns the selectedMachine's name and error (if there's any).
func (k *kubeScheduler) Schedule(pod *api.Pod, unused algorithm.NodeLister) (string, error) {
log.Infof("Try to schedule pod %v\n", pod.Name)
ctx := api.WithNamespace(api.NewDefaultContext(), pod.Namespace)
// default upstream scheduler passes pod.Name as binding.PodID
podKey, err := podtask.MakePodKey(ctx, pod.Name)
if err != nil {
return "", err
}
k.api.Lock()
defer k.api.Unlock()
switch task, state := k.api.tasks().ForPod(podKey); state {
case podtask.StateUnknown:
// There's a bit of a potential race here, a pod could have been yielded() and
// then before we get *here* it could be deleted.
// We use meta to index the pod in the store since that's what k8s reflector does.
podName, err := cache.MetaNamespaceKeyFunc(pod)
if err != nil {
log.Warningf("aborting Schedule, unable to understand pod object %+v", pod)
return "", noSuchPodErr
}
if deleted := k.podUpdates.Poll(podName, queue.DELETE_EVENT); deleted {
// avoid scheduling a pod that's been deleted between yieldPod() and Schedule()
log.Infof("aborting Schedule, pod has been deleted %+v", pod)
return "", noSuchPodErr
}
return k.doSchedule(k.api.tasks().Register(k.api.createPodTask(ctx, pod)))
//TODO(jdef) it's possible that the pod state has diverged from what
//we knew previously, we should probably update the task.Pod state here
//before proceeding with scheduling
case podtask.StatePending:
if pod.UID != task.Pod.UID {
// we're dealing with a brand new pod spec here, so the old one must have been
// deleted -- and so our task store is out of sync w/ respect to reality
//TODO(jdef) reconcile task
return "", fmt.Errorf("task %v spec is out of sync with pod %v spec, aborting schedule", task.ID, pod.Name)
} else if task.Has(podtask.Launched) {
// task has been marked as "launched" but the pod binding creation may have failed in k8s,
// but we're going to let someone else handle it, probably the mesos task error handler
return "", fmt.Errorf("task %s has already been launched, aborting schedule", task.ID)
} else {
return k.doSchedule(task, nil)
}
default:
return "", fmt.Errorf("task %s is not pending, nothing to schedule", task.ID)
}
}
// implements binding.Registry, launches the pod-associated-task in mesos
func (b *binder) Bind(binding *api.Binding) error {
ctx := api.WithNamespace(api.NewContext(), binding.Namespace)
// default upstream scheduler passes pod.Name as binding.Name
podKey, err := podtask.MakePodKey(ctx, binding.Name)
if err != nil {
return err
}
b.api.Lock()
defer b.api.Unlock()
switch task, state := b.api.tasks().ForPod(podKey); state {
case podtask.StatePending:
return b.bind(ctx, binding, task)
default:
// in this case it's likely that the pod has been deleted between Schedule
// and Bind calls
log.Infof("No pending task for pod %s", podKey)
return noSuchPodErr //TODO(jdef) this error is somewhat misleading since the task could be running?!
}
}
// Test to create the scheduler plugin with the config returned by the scheduler,
// and play through the whole life cycle of the plugin while creating pods, deleting
// and failing them.
func TestPlugin_LifeCycle(t *testing.T) {
t.Skip("This test is flaky, see #11901")
assert := &EventAssertions{*assert.New(t)}
// create a fake pod watch. We use that below to submit new pods to the scheduler
podListWatch := NewMockPodsListWatch(api.PodList{})
// create fake apiserver
testApiServer := NewTestServer(t, api.NamespaceDefault, podListWatch)
defer testApiServer.server.Close()
// create executor with some data for static pods if set
executor := util.NewExecutorInfo(
util.NewExecutorID("executor-id"),
util.NewCommandInfo("executor-cmd"),
)
executor.Data = []byte{0, 1, 2}
// create scheduler
as := NewAllocationStrategy(
podtask.DefaultPredicate,
podtask.NewDefaultProcurement(mresource.DefaultDefaultContainerCPULimit, mresource.DefaultDefaultContainerMemLimit))
testScheduler := New(Config{
Executor: executor,
Client: client.NewOrDie(&client.Config{Host: testApiServer.server.URL, Version: testapi.Default.Version()}),
Scheduler: NewFCFSPodScheduler(as),
Schedcfg: *schedcfg.CreateDefaultConfig(),
})
assert.NotNil(testScheduler.client, "client is nil")
assert.NotNil(testScheduler.executor, "executor is nil")
assert.NotNil(testScheduler.offers, "offer registry is nil")
// create scheduler process
schedulerProcess := ha.New(testScheduler)
// get plugin config from it
c := testScheduler.NewPluginConfig(schedulerProcess.Terminal(), http.DefaultServeMux, &podListWatch.ListWatch)
assert.NotNil(c)
// make events observable
eventObserver := NewEventObserver()
c.Recorder = eventObserver
// create plugin
p := NewPlugin(c).(*schedulingPlugin)
assert.NotNil(p)
// run plugin
p.Run(schedulerProcess.Terminal())
defer schedulerProcess.End()
// init scheduler
err := testScheduler.Init(schedulerProcess.Master(), p, http.DefaultServeMux)
assert.NoError(err)
// create mock mesos scheduler driver
mockDriver := &joinableDriver{}
mockDriver.On("Start").Return(mesos.Status_DRIVER_RUNNING, nil).Once()
started := mockDriver.Upon()
mAny := mock.AnythingOfType
mockDriver.On("ReconcileTasks", mAny("[]*mesosproto.TaskStatus")).Return(mesos.Status_DRIVER_RUNNING, nil)
mockDriver.On("SendFrameworkMessage", mAny("*mesosproto.ExecutorID"), mAny("*mesosproto.SlaveID"), mAny("string")).
