Merge pull request #23567 from smarterclayton/init_containers

Automatic merge from submit-queue

Add init containers to pods

This implements #1589 as per proposal #23666

Incorporates feedback on #1589, creates parallel structure for InitContainers and Containers, adds validation for InitContainers that requires name uniqueness, and comments on a number of implications of init containers.

This is a complete alpha implementation.

<!-- Reviewable:start -->
---
This change is [<img src="http://reviewable.k8s.io/review_button.svg" height="35" align="absmiddle" alt="Reviewable"/>](http://reviewable.k8s.io/reviews/kubernetes/kubernetes/23567)
<!-- Reviewable:end -->

CHANGELOG.md

@@ -11,6 +11,7 @@
     - [Other notable changes](#other-notable-changes)
 - [v1.3.0-alpha.3](#v130-alpha3)
   - [Downloads](#downloads)
+- [v1.3.0-alpha.3](#v130-alpha3)
   - [Changes since v1.3.0-alpha.2](#changes-since-v130-alpha2)
     - [Action Required](#action-required)
     - [Other notable changes](#other-notable-changes)

api/swagger-spec/apps_v1alpha1.json

@@ -1854,7 +1854,7 @@
       "items": {
        "$ref": "v1.VolumeMount"
       },
-      "description": "Pod volumes to mount into the container's filesyste. Cannot be updated."
+      "description": "Pod volumes to mount into the container's filesystem. Cannot be updated."
      },
      "livenessProbe": {
       "$ref": "v1.Probe",

api/swagger-spec/batch_v1.json

@@ -1859,7 +1859,7 @@
       "items": {
        "$ref": "v1.VolumeMount"
       },
-      "description": "Pod volumes to mount into the container's filesyste. Cannot be updated."
+      "description": "Pod volumes to mount into the container's filesystem. Cannot be updated."
      },
      "livenessProbe": {
       "$ref": "v1.Probe",

api/swagger-spec/extensions_v1beta1.json

@@ -7166,7 +7166,7 @@
       "items": {
        "$ref": "v1.VolumeMount"
       },
-      "description": "Pod volumes to mount into the container's filesyste. Cannot be updated."
+      "description": "Pod volumes to mount into the container's filesystem. Cannot be updated."
      },
      "livenessProbe": {
       "$ref": "v1.Probe",

api/swagger-spec/v1.json

@@ -17177,7 +17177,7 @@
       "items": {
        "$ref": "v1.VolumeMount"
       },
-      "description": "Pod volumes to mount into the container's filesyste. Cannot be updated."
+      "description": "Pod volumes to mount into the container's filesystem. Cannot be updated."
      },
      "livenessProbe": {
       "$ref": "v1.Probe",

docs/api-reference/batch/v1/definitions.html

@@ -1375,7 +1375,7 @@ Examples:<br>
 </tr>
 <tr>
 <td class="tableblock halign-left valign-top"><p class="tableblock">volumeMounts</p></td>
-<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesyste. Cannot be updated.</p></td>
+<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesystem. Cannot be updated.</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock">false</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock"><a href="#_v1_volumemount">v1.VolumeMount</a> array</p></td>
 <td class="tableblock halign-left valign-top"></td>
@@ -3979,7 +3979,7 @@ Populated by the system when a graceful deletion is requested. Read-only. More i
 </div>
 <div id="footer">
 <div id="footer-text">
-Last updated 2016-05-06 04:28:34 UTC
+Last updated 2016-05-06 14:18:34 UTC
 </div>
 </div>
 </body>

docs/api-reference/extensions/v1beta1/definitions.html

@@ -4706,7 +4706,7 @@ Both these may change in the future. Incoming requests are matched against the h
 </tr>
 <tr>
 <td class="tableblock halign-left valign-top"><p class="tableblock">volumeMounts</p></td>
-<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesyste. Cannot be updated.</p></td>
+<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesystem. Cannot be updated.</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock">false</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock"><a href="#_v1_volumemount">v1.VolumeMount</a> array</p></td>
 <td class="tableblock halign-left valign-top"></td>

docs/api-reference/v1/definitions.html

@@ -5440,7 +5440,7 @@ The resulting set of endpoints can be viewed as:<br>
 </tr>
 <tr>
 <td class="tableblock halign-left valign-top"><p class="tableblock">volumeMounts</p></td>
-<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesyste. Cannot be updated.</p></td>
+<td class="tableblock halign-left valign-top"><p class="tableblock">Pod volumes to mount into the container&#8217;s filesystem. Cannot be updated.</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock">false</p></td>
 <td class="tableblock halign-left valign-top"><p class="tableblock"><a href="#_v1_volumemount">v1.VolumeMount</a> array</p></td>
 <td class="tableblock halign-left valign-top"></td>

pkg/api/deep_copy_generated.go

@@ -2232,6 +2232,17 @@ func DeepCopy_api_PodSpec(in PodSpec, out *PodSpec, c *conversion.Cloner) error
 	} else {
 		out.Volumes = nil
 	}
+	if in.InitContainers != nil {
+		in, out := in.InitContainers, &out.InitContainers
+		*out = make([]Container, len(in))
+		for i := range in {
+			if err := DeepCopy_api_Container(in[i], &(*out)[i], c); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainers = nil
+	}
 	if in.Containers != nil {
 		in, out := in.Containers, &out.Containers
 		*out = make([]Container, len(in))
@@ -2321,6 +2332,17 @@ func DeepCopy_api_PodStatus(in PodStatus, out *PodStatus, c *conversion.Cloner)
 	} else {
 		out.StartTime = nil
 	}
+	if in.InitContainerStatuses != nil {
+		in, out := in.InitContainerStatuses, &out.InitContainerStatuses
+		*out = make([]ContainerStatus, len(in))
+		for i := range in {
+			if err := DeepCopy_api_ContainerStatus(in[i], &(*out)[i], c); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainerStatuses = nil
+	}
 	if in.ContainerStatuses != nil {
 		in, out := in.ContainerStatuses, &out.ContainerStatuses
 		*out = make([]ContainerStatus, len(in))

pkg/api/resource_helpers.go

@@ -92,6 +92,8 @@ func GetPodReadyCondition(status PodStatus) *PodCondition {
 	return condition
 }
 
