Re-run init containers if the pod sandbox needs to be recreated

pkg/kubelet/apis/cri/testing/fake_runtime_service.go

@@ -55,6 +55,18 @@ type FakeRuntimeService struct {
 	Sandboxes  map[string]*FakePodSandbox
 }
 
+func (r *FakeRuntimeService) GetContainerID(sandboxID, name string, attempt uint32) (string, error) {
+	r.Lock()
+	defer r.Unlock()
+
+	for id, c := range r.Containers {
+		if c.SandboxID == sandboxID && c.Metadata.Name == name && c.Metadata.Attempt == attempt {
+			return id, nil
+		}
+	}
+	return "", fmt.Errorf("container (name, attempt, sandboxID)=(%q, %d, %q) not found", name, attempt, sandboxID)
+}
+
 func (r *FakeRuntimeService) SetFakeSandboxes(sandboxes []*FakePodSandbox) {
 	r.Lock()
 	defer r.Unlock()

pkg/kubelet/kuberuntime/kuberuntime_container.go

@@ -631,10 +631,11 @@ func (m *kubeGenericRuntimeManager) killContainersWithSyncResult(pod *v1.Pod, ru
 	return
 }
 
-// pruneInitContainers ensures that before we begin creating init containers, we have reduced the number
-// of outstanding init containers still present. This reduces load on the container garbage collector
-// by only preserving the most recent terminated init container.
-func (m *kubeGenericRuntimeManager) pruneInitContainersBeforeStart(pod *v1.Pod, podStatus *kubecontainer.PodStatus, initContainersToKeep map[kubecontainer.ContainerID]int) {
+// pruneInitContainersBeforeStart ensures that before we begin creating init
+// containers, we have reduced the number of outstanding init containers still
+// present. This reduces load on the container garbage collector by only
+// preserving the most recent terminated init container.
+func (m *kubeGenericRuntimeManager) pruneInitContainersBeforeStart(pod *v1.Pod, podStatus *kubecontainer.PodStatus) {
 	// only the last execution of each init container should be preserved, and only preserve it if it is in the
 	// list of init containers to keep.
 	initContainerNames := sets.NewString()
@@ -652,14 +653,9 @@ func (m *kubeGenericRuntimeManager) pruneInitContainersBeforeStart(pod *v1.Pod,
 			if count == 1 {
 				continue
 			}
-			// if there is a reason to preserve the older container, do so
-			if _, ok := initContainersToKeep[status.ID]; ok {
-				continue
-			}
-
 			// prune all other init containers that match this container name
 			glog.V(4).Infof("Removing init container %q instance %q %d", status.Name, status.ID.ID, count)
-			if err := m.runtimeService.RemoveContainer(status.ID.ID); err != nil {
+			if err := m.removeContainer(status.ID.ID); err != nil {
 				utilruntime.HandleError(fmt.Errorf("failed to remove pod init container %q: %v; Skipping pod %q", status.Name, err, format.Pod(pod)))
 				continue
 			}
@@ -674,6 +670,37 @@ func (m *kubeGenericRuntimeManager) pruneInitContainersBeforeStart(pod *v1.Pod,
 	}
 }
 
+// Remove all init containres. Note that this function does not check the state
+// of the container because it assumes all init containers have been stopped
+// before the call happens.
+func (m *kubeGenericRuntimeManager) purgeInitContainers(pod *v1.Pod, podStatus *kubecontainer.PodStatus) {
+	initContainerNames := sets.NewString()
+	for _, container := range pod.Spec.InitContainers {
+		initContainerNames.Insert(container.Name)
+	}
+	for name := range initContainerNames {
+		count := 0
+		for _, status := range podStatus.ContainerStatuses {
+			if status.Name != name || !initContainerNames.Has(status.Name) {
+				continue
+			}
+			count++
+			// Purge all init containers that match this container name
+			glog.V(4).Infof("Removing init container %q instance %q %d", status.Name, status.ID.ID, count)
+			if err := m.removeContainer(status.ID.ID); err != nil {
+				utilruntime.HandleError(fmt.Errorf("failed to remove pod init container %q: %v; Skipping pod %q", status.Name, err, format.Pod(pod)))
+				continue
+			}
+			// Remove any references to this container
+			if _, ok := m.containerRefManager.GetRef(status.ID); ok {
+				m.containerRefManager.ClearRef(status.ID)
+			} else {
+				glog.Warningf("No ref for container %q", status.ID)
+			}
+		}
+	}
+}
+
 // findNextInitContainerToRun returns the status of the last failed container, the
 // next init container to start, or done if there are no further init containers.
 // Status is only returned if an init container is failed, in which case next will