Return(mesos.Status_DRIVER_RUNNING, nil)
type LaunchedTask struct {
offerId mesos.OfferID
taskInfo *mesos.TaskInfo
}
launchedTasks := make(chan LaunchedTask, 1)
launchTasksCalledFunc := func(args mock.Arguments) {
offerIDs := args.Get(0).([]*mesos.OfferID)
taskInfos := args.Get(1).([]*mesos.TaskInfo)
assert.Equal(1, len(offerIDs))
assert.Equal(1, len(taskInfos))
launchedTasks <- LaunchedTask{
offerId: *offerIDs[0],
taskInfo: taskInfos[0],
}
}
mockDriver.On("LaunchTasks", mAny("[]*mesosproto.OfferID"), mAny("[]*mesosproto.TaskInfo"), mAny("*mesosproto.Filters")).
Return(mesos.Status_DRIVER_RUNNING, nil).Run(launchTasksCalledFunc)
mockDriver.On("DeclineOffer", mAny("*mesosproto.OfferID"), mAny("*mesosproto.Filters")).
Return(mesos.Status_DRIVER_RUNNING, nil)
// elect master with mock driver
driverFactory := ha.DriverFactory(func() (bindings.SchedulerDriver, error) {
return mockDriver, nil
})
schedulerProcess.Elect(driverFactory)
elected := schedulerProcess.Elected()
// driver will be started
<-started
// tell scheduler to be registered
testScheduler.Registered(
mockDriver,
util.NewFrameworkID("kubernetes-id"),
util.NewMasterInfo("master-id", (192<<24)+(168<<16)+(0<<8)+1, 5050),
)
// wait for being elected
<-elected
//TODO(jdef) refactor things above here into a test suite setup of some sort
// fake new, unscheduled pod
pod, i := NewTestPod()
podListWatch.Add(pod, true) // notify watchers
// wait for failedScheduling event because there is no offer
assert.EventWithReason(eventObserver, "failedScheduling", "failedScheduling event not received")
// add some matching offer
offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
testScheduler.ResourceOffers(nil, offers)
// and wait for scheduled pod
assert.EventWithReason(eventObserver, "scheduled")
select {
case launchedTask := <-launchedTasks:
// report back that the task has been staged, and then started by mesos
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_STAGING))
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_RUNNING))
// check that ExecutorInfo.data has the static pod data
assert.Len(launchedTask.taskInfo.Executor.Data, 3)
// report back that the task has been lost
mockDriver.AssertNumberOfCalls(t, "SendFrameworkMessage", 0)
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_LOST))
// and wait that framework message is sent to executor
mockDriver.AssertNumberOfCalls(t, "SendFrameworkMessage", 1)
case <-time.After(5 * time.Second):
t.Fatalf("timed out waiting for launchTasks call")
}
// Launch a pod and wait until the scheduler driver is called
schedulePodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
// wait for failedScheduling event because there is no offer
assert.EventWithReason(eventObserver, "failedScheduling", "failedScheduling event not received")
// supply a matching offer
testScheduler.ResourceOffers(mockDriver, offers)
// and wait to get scheduled
assert.EventWithReason(eventObserver, "scheduled")
// wait for driver.launchTasks call
select {
case launchedTask := <-launchedTasks:
for _, offer := range offers {
if offer.Id.GetValue() == launchedTask.offerId.GetValue() {
return pod, &launchedTask, offer
}
}
t.Fatalf("unknown offer used to start a pod")
return nil, nil, nil
case <-time.After(5 * time.Second):
t.Fatal("timed out waiting for launchTasks")
return nil, nil, nil
}
}
// Launch a pod and wait until the scheduler driver is called
launchPodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
podListWatch.Add(pod, true)
return schedulePodWithOffers(pod, offers)
}
// Launch a pod, wait until the scheduler driver is called and report back that it is running
startPodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
// notify about pod, offer resources and wait for scheduling
pod, launchedTask, offer := launchPodWithOffers(pod, offers)