+// GetPodCondition extracts the provided condition from the given status and returns that.
+// Returns nil and -1 if the condition is not present, and the the index of the located condition.
 func GetPodCondition(status *PodStatus, conditionType PodConditionType) (int, *PodCondition) {
 	for i, c := range status.Conditions {
 		if c.Type == conditionType {
@@ -161,5 +163,28 @@ func PodRequestsAndLimits(pod *Pod) (reqs map[ResourceName]resource.Quantity, li
 			}
 		}
 	}
+	// init containers define the minimum of any resource
+	for _, container := range pod.Spec.InitContainers {
+		for name, quantity := range container.Resources.Requests {
+			value, ok := reqs[name]
+			if !ok {
+				reqs[name] = *quantity.Copy()
+				continue
+			}
+			if quantity.Cmp(value) > 0 {
+				reqs[name] = *quantity.Copy()
+			}
+		}
+		for name, quantity := range container.Resources.Limits {
+			value, ok := limits[name]
+			if !ok {
+				limits[name] = *quantity.Copy()
+				continue
+			}
+			if quantity.Cmp(value) > 0 {
+				limits[name] = *quantity.Copy()
+			}
+		}
+	}
 	return
 }

pkg/api/serialization_proto_test.go

@@ -67,6 +67,9 @@ func TestUniversalDeserializer(t *testing.T) {
 func TestProtobufRoundTrip(t *testing.T) {
 	obj := &v1.Pod{}
 	apitesting.FuzzerFor(t, v1.SchemeGroupVersion, rand.NewSource(benchmarkSeed)).Fuzz(obj)
+	// InitContainers are turned into annotations by conversion.
+	obj.Spec.InitContainers = nil
+	obj.Status.InitContainerStatuses = nil
 	data, err := obj.Marshal()
 	if err != nil {
 		t.Fatal(err)
@@ -77,7 +80,7 @@ func TestProtobufRoundTrip(t *testing.T) {
 	}
 	if !api.Semantic.Equalities.DeepEqual(out, obj) {
 		t.Logf("marshal\n%s", hex.Dump(data))
-		t.Fatalf("Unmarshal is unequal\n%s", diff.ObjectGoPrintSideBySide(out, obj))
+		t.Fatalf("Unmarshal is unequal\n%s", diff.ObjectGoPrintDiff(out, obj))
 	}
 }
 

pkg/api/types.generated.go

@@ -52540,7 +52540,7 @@ func (x codecSelfer1234) decSlicePod(v *[]Pod, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 576)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 624)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -53849,7 +53849,7 @@ func (x codecSelfer1234) decSlicePodTemplate(v *[]PodTemplate, d *codec1978.Deco
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 648)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 672)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/api/types.go

@@ -1076,6 +1076,8 @@ const (
 	// PodReady means the pod is able to service requests and should be added to the
 	// load balancing pools of all matching services.
 	PodReady PodConditionType = "Ready"
+	// PodInitialized means that all init containers in the pod have started successfully.
+	PodInitialized PodConditionType = "Initialized"
 )
 
 type PodCondition struct {
@@ -1309,7 +1311,9 @@ type PreferredSchedulingTerm struct {
 // PodSpec is a description of a pod
 type PodSpec struct {
 	Volumes []Volume `json:"volumes"`
-	// Required: there must be at least one container in a pod.
+	// List of initialization containers belonging to the pod.
+	InitContainers []Container `json:"-"`
+	// List of containers belonging to the pod.
 	Containers    []Container   `json:"containers"`
 	RestartPolicy RestartPolicy `json:"restartPolicy,omitempty"`
 	// Optional duration in seconds the pod needs to terminate gracefully. May be decreased in delete request.
@@ -1416,6 +1420,11 @@ type PodStatus struct {
 	// This is before the Kubelet pulled the container image(s) for the pod.
 	StartTime *unversioned.Time `json:"startTime,omitempty"`
 
+	// The list has one entry per init container in the manifest. The most recent successful
+	// init container will have ready = true, the most recently started container will have
+	// startTime set.
+	// More info: http://releases.k8s.io/HEAD/docs/user-guide/pod-states.md#container-statuses
+	InitContainerStatuses []ContainerStatus `json:"-"`
 	// The list has one entry per container in the manifest. Each entry is
 	// currently the output of `docker inspect`. This output format is *not*
 	// final and should not be relied upon.

pkg/api/v1/conversion.go

@@ -17,6 +17,7 @@ limitations under the License.
 package v1
 
 import (
+	"encoding/json"
 	"fmt"
 
 	inf "gopkg.in/inf.v0"
@@ -258,6 +259,75 @@ func Convert_v1_ReplicationControllerSpec_To_api_ReplicationControllerSpec(in *R
 	return nil
 }
 
