// Generates skydns records for a headless service.
func (ks *qing2sky) newHeadlessService(subdomain string, service *kapi.Service, isNewStyleFormat bool) error {
// Create an A record for every pod in the service.
// This record must be periodically updated.
// Format is as follows:
// For a service x, with pods a and b create DNS records,
// a.x.ns.domain. and, b.x.ns.domain.
// TODO: Handle multi-port services.
ks.mlock.Lock()
defer ks.mlock.Unlock()
key, err := kcache.MetaNamespaceKeyFunc(service)
if err != nil {
return err
}
e, exists, err := ks.endpointsStore.GetByKey(key)
if err != nil {
return fmt.Errorf("failed to get endpoints object from endpoints store - %v", err)
}
if !exists {
glog.V(1).Infof("could not find endpoints for service %q in namespace %q. DNS records will be created once endpoints show up.", service.Name, service.Namespace)
return nil
}
if e, ok := e.(*kapi.Endpoints); ok {
return ks.generateRecordsForHeadlessService(subdomain, e, service, isNewStyleFormat)
}
return nil
}
// implements Unique
func (p *Pod) GetUID() string {
if id, err := cache.MetaNamespaceKeyFunc(p.Pod); err != nil {
panic(fmt.Sprintf("failed to determine pod id for '%+v'", p.Pod))
} else {
return id
}
}
// Schedule implements the Scheduler interface of QingYuan.
// It returns the selectedMachine's name and error (if there's any).
func (k *qingScheduler) Schedule(pod *api.Pod, unused algorithm.MinionLister) (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)
}
}
func (ks *qing2sky) getServiceFromEndpoints(e *kapi.Endpoints) (*kapi.Service, error) {
key, err := kcache.MetaNamespaceKeyFunc(e)
if err != nil {
return nil, err
}
obj, exists, err := ks.servicesStore.GetByKey(key)
if err != nil {
return nil, fmt.Errorf("failed to get service object from services store - %v", err)
}
if !exists {
glog.V(1).Infof("could not find service for endpoint %q in namespace %q", e.Name, e.Namespace)
return nil, nil
}
if svc, ok := obj.(*kapi.Service); ok {
return svc, nil
}
return nil, fmt.Errorf("got a non service object in services store %v", obj)
}
// implementation of scheduling plugin's NextPod func; see k8s plugin/pkg/scheduler
func (q *queuer) yield() *api.Pod {
log.V(2).Info("attempting to yield a pod")
q.lock.Lock()
defer q.lock.Unlock()
for {
// limit blocking here to short intervals so that we don't block the
// enqueuer Run() routine for very long
kpod := q.podQueue.Await(yieldPopTimeout)
if kpod == nil {
signalled := runtime.After(q.unscheduledCond.Wait)
// lock is yielded at this point and we're going to wait for either
// a timeout, or a signal that there's data
select {
case <-time.After(yieldWaitTimeout):
q.unscheduledCond.Broadcast() // abort Wait()
<-signalled // wait for the go-routine, and the lock
log.V(4).Infoln("timed out waiting for a pod to yield")
case <-signalled:
// we have acquired the lock, and there
// may be a pod for us to pop now
}
continue
}
pod := kpod.(*Pod).Pod
if podName, err := cache.MetaNamespaceKeyFunc(pod); err != nil {
log.Warningf("yield unable to understand pod object %+v, will skip: %v", pod, err)
} else if !q.podUpdates.Poll(podName, queue.POP_EVENT) {
log.V(1).Infof("yield popped a transitioning pod, skipping: %+v", pod)
} else if pod.Spec.NodeName != "" {
// should never happen if enqueuePods is filtering properly
log.Warningf("yield popped an already-scheduled pod, skipping: %+v", pod)
} else {
return pod
}
}
}
func DoTestUnschedulableNodes(t *testing.T, restClient *client.Client, nodeStore cache.Store) {
goodCondition := api.NodeCondition{
Type: api.NodeReady,
Status: api.ConditionTrue,
Reason: fmt.Sprintf("schedulable condition"),
LastHeartbeatTime: util.Time{time.Now()},
}
badCondition := api.NodeCondition{
Type: api.NodeReady,
Status: api.ConditionUnknown,
Reason: fmt.Sprintf("unschedulable condition"),
LastHeartbeatTime: util.Time{time.Now()},
}
// Create a new schedulable node, since we're first going to apply
// the unschedulable condition and verify that pods aren't scheduled.
node := &api.Node{
ObjectMeta: api.ObjectMeta{Name: "node-scheduling-test-node"},
Spec: api.NodeSpec{Unschedulable: false},
Status: api.NodeStatus{
Capacity: api.ResourceList{
api.ResourcePods: *resource.NewQuantity(32, resource.DecimalSI),
},
Conditions: []api.NodeCondition{goodCondition},
},
}
nodeKey, err := cache.MetaNamespaceKeyFunc(node)
if err != nil {
t.Fatalf("Couldn't retrieve key for node %v", node.Name)
}
// The test does the following for each nodeStateManager in this list:
// 1. Create a new node
// 2. Apply the makeUnSchedulable function
// 3. Create a new pod
// 4. Check that the pod doesn't get assigned to the node
// 5. Apply the schedulable function
// 6. Check that the pod *does* get assigned to the node
// 7. Delete the pod and node.
