kubelet: fix /stats/summary endpoint on Windows when init-container…

…s are present on the node

Following changes in kubernetes#87730, Kubelet is directly hcsshim to gather stats.
However, unlike `docker stats` API that was used before, hcsshim does not
keep information about exited containers.

When the Kubelet lists containers (`docker_container.go:ListContainers()`),
it sets `All: true`, retrieving non-running containers.

When docker stats is called with such container id, it'll return a valid JSON
with all values set to 0. The non-running containers are filtered later on in the process.

When the hcsshim is called with such container id, it'll return an error, effectively
stopping the stats retrieval for all containers.

pkg/kubelet/dockershim/docker_stats.go

@@ -53,8 +53,9 @@ func (ds *dockerService) ListContainerStats(ctx context.Context, r *runtimeapi.L
 		if err != nil {
 			return nil, err
 		}
-
-		stats = append(stats, containerStats)
+		if containerStats != nil {
+			stats = append(stats, containerStats)
+		}
 	}
 
 	return &runtimeapi.ListContainerStatsResponse{Stats: stats}, nil

pkg/kubelet/dockershim/docker_stats_windows.go

@@ -35,7 +35,13 @@ func (ds *dockerService) getContainerStats(containerID string) (*runtimeapi.Cont
 
 	hcsshim_container, err := hcsshim.OpenContainer(containerID)
 	if err != nil {
-		return nil, err
+		// As we moved from using Docker stats to hcsshim directly, we may query HCS with already exited container IDs.
+		// That will typically happen with init-containers in Exited state. Docker still knows about them but the HCS does not.
+		// As we don't want to block stats retrieval for other containers, we only log errors.
+		if !hcsshim.IsNotExist(err) && !hcsshim.IsAlreadyStopped(err) {
+			klog.Errorf("Error opening container (stats will be missing) '%s': %v", containerID, err)
+		}
+		return nil, nil
 	}
 	defer func() {
 		closeErr := hcsshim_container.Close()

test/e2e/windows/kubelet_stats.go

@@ -144,7 +144,17 @@ func newKubeletStatsTestPods(numPods int, image imageutils.Config, nodeName stri
 						},
 					},
 				},
-
+				InitContainers: []v1.Container{
+					{
+						Image: image.GetE2EImage(),
+						Name:  podName,
+						Command: []string{
+							"powershell.exe",
+							"-Command",
+							"sleep -Seconds 1",
+						},
+					},
+				},
 				NodeName: nodeName,
 			},
 		}