+func Convert_api_PodStatusResult_To_v1_PodStatusResult(in *api.PodStatusResult, out *PodStatusResult, s conversion.Scope) error {
+	if err := autoConvert_api_PodStatusResult_To_v1_PodStatusResult(in, out, s); err != nil {
+		return err
+	}
+
+	if len(out.Status.InitContainerStatuses) > 0 {
+		if out.Annotations == nil {
+			out.Annotations = make(map[string]string)
+		}
+		value, err := json.Marshal(out.Status.InitContainerStatuses)
+		if err != nil {
+			return err
+		}
+		out.Annotations[PodInitContainerStatusesAnnotationKey] = string(value)
+	} else {
+		delete(in.Annotations, PodInitContainerStatusesAnnotationKey)
+	}
+	return nil
+}
+
+func Convert_v1_PodStatusResult_To_api_PodStatusResult(in *PodStatusResult, out *api.PodStatusResult, s conversion.Scope) error {
+	// TODO: when we move init container to beta, remove these conversions
+	if value := in.Annotations[PodInitContainerStatusesAnnotationKey]; len(value) > 0 {
+		delete(in.Annotations, PodInitContainerStatusesAnnotationKey)
+		var values []ContainerStatus
+		if err := json.Unmarshal([]byte(value), &values); err != nil {
+			return err
+		}
+		in.Status.InitContainerStatuses = values
+	}
+
+	return autoConvert_v1_PodStatusResult_To_api_PodStatusResult(in, out, s)
+}
+
+func Convert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in *api.PodTemplateSpec, out *PodTemplateSpec, s conversion.Scope) error {
+	if err := autoConvert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in, out, s); err != nil {
+		return err
+	}
+
+	// TODO: when we move init container to beta, remove these conversions
+	if len(out.Spec.InitContainers) > 0 {
+		if out.Annotations == nil {
+			out.Annotations = make(map[string]string)
+		}
+		value, err := json.Marshal(out.Spec.InitContainers)
+		if err != nil {
+			return err
+		}
+		out.Annotations[PodInitContainersAnnotationKey] = string(value)
+	} else {
+		delete(out.Annotations, PodInitContainersAnnotationKey)
+	}
+	return nil
+}
+
+func Convert_v1_PodTemplateSpec_To_api_PodTemplateSpec(in *PodTemplateSpec, out *api.PodTemplateSpec, s conversion.Scope) error {
+	// TODO: when we move init container to beta, remove these conversions
+	if value := in.Annotations[PodInitContainersAnnotationKey]; len(value) > 0 {
+		delete(in.Annotations, PodInitContainersAnnotationKey)
+		var values []Container
+		if err := json.Unmarshal([]byte(value), &values); err != nil {
+			return err
+		}
+		in.Spec.InitContainers = values
+	}
+
+	return autoConvert_v1_PodTemplateSpec_To_api_PodTemplateSpec(in, out, s)
+}
+
 // The following two PodSpec conversions are done here to support ServiceAccount
 // as an alias for ServiceAccountName.
 func Convert_api_PodSpec_To_v1_PodSpec(in *api.PodSpec, out *PodSpec, s conversion.Scope) error {
@@ -271,6 +341,16 @@ func Convert_api_PodSpec_To_v1_PodSpec(in *api.PodSpec, out *PodSpec, s conversi
 	} else {
 		out.Volumes = nil
 	}
+	if in.InitContainers != nil {
+		out.InitContainers = make([]Container, len(in.InitContainers))
+		for i := range in.InitContainers {
+			if err := Convert_api_Container_To_v1_Container(&in.InitContainers[i], &out.InitContainers[i], s); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainers = nil
+	}
 	if in.Containers != nil {
 		out.Containers = make([]Container, len(in.Containers))
 		for i := range in.Containers {
@@ -346,6 +426,16 @@ func Convert_v1_PodSpec_To_api_PodSpec(in *PodSpec, out *api.PodSpec, s conversi
 	} else {
 		out.Volumes = nil
 	}
+	if in.InitContainers != nil {
+		out.InitContainers = make([]api.Container, len(in.InitContainers))
+		for i := range in.InitContainers {
+			if err := Convert_v1_Container_To_api_Container(&in.InitContainers[i], &out.InitContainers[i], s); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainers = nil
+	}
 	if in.Containers != nil {
 		out.Containers = make([]api.Container, len(in.Containers))
 		for i := range in.Containers {
@@ -419,6 +509,33 @@ func Convert_api_Pod_To_v1_Pod(in *api.Pod, out *Pod, s conversion.Scope) error
 	if err := autoConvert_api_Pod_To_v1_Pod(in, out, s); err != nil {
 		return err
 	}
+
+	// TODO: when we move init container to beta, remove these conversions
+	if len(out.Spec.InitContainers) > 0 {
+		if out.Annotations == nil {
+			out.Annotations = make(map[string]string)
+		}
+		value, err := json.Marshal(out.Spec.InitContainers)
+		if err != nil {
+			return err
+		}
+		out.Annotations[PodInitContainersAnnotationKey] = string(value)
+	} else {
+		delete(out.Annotations, PodInitContainersAnnotationKey)
+	}
+	if len(out.Status.InitContainerStatuses) > 0 {
+		if out.Annotations == nil {
+			out.Annotations = make(map[string]string)
+		}
+		value, err := json.Marshal(out.Status.InitContainerStatuses)
+		if err != nil {
+			return err
+		}
+		out.Annotations[PodInitContainerStatusesAnnotationKey] = string(value)
+	} else {
+		delete(in.Annotations, PodInitContainerStatusesAnnotationKey)
+	}
+
 	// We need to reset certain fields for mirror pods from pre-v1.1 kubelet
 	// (#15960).
 	// TODO: Remove this code after we drop support for v1.0 kubelets.
@@ -434,6 +551,24 @@ func Convert_api_Pod_To_v1_Pod(in *api.Pod, out *Pod, s conversion.Scope) error
 }
 