nodeModifications := []nodeStateManager{
// Test node.Spec.Unschedulable=true/false
{
makeUnSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
n.Spec.Unschedulable = true
if _, err := c.Nodes().Update(n); err != nil {
t.Fatalf("Failed to update node with unschedulable=true: %v", err)
}
err = waitForReflection(s, nodeKey, func(node interface{}) bool {
// An unschedulable node should get deleted from the store
return node == nil
})
if err != nil {
t.Fatalf("Failed to observe reflected update for setting unschedulable=true: %v", err)
}
},
makeSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
n.Spec.Unschedulable = false
if _, err := c.Nodes().Update(n); err != nil {
t.Fatalf("Failed to update node with unschedulable=false: %v", err)
}
err = waitForReflection(s, nodeKey, func(node interface{}) bool {
return node != nil && node.(*api.Node).Spec.Unschedulable == false
})
if err != nil {
t.Fatalf("Failed to observe reflected update for setting unschedulable=false: %v", err)
}
},
},
// Test node.Status.Conditions=ConditionTrue/Unknown
{
makeUnSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
n.Status = api.NodeStatus{
Capacity: api.ResourceList{
api.ResourcePods: *resource.NewQuantity(32, resource.DecimalSI),
},
Conditions: []api.NodeCondition{badCondition},
}
if _, err = c.Nodes().UpdateStatus(n); err != nil {
t.Fatalf("Failed to update node with bad status condition: %v", err)
}
err = waitForReflection(s, nodeKey, func(node interface{}) bool {
return node != nil && node.(*api.Node).Status.Conditions[0].Status == api.ConditionUnknown
})
if err != nil {
t.Fatalf("Failed to observe reflected update for status condition update: %v", err)
}
},
makeSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
n.Status = api.NodeStatus{
Capacity: api.ResourceList{
api.ResourcePods: *resource.NewQuantity(32, resource.DecimalSI),
},
Conditions: []api.NodeCondition{goodCondition},
}
if _, err = c.Nodes().UpdateStatus(n); err != nil {
t.Fatalf("Failed to update node with healthy status condition: %v", err)
}
waitForReflection(s, nodeKey, func(node interface{}) bool {
return node != nil && node.(*api.Node).Status.Conditions[0].Status == api.ConditionTrue
})
if err != nil {
t.Fatalf("Failed to observe reflected update for status condition update: %v", err)
}
},
},
}
for i, mod := range nodeModifications {
unSchedNode, err := restClient.Nodes().Create(node)
if err != nil {
t.Fatalf("Failed to create node: %v", err)
}
// Apply the unschedulable modification to the node, and wait for the reflection
mod.makeUnSchedulable(t, unSchedNode, nodeStore, restClient)
// Create the new pod, note that this needs to happen post unschedulable
// modification or we have a race in the test.
pod := &api.Pod{
ObjectMeta: api.ObjectMeta{Name: "node-scheduling-test-pod"},
Spec: api.PodSpec{
Containers: []api.Container{{Name: "container", Image: "qingyuan/pause:go"}},
},
}
myPod, err := restClient.Pods(api.NamespaceDefault).Create(pod)
if err != nil {
t.Fatalf("Failed to create pod: %v", err)
}
// There are no schedulable nodes - the pod shouldn't be scheduled.
err = wait.Poll(time.Second, time.Second*10, podScheduled(restClient, myPod.Namespace, myPod.Name))
if err == nil {
t.Errorf("Pod scheduled successfully on unschedulable nodes")
}
if err != wait.ErrWaitTimeout {
t.Errorf("Test %d: failed while trying to confirm the pod does not get scheduled on the node: %v", i, err)
} else {
t.Logf("Test %d: Pod did not get scheduled on an unschedulable node", i)
}
// Apply the schedulable modification to the node, and wait for the reflection
schedNode, err := restClient.Nodes().Get(unSchedNode.Name)
if err != nil {
t.Fatalf("Failed to get node: %v", err)
}
mod.makeSchedulable(t, schedNode, nodeStore, restClient)
// Wait until the pod is scheduled.
err = wait.Poll(time.Second, time.Second*10, podScheduled(restClient, myPod.Namespace, myPod.Name))
if err != nil {
t.Errorf("Test %d: failed to schedule a pod: %v", i, err)
} else {
t.Logf("Test %d: Pod got scheduled on a schedulable node", i)
}
err = restClient.Pods(api.NamespaceDefault).Delete(myPod.Name, nil)
if err != nil {
t.Errorf("Failed to delete pod: %v", err)
}
err = restClient.Nodes().Delete(schedNode.Name)
if err != nil {
t.Errorf("Failed to delete node: %v", err)
}
}
}
// DeletionHandlingMetaNamespaceKeyFunc checks for
// cache.DeletedFinalStateUnknown objects before calling
// cache.MetaNamespaceKeyFunc.
func DeletionHandlingMetaNamespaceKeyFunc(obj interface{}) (string, error) {
if d, ok := obj.(cache.DeletedFinalStateUnknown); ok {
return d.Key, nil
}
return cache.MetaNamespaceKeyFunc(obj)
}