Alleviate race between scheduler and kubelet

If the scheduler learns of a node change and schedules based on it
before the kubelet learns of the node change, the pod will fail.

pkg/kubelet/lifecycle/predicate.go

@@ -18,6 +18,7 @@ package lifecycle
 
 import (
 	"fmt"
+	"time"
 
 	"k8s.io/klog"
 
@@ -55,6 +56,19 @@ func NewPredicateAdmitHandler(getNodeAnyWayFunc getNodeAnyWayFuncType, admission
 }
 
 func (w *predicateAdmitHandler) Admit(attrs *PodAdmitAttributes) PodAdmitResult {
+	result := w.admit(attrs)
+	for i := 0; i < 10; i++ {
+		if result.Admit {
+			break
+		}
+		time.Sleep(time.Second)
+		result = w.admit(attrs)
+	}
+
+	return result
+}
+
+func (w *predicateAdmitHandler) admit(attrs *PodAdmitAttributes) PodAdmitResult {
 	node, err := w.getNodeAnyWayFunc()
 	if err != nil {
 		klog.Errorf("Cannot get Node info: %v", err)