 func Convert_v1_Pod_To_api_Pod(in *Pod, out *api.Pod, s conversion.Scope) error {
+	// TODO: when we move init container to beta, remove these conversions
+	if value := in.Annotations[PodInitContainersAnnotationKey]; len(value) > 0 {
+		delete(in.Annotations, PodInitContainersAnnotationKey)
+		var values []Container
+		if err := json.Unmarshal([]byte(value), &values); err != nil {
+			return err
+		}
+		in.Spec.InitContainers = values
+	}
+	if value := in.Annotations[PodInitContainerStatusesAnnotationKey]; len(value) > 0 {
+		delete(in.Annotations, PodInitContainerStatusesAnnotationKey)
+		var values []ContainerStatus
+		if err := json.Unmarshal([]byte(value), &values); err != nil {
+			return err
+		}
+		in.Status.InitContainerStatuses = values
+	}
+
 	return autoConvert_v1_Pod_To_api_Pod(in, out, s)
 }
 

pkg/api/v1/conversion_generated.go

@@ -5012,6 +5012,17 @@ func autoConvert_api_PodSpec_To_v1_PodSpec(in *api.PodSpec, out *PodSpec, s conv
 	} else {
 		out.Volumes = nil
 	}
+	if in.InitContainers != nil {
+		in, out := &in.InitContainers, &out.InitContainers
+		*out = make([]Container, len(*in))
+		for i := range *in {
+			if err := Convert_api_Container_To_v1_Container(&(*in)[i], &(*out)[i], s); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainers = nil
+	}
 	if in.Containers != nil {
 		in, out := &in.Containers, &out.Containers
 		*out = make([]Container, len(*in))
@@ -5101,6 +5112,17 @@ func autoConvert_v1_PodStatus_To_api_PodStatus(in *PodStatus, out *api.PodStatus
 	} else {
 		out.StartTime = nil
 	}
+	if in.InitContainerStatuses != nil {
+		in, out := &in.InitContainerStatuses, &out.InitContainerStatuses
+		*out = make([]api.ContainerStatus, len(*in))
+		for i := range *in {
+			if err := Convert_v1_ContainerStatus_To_api_ContainerStatus(&(*in)[i], &(*out)[i], s); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainerStatuses = nil
+	}
 	if in.ContainerStatuses != nil {
 		in, out := &in.ContainerStatuses, &out.ContainerStatuses
 		*out = make([]api.ContainerStatus, len(*in))
@@ -5145,6 +5167,17 @@ func autoConvert_api_PodStatus_To_v1_PodStatus(in *api.PodStatus, out *PodStatus
 	} else {
 		out.StartTime = nil
 	}
+	if in.InitContainerStatuses != nil {
+		in, out := &in.InitContainerStatuses, &out.InitContainerStatuses
+		*out = make([]ContainerStatus, len(*in))
+		for i := range *in {
+			if err := Convert_api_ContainerStatus_To_v1_ContainerStatus(&(*in)[i], &(*out)[i], s); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainerStatuses = nil
+	}
 	if in.ContainerStatuses != nil {
 		in, out := &in.ContainerStatuses, &out.ContainerStatuses
 		*out = make([]ContainerStatus, len(*in))
@@ -5176,10 +5209,6 @@ func autoConvert_v1_PodStatusResult_To_api_PodStatusResult(in *PodStatusResult,
 	return nil
 }
 
-func Convert_v1_PodStatusResult_To_api_PodStatusResult(in *PodStatusResult, out *api.PodStatusResult, s conversion.Scope) error {
-	return autoConvert_v1_PodStatusResult_To_api_PodStatusResult(in, out, s)
-}
-
 func autoConvert_api_PodStatusResult_To_v1_PodStatusResult(in *api.PodStatusResult, out *PodStatusResult, s conversion.Scope) error {
 	if err := api.Convert_unversioned_TypeMeta_To_unversioned_TypeMeta(&in.TypeMeta, &out.TypeMeta, s); err != nil {
 		return err
@@ -5193,10 +5222,6 @@ func autoConvert_api_PodStatusResult_To_v1_PodStatusResult(in *api.PodStatusResu
 	return nil
 }
 
-func Convert_api_PodStatusResult_To_v1_PodStatusResult(in *api.PodStatusResult, out *PodStatusResult, s conversion.Scope) error {
-	return autoConvert_api_PodStatusResult_To_v1_PodStatusResult(in, out, s)
-}
-
 func autoConvert_v1_PodTemplate_To_api_PodTemplate(in *PodTemplate, out *api.PodTemplate, s conversion.Scope) error {
 	if err := api.Convert_unversioned_TypeMeta_To_unversioned_TypeMeta(&in.TypeMeta, &out.TypeMeta, s); err != nil {
 		return err
@@ -5291,10 +5316,6 @@ func autoConvert_v1_PodTemplateSpec_To_api_PodTemplateSpec(in *PodTemplateSpec,
 	return nil
 }
 