if pod != nil {
// report back status
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_STAGING))
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_RUNNING))
return pod, launchedTask, offer
}
return nil, nil, nil
}
startTestPod := func() (*api.Pod, *LaunchedTask, *mesos.Offer) {
pod, i := NewTestPod()
offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
return startPodWithOffers(pod, offers)
}
// start another pod
pod, launchedTask, _ := startTestPod()
// mock drvier.KillTask, should be invoked when a pod is deleted
mockDriver.On("KillTask", mAny("*mesosproto.TaskID")).Return(mesos.Status_DRIVER_RUNNING, nil).Run(func(args mock.Arguments) {
killedTaskId := *(args.Get(0).(*mesos.TaskID))
assert.Equal(*launchedTask.taskInfo.TaskId, killedTaskId, "expected same TaskID as during launch")
})
killTaskCalled := mockDriver.Upon()
// stop it again via the apiserver mock
podListWatch.Delete(pod, true) // notify watchers
// and wait for the driver killTask call with the correct TaskId
select {
case <-killTaskCalled:
// report back that the task is finished
testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_FINISHED))
case <-time.After(5 * time.Second):
t.Fatal("timed out waiting for KillTask")
}
// start a pod with on a given NodeName and check that it is scheduled to the right host
pod, i = NewTestPod()
pod.Spec.NodeName = "hostname1"
offers = []*mesos.Offer{}
for j := 0; j < 3; j++ {
offer := NewTestOffer(fmt.Sprintf("offer%d_%d", i, j))
hostname := fmt.Sprintf("hostname%d", j)
offer.Hostname = &hostname
offers = append(offers, offer)
}
_, _, usedOffer := startPodWithOffers(pod, offers)
assert.Equal(offers[1].Id.GetValue(), usedOffer.Id.GetValue())
assert.Equal(pod.Spec.NodeName, *usedOffer.Hostname)
testScheduler.OfferRescinded(mockDriver, offers[0].Id)
testScheduler.OfferRescinded(mockDriver, offers[2].Id)
// start pods:
// - which are failing while binding,
// - leading to reconciliation
// - with different states on the apiserver
failPodFromExecutor := func(task *mesos.TaskInfo) {
beforePodLookups := testApiServer.Stats(pod.Name)
status := newTaskStatusForTask(task, mesos.TaskState_TASK_FAILED)
message := messages.CreateBindingFailure
status.Message = &message
testScheduler.StatusUpdate(mockDriver, status)
// wait until pod is looked up at the apiserver
assertext.EventuallyTrue(t, time.Second, func() bool {
return testApiServer.Stats(pod.Name) == beforePodLookups+1
}, "expect that reconcileTask will access apiserver for pod %v", pod.Name)
}
launchTestPod := func() (*api.Pod, *LaunchedTask, *mesos.Offer) {
pod, i := NewTestPod()
offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
return launchPodWithOffers(pod, offers)
}
// 1. with pod deleted from the apiserver
// expected: pod is removed from internal task registry
pod, launchedTask, _ = launchTestPod()
podListWatch.Delete(pod, false) // not notifying the watchers
failPodFromExecutor(launchedTask.taskInfo)
podKey, _ := podtask.MakePodKey(api.NewDefaultContext(), pod.Name)
assertext.EventuallyTrue(t, time.Second, func() bool {
t, _ := p.api.tasks().ForPod(podKey)
return t == nil
})
// 2. with pod still on the apiserver, not bound
// expected: pod is rescheduled
pod, launchedTask, _ = launchTestPod()
failPodFromExecutor(launchedTask.taskInfo)
retryOffers := []*mesos.Offer{NewTestOffer("retry-offer")}
schedulePodWithOffers(pod, retryOffers)
// 3. with pod still on the apiserver, bound, notified via ListWatch
// expected: nothing, pod updates not supported, compare ReconcileTask function
pod, launchedTask, usedOffer = startTestPod()
pod.Annotations = map[string]string{
meta.BindingHostKey: *usedOffer.Hostname,
}
pod.Spec.NodeName = *usedOffer.Hostname
podListWatch.Modify(pod, true) // notifying the watchers
time.Sleep(time.Second / 2)
failPodFromExecutor(launchedTask.taskInfo)
}