Work around the container name conflict bug

pkg/kubelet/dockershim/docker_container.go

@@ -176,7 +176,9 @@ func (ds *dockerService) CreateContainer(podSandboxID string, config *runtimeapi
 
 	createConfig.HostConfig = hc
 	createResp, err := ds.client.CreateContainer(createConfig)
-	recoverFromConflictIfNeeded(ds.client, err)
+	if err != nil {
+		createResp, err = recoverFromCreationConflictIfNeeded(ds.client, createConfig, err)
+	}
 
 	if createResp != nil {
 		return createResp.ID, err

pkg/kubelet/dockershim/docker_sandbox.go

@@ -69,7 +69,9 @@ func (ds *dockerService) RunPodSandbox(config *runtimeapi.PodSandboxConfig) (str
 		return "", fmt.Errorf("failed to make sandbox docker config for pod %q: %v", config.Metadata.Name, err)
 	}
 	createResp, err := ds.client.CreateContainer(*createConfig)
-	recoverFromConflictIfNeeded(ds.client, err)
+	if err != nil {
+		createResp, err = recoverFromCreationConflictIfNeeded(ds.client, *createConfig, err)
+	}
 
 	if err != nil || createResp == nil {
 		return "", fmt.Errorf("failed to create a sandbox for pod %q: %v", config.Metadata.Name, err)

pkg/kubelet/dockershim/helpers.go

@@ -39,7 +39,8 @@ const (
 )
 
 var (
-	conflictRE = regexp.MustCompile(`Conflict. (?:.)+ is already in use by container ([0-9a-z]+)`)
+	conflictRE    = regexp.MustCompile(`Conflict. (?:.)+ is already in use by container ([0-9a-z]+)`)
+	noContainerRE = regexp.MustCompile(`No such container: [0-9a-z]+`)
 )
 
 // apiVersion implements kubecontainer.Version interface by implementing
@@ -295,22 +296,31 @@ func getUserFromImageUser(imageUser string) (*int64, string) {
 // create a new container named FOO. To work around this, we parse the error
 // message to identify failure caused by naming conflict, and try to remove
 // the old container FOO.
+// See #40443. Sometimes even removal may fail with "no such container" error.
+// In that case we have to create the container with a randomized name.
+// TODO(random-liu): Remove this work around after docker 1.11 is deprecated.
 // TODO(#33189): Monitor the tests to see if the fix is sufficent.
-func recoverFromConflictIfNeeded(client dockertools.DockerInterface, err error) {
-	if err == nil {
-		return
-	}
-
+func recoverFromCreationConflictIfNeeded(client dockertools.DockerInterface, createConfig dockertypes.ContainerCreateConfig, err error) (*dockertypes.ContainerCreateResponse, error) {
 	matches := conflictRE.FindStringSubmatch(err.Error())
 	if len(matches) != 2 {
-		return
+		return nil, err
 	}
 
 	id := matches[1]
 	glog.Warningf("Unable to create pod sandbox due to conflict. Attempting to remove sandbox %q", id)
-	if err := client.RemoveContainer(id, dockertypes.ContainerRemoveOptions{RemoveVolumes: true}); err != nil {
-		glog.Errorf("Failed to remove the conflicting sandbox container: %v", err)
+	if rmErr := client.RemoveContainer(id, dockertypes.ContainerRemoveOptions{RemoveVolumes: true}); rmErr == nil {
+		glog.V(2).Infof("Successfully removed conflicting container %q", id)
+		return nil, err
 	} else {
-		glog.V(2).Infof("Successfully removed conflicting sandbox %q", id)
+		glog.Errorf("Failed to remove the conflicting container %q: %v", id, rmErr)
+		// Return if the error is not "No such container" error.
+		if !noContainerRE.MatchString(rmErr.Error()) {
+			return nil, err
+		}
 	}
+
+	// randomize the name to avoid conflict.
+	createConfig.Name = randomizeName(createConfig.Name)
+	glog.V(2).Infof("Create the container with randomized name %s", createConfig.Name)
+	return client.CreateContainer(createConfig)
 }

pkg/kubelet/dockershim/naming.go

@@ -18,6 +18,7 @@ package dockershim
 
 import (
 	"fmt"
+	"math/rand"
 	"strconv"
 	"strings"
 
@@ -78,6 +79,15 @@ func makeContainerName(s *runtimeapi.PodSandboxConfig, c *runtimeapi.ContainerCo
 
 }
 
+// randomizeName randomizes the container name. This should only be used when we hit the
+// docker container name conflict bug.
+func randomizeName(name string) string {
+	return strings.Join([]string{
+		name,
+		fmt.Sprintf("%08x", rand.Uint32()),
+	}, nameDelimiter)
+}
+
 func parseUint32(s string) (uint32, error) {
 	n, err := strconv.ParseUint(s, 10, 32)
 	if err != nil {
@@ -92,7 +102,9 @@ func parseSandboxName(name string) (*runtimeapi.PodSandboxMetadata, error) {
 	name = strings.TrimPrefix(name, "/")
 
 	parts := strings.Split(name, nameDelimiter)
-	if len(parts) != 6 {
+	// Tolerate the random suffix.
+	// TODO(random-liu): Remove 7 field case when docker 1.11 is deprecated.
+	if len(parts) != 6 && len(parts) != 7 {
 		return nil, fmt.Errorf("failed to parse the sandbox name: %q", name)
 	}
 	if parts[0] != kubePrefix {
@@ -118,7 +130,9 @@ func parseContainerName(name string) (*runtimeapi.ContainerMetadata, error) {
 	name = strings.TrimPrefix(name, "/")
 
 	parts := strings.Split(name, nameDelimiter)
-	if len(parts) != 6 {
+	// Tolerate the random suffix.
+	// TODO(random-liu): Remove 7 field case when docker 1.11 is deprecated.
+	if len(parts) != 6 && len(parts) != 7 {
 		return nil, fmt.Errorf("failed to parse the container name: %q", name)
 	}
 	if parts[0] != kubePrefix {