-func Convert_v1_PodTemplateSpec_To_api_PodTemplateSpec(in *PodTemplateSpec, out *api.PodTemplateSpec, s conversion.Scope) error {
-	return autoConvert_v1_PodTemplateSpec_To_api_PodTemplateSpec(in, out, s)
-}
-
 func autoConvert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in *api.PodTemplateSpec, out *PodTemplateSpec, s conversion.Scope) error {
 	if err := Convert_api_ObjectMeta_To_v1_ObjectMeta(&in.ObjectMeta, &out.ObjectMeta, s); err != nil {
 		return err
@@ -5305,10 +5326,6 @@ func autoConvert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in *api.PodTemplateSp
 	return nil
 }
 
-func Convert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in *api.PodTemplateSpec, out *PodTemplateSpec, s conversion.Scope) error {
-	return autoConvert_api_PodTemplateSpec_To_v1_PodTemplateSpec(in, out, s)
-}
-
 func autoConvert_v1_Preconditions_To_api_Preconditions(in *Preconditions, out *api.Preconditions, s conversion.Scope) error {
 	if in.UID != nil {
 		in, out := &in.UID, &out.UID

pkg/api/v1/deep_copy_generated.go

@@ -2176,6 +2176,17 @@ func DeepCopy_v1_PodSpec(in PodSpec, out *PodSpec, c *conversion.Cloner) error {
 	} else {
 		out.Volumes = nil
 	}
+	if in.InitContainers != nil {
+		in, out := in.InitContainers, &out.InitContainers
+		*out = make([]Container, len(in))
+		for i := range in {
+			if err := DeepCopy_v1_Container(in[i], &(*out)[i], c); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainers = nil
+	}
 	if in.Containers != nil {
 		in, out := in.Containers, &out.Containers
 		*out = make([]Container, len(in))
@@ -2269,6 +2280,17 @@ func DeepCopy_v1_PodStatus(in PodStatus, out *PodStatus, c *conversion.Cloner) e
 	} else {
 		out.StartTime = nil
 	}
+	if in.InitContainerStatuses != nil {
+		in, out := in.InitContainerStatuses, &out.InitContainerStatuses
+		*out = make([]ContainerStatus, len(in))
+		for i := range in {
+			if err := DeepCopy_v1_ContainerStatus(in[i], &(*out)[i], c); err != nil {
+				return err
+			}
+		}
+	} else {
+		out.InitContainerStatuses = nil
+	}
 	if in.ContainerStatuses != nil {
 		in, out := in.ContainerStatuses, &out.ContainerStatuses
 		*out = make([]ContainerStatus, len(in))

pkg/api/v1/generated.proto

@@ -295,7 +295,7 @@ message Container {
   // More info: http://releases.k8s.io/HEAD/docs/user-guide/persistent-volumes.md#resources
   optional ResourceRequirements resources = 8;
 
-  // Pod volumes to mount into the container's filesyste.
+  // Pod volumes to mount into the container's filesystem.
   // Cannot be updated.
   repeated VolumeMount volumeMounts = 9;
 

pkg/api/v1/types.generated.go

@@ -53783,7 +53783,7 @@ func (x codecSelfer1234) decSlicePod(v *[]Pod, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 600)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 648)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -53902,7 +53902,7 @@ func (x codecSelfer1234) decSlicePodTemplate(v *[]PodTemplate, d *codec1978.Deco
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 672)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 696)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/api/v1/types.go

@@ -1106,7 +1106,7 @@ type Container struct {
 	// Cannot be updated.
 	// More info: http://releases.k8s.io/HEAD/docs/user-guide/persistent-volumes.md#resources
 	Resources ResourceRequirements `json:"resources,omitempty" protobuf:"bytes,8,opt,name=resources"`
-	// Pod volumes to mount into the container's filesyste.
+	// Pod volumes to mount into the container's filesystem.
 	// Cannot be updated.
 	VolumeMounts []VolumeMount `json:"volumeMounts,omitempty" patchStrategy:"merge" patchMergeKey:"name" protobuf:"bytes,9,rep,name=volumeMounts"`
 	// Periodic probe of container liveness.
@@ -1541,11 +1541,36 @@ type PreferredSchedulingTerm struct {
 	Preference NodeSelectorTerm `json:"preference" protobuf:"bytes,2,opt,name=preference"`
 }
 
+const (
+	// This annotation key will be used to contain an array of v1 JSON encoded Containers
+	// for init containers. The annotation will be placed into the internal type and cleared.
+	PodInitContainersAnnotationKey = "pod.alpha.kubernetes.io/init-containers"
+	// This annotation key will be used to contain an array of v1 JSON encoded
+	// ContainerStatuses for init containers. The annotation will be placed into the internal
+	// type and cleared.
+	PodInitContainerStatusesAnnotationKey = "pod.alpha.kubernetes.io/init-container-statuses"
+)
+
 // PodSpec is a description of a pod.
 type PodSpec struct {
 	// List of volumes that can be mounted by containers belonging to the pod.
 	// More info: http://releases.k8s.io/HEAD/docs/user-guide/volumes.md
 	Volumes []Volume `json:"volumes,omitempty" patchStrategy:"merge" patchMergeKey:"name" protobuf:"bytes,1,rep,name=volumes"`
+	// List of initialization containers belonging to the pod.
+	// Init containers are executed in order prior to containers being started. If any
+	// init container fails, the pod is considered to have failed and is handled according
+	// to its restartPolicy. The name for an init container or normal container must be
+	// unique among all containers.
+	// Init containers may not have Lifecycle actions, Readiness probes, or Liveness probes.
+	// The resourceRequirements of an init container are taken into account during scheduling
+	// by finding the highest request/limit for each resource type, and then using the max of
+	// of that value or the sum of the normal containers. Limits are applied to init containers
+	// in a similar fashion.
+	// Init containers cannot currently be added or removed.
+	// Init containers are in alpha state and may change without notice.
+	// Cannot be updated.
+	// More info: http://releases.k8s.io/HEAD/docs/user-guide/containers.md
+	InitContainers []Container `json:"-"  patchStrategy:"merge" patchMergeKey:"name"`
 	// List of containers belonging to the pod.
 	// Containers cannot currently be added or removed.
 	// There must be at least one container in a Pod.
@@ -1679,6 +1704,12 @@ type PodStatus struct {
 	// This is before the Kubelet pulled the container image(s) for the pod.
 	StartTime *unversioned.Time `json:"startTime,omitempty" protobuf:"bytes,7,opt,name=startTime"`
 
+	// The list has one entry per init container in the manifest. The most recent successful
+	// init container will have ready = true, the most recently started container will have
+	// startTime set.
+	// Init containers are in alpha state and may change without notice.
+	// More info: http://releases.k8s.io/HEAD/docs/user-guide/pod-states.md#container-statuses
+	InitContainerStatuses []ContainerStatus `json:"-"`
 	// The list has one entry per container in the manifest. Each entry is currently the output
 	// of `docker inspect`.
 	// More info: http://releases.k8s.io/HEAD/docs/user-guide/pod-states.md#container-statuses

pkg/api/v1/types_swagger_doc_generated.go

@@ -186,7 +186,7 @@ var map_Container = map[string]string{
 	"ports":                  "List of ports to expose from the container. Exposing a port here gives the system additional information about the network connections a container uses, but is primarily informational. Not specifying a port here DOES NOT prevent that port from being exposed. Any port which is listening on the default \"0.0.0.0\" address inside a container will be accessible from the network. Cannot be updated.",
 	"env":                    "List of environment variables to set in the container. Cannot be updated.",
 	"resources":              "Compute Resources required by this container. Cannot be updated. More info: http://releases.k8s.io/HEAD/docs/user-guide/persistent-volumes.md#resources",
-	"volumeMounts":           "Pod volumes to mount into the container's filesyste. Cannot be updated.",
+	"volumeMounts":           "Pod volumes to mount into the container's filesystem. Cannot be updated.",
 	"livenessProbe":          "Periodic probe of container liveness. Container will be restarted if the probe fails. Cannot be updated. More info: http://releases.k8s.io/HEAD/docs/user-guide/pod-states.md#container-probes",
 	"readinessProbe":         "Periodic probe of container service readiness. Container will be removed from service endpoints if the probe fails. Cannot be updated. More info: http://releases.k8s.io/HEAD/docs/user-guide/pod-states.md#container-probes",
 	"lifecycle":              "Actions that the management system should take in response to container lifecycle events. Cannot be updated.",

pkg/api/validation/validation.go

@@ -1326,6 +1326,37 @@ func validatePullPolicy(policy api.PullPolicy, fldPath *field.Path) field.ErrorL
 	return allErrors
 }
 
+func validateInitContainers(containers, otherContainers []api.Container, volumes sets.String, fldPath *field.Path) field.ErrorList {
+	var allErrs field.ErrorList
+	if len(containers) > 0 {
+		allErrs = append(allErrs, validateContainers(containers, volumes, fldPath)...)
+	}
+
+	allNames := sets.String{}
+	for _, ctr := range otherContainers {
+		allNames.Insert(ctr.Name)
+	}
+	for i, ctr := range containers {
+		idxPath := fldPath.Index(i)
+		if allNames.Has(ctr.Name) {
+			allErrs = append(allErrs, field.Duplicate(idxPath.Child("name"), ctr.Name))
+		}
+		if len(ctr.Name) > 0 {
+			allNames.Insert(ctr.Name)
+		}
+		if ctr.Lifecycle != nil {
+			allErrs = append(allErrs, field.Invalid(idxPath.Child("lifecycle"), ctr.Lifecycle, "must not be set for init containers"))
+		}
+		if ctr.LivenessProbe != nil {
+			allErrs = append(allErrs, field.Invalid(idxPath.Child("livenessProbe"), ctr.LivenessProbe, "must not be set for init containers"))
+		}
+		if ctr.ReadinessProbe != nil {
+			allErrs = append(allErrs, field.Invalid(idxPath.Child("readinessProbe"), ctr.ReadinessProbe, "must not be set for init containers"))
+		}
+	}
+	return allErrs
+}
+
 func validateContainers(containers []api.Container, volumes sets.String, fldPath *field.Path) field.ErrorList {
 	allErrs := field.ErrorList{}
 
@@ -1451,6 +1482,7 @@ func ValidatePodSpec(spec *api.PodSpec, fldPath *field.Path) field.ErrorList {
 	allVolumes, vErrs := validateVolumes(spec.Volumes, fldPath.Child("volumes"))
 	allErrs = append(allErrs, vErrs...)
 	allErrs = append(allErrs, validateContainers(spec.Containers, allVolumes, fldPath.Child("containers"))...)
+	allErrs = append(allErrs, validateInitContainers(spec.InitContainers, spec.Containers, allVolumes, fldPath.Child("initContainers"))...)
 	allErrs = append(allErrs, validateRestartPolicy(&spec.RestartPolicy, fldPath.Child("restartPolicy"))...)
 	allErrs = append(allErrs, validateDNSPolicy(&spec.DNSPolicy, fldPath.Child("dnsPolicy"))...)
 	allErrs = append(allErrs, unversionedvalidation.ValidateLabels(spec.NodeSelector, fldPath.Child("nodeSelector"))...)

pkg/apis/apps/types.generated.go

@@ -1555,7 +1555,7 @@ func (x codecSelfer1234) decSlicePetSet(v *[]PetSet, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 720)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 744)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/apps/v1alpha1/types.generated.go

@@ -1585,7 +1585,7 @@ func (x codecSelfer1234) decSlicePetSet(v *[]PetSet, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 744)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/batch/types.generated.go

@@ -4205,7 +4205,7 @@ func (x codecSelfer1234) decSliceJob(v *[]Job, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 744)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -4443,7 +4443,7 @@ func (x codecSelfer1234) decSliceScheduledJob(v *[]ScheduledJob, d *codec1978.De
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 976)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 1000)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/batch/v1/conversion_generated.go

@@ -217,8 +217,7 @@ func autoConvert_v1_JobSpec_To_batch_JobSpec(in *JobSpec, out *batch.JobSpec, s
 	} else {
 		out.ManualSelector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := api_v1.Convert_v1_PodTemplateSpec_To_api_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil
@@ -262,8 +261,7 @@ func autoConvert_batch_JobSpec_To_v1_JobSpec(in *batch.JobSpec, out *JobSpec, s
 	} else {
 		out.ManualSelector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := api_v1.Convert_api_PodTemplateSpec_To_v1_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil

pkg/apis/batch/v1/types.generated.go

@@ -2867,7 +2867,7 @@ func (x codecSelfer1234) decSliceJob(v *[]Job, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 792)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/batch/v2alpha1/conversion_generated.go

@@ -223,8 +223,7 @@ func autoConvert_v2alpha1_JobSpec_To_batch_JobSpec(in *JobSpec, out *batch.JobSp
 	} else {
 		out.ManualSelector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := v1.Convert_v1_PodTemplateSpec_To_api_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil
@@ -268,8 +267,7 @@ func autoConvert_batch_JobSpec_To_v2alpha1_JobSpec(in *batch.JobSpec, out *JobSp
 	} else {
 		out.ManualSelector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := v1.Convert_api_PodTemplateSpec_To_v1_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil

pkg/apis/batch/v2alpha1/types.generated.go

@@ -4739,7 +4739,7 @@ func (x codecSelfer1234) decSliceJob(v *[]Job, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 792)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/extensions/types.generated.go

@@ -14325,7 +14325,7 @@ func (x codecSelfer1234) decSliceDeployment(v *[]Deployment, d *codec1978.Decode
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 744)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -14444,7 +14444,7 @@ func (x codecSelfer1234) decSliceDaemonSet(v *[]DaemonSet, d *codec1978.Decoder)
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 672)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 696)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -15158,7 +15158,7 @@ func (x codecSelfer1234) decSliceReplicaSet(v *[]ReplicaSet, d *codec1978.Decode
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 680)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 704)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/apis/extensions/v1beta1/conversion_generated.go

@@ -401,8 +401,7 @@ func autoConvert_v1beta1_DaemonSetSpec_To_extensions_DaemonSetSpec(in *DaemonSet
 	} else {
 		out.Selector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := v1.Convert_v1_PodTemplateSpec_To_api_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil
@@ -422,8 +421,7 @@ func autoConvert_extensions_DaemonSetSpec_To_v1beta1_DaemonSetSpec(in *extension
 	} else {
 		out.Selector = nil
 	}
-	// TODO: Inefficient conversion - can we improve it?
+	if err := v1.Convert_api_PodTemplateSpec_To_v1_PodTemplateSpec(&in.Template, &out.Template, s); err != nil {
-	if err := s.Convert(&in.Template, &out.Template, 0); err != nil {
 		return err
 	}
 	return nil

pkg/apis/extensions/v1beta1/types.generated.go

@@ -19988,7 +19988,7 @@ func (x codecSelfer1234) decSliceDeployment(v *[]Deployment, d *codec1978.Decode
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 792)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -20107,7 +20107,7 @@ func (x codecSelfer1234) decSliceDaemonSet(v *[]DaemonSet, d *codec1978.Decoder)
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 696)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 720)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -20345,7 +20345,7 @@ func (x codecSelfer1234) decSliceJob(v *[]Job, d *codec1978.Decoder) {
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 768)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 792)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]
@@ -21178,7 +21178,7 @@ func (x codecSelfer1234) decSliceReplicaSet(v *[]ReplicaSet, d *codec1978.Decode
 
 			yyrg1 := len(yyv1) > 0
 			yyv21 := yyv1
-			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 704)
+			yyrl1, yyrt1 = z.DecInferLen(yyl1, z.DecBasicHandle().MaxInitLen, 728)
 			if yyrt1 {
 				if yyrl1 <= cap(yyv1) {
 					yyv1 = yyv1[:yyrl1]

pkg/kubectl/describe.go

@@ -524,7 +524,10 @@ func describePod(pod *api.Pod, events *api.EventList) (string, error) {
 		}
 		fmt.Fprintf(out, "IP:\t%s\n", pod.Status.PodIP)
 		fmt.Fprintf(out, "Controllers:\t%s\n", printControllers(pod.Annotations))
-		describeContainers(pod.Spec.Containers, pod.Status.ContainerStatuses, EnvValueRetriever(pod), out, "")
+		if len(pod.Spec.InitContainers) > 0 {
+			describeContainers("Init Containers", pod.Spec.InitContainers, pod.Status.InitContainerStatuses, EnvValueRetriever(pod), out, "")
+		}
+		describeContainers("Containers", pod.Spec.Containers, pod.Status.ContainerStatuses, EnvValueRetriever(pod), out, "")
 		if len(pod.Status.Conditions) > 0 {
 			fmt.Fprint(out, "Conditions:\n  Type\tStatus\n")
 			for _, c := range pod.Status.Conditions {
@@ -782,12 +785,16 @@ func (d *PersistentVolumeClaimDescriber) Describe(namespace, name string, descri
 }
 
 // TODO: Do a better job at indenting, maybe by using a prefix writer
-func describeContainers(containers []api.Container, containerStatuses []api.ContainerStatus, resolverFn EnvVarResolverFunc, out io.Writer, space string) {
+func describeContainers(label string, containers []api.Container, containerStatuses []api.ContainerStatus, resolverFn EnvVarResolverFunc, out io.Writer, space string) {
 	statuses := map[string]api.ContainerStatus{}
 	for _, status := range containerStatuses {
 		statuses[status.Name] = status
 	}
-	fmt.Fprintf(out, "%sContainers:\n", space)
+	if len(containers) == 0 {
+		fmt.Fprintf(out, "%s%s: <none>\n", space, label)
+	} else {
+		fmt.Fprintf(out, "%s%s:\n", space, label)
+	}
 	for _, container := range containers {
 		status, ok := statuses[container.Name]
 		nameIndent := ""
@@ -1037,7 +1044,10 @@ func DescribePodTemplate(template *api.PodTemplateSpec, out io.Writer) {
 	if len(template.Spec.ServiceAccountName) > 0 {
 		fmt.Fprintf(out, "  Service Account:\t%s\n", template.Spec.ServiceAccountName)
 	}
-	describeContainers(template.Spec.Containers, nil, nil, out, "  ")
+	if len(template.Spec.InitContainers) > 0 {
+		describeContainers("Init Containers", template.Spec.InitContainers, nil, nil, out, "  ")
+	}
+	describeContainers("Containers", template.Spec.Containers, nil, nil, out, "  ")
 	describeVolumes(template.Spec.Volumes, out, "  ")
 }
 

pkg/kubectl/describe_test.go

@@ -268,7 +268,7 @@ func TestDescribeContainers(t *testing.T) {
 				ContainerStatuses: []api.ContainerStatus{testCase.status},
 			},
 		}
-		describeContainers(pod.Spec.Containers, pod.Status.ContainerStatuses, EnvValueRetriever(&pod), out, "")
+		describeContainers("Containers", pod.Spec.Containers, pod.Status.ContainerStatuses, EnvValueRetriever(&pod), out, "")
 		output := out.String()
 		for _, expected := range testCase.expectedElements {
 			if !strings.Contains(output, expected) {

pkg/kubectl/resource_printer.go

@@ -594,22 +594,51 @@ func printPodBase(pod *api.Pod, w io.Writer, options PrintOptions) error {
 		reason = pod.Status.Reason
 	}
 
-	for i := len(pod.Status.ContainerStatuses) - 1; i >= 0; i-- {
+	initializing := false
-		container := pod.Status.ContainerStatuses[i]
+	for i := range pod.Status.InitContainerStatuses {
-
+		container := pod.Status.InitContainerStatuses[i]
-		restarts += int(container.RestartCount)
+		switch {
-		if container.State.Waiting != nil && container.State.Waiting.Reason != "" {
+		case container.State.Terminated != nil && container.State.Terminated.ExitCode == 0:
-			reason = container.State.Waiting.Reason
+			continue
-		} else if container.State.Terminated != nil && container.State.Terminated.Reason != "" {
+		case container.State.Terminated != nil:
-			reason = container.State.Terminated.Reason
+			// initialization is failed
-		} else if container.State.Terminated != nil && container.State.Terminated.Reason == "" {
+			if len(container.State.Terminated.Reason) == 0 {
-			if container.State.Terminated.Signal != 0 {
+				if container.State.Terminated.Signal != 0 {
-				reason = fmt.Sprintf("Signal:%d", container.State.Terminated.Signal)
+					reason = fmt.Sprintf("Init:Signal:%d", container.State.Terminated.Signal)
+				} else {
+					reason = fmt.Sprintf("Init:ExitCode:%d", container.State.Terminated.ExitCode)
+				}
 			} else {
-				reason = fmt.Sprintf("ExitCode:%d", container.State.Terminated.ExitCode)
+				reason = "Init:" + container.State.Terminated.Reason
+			}
+			initializing = true
+		case container.State.Waiting != nil && len(container.State.Waiting.Reason) > 0 && container.State.Waiting.Reason != "PodInitializing":
+			reason = "Init:" + container.State.Waiting.Reason
+			initializing = true
+		default:
+			reason = fmt.Sprintf("Init:%d/%d", i, len(pod.Spec.InitContainers))
+			initializing = true
+		}
+		break
+	}
+	if !initializing {
+		for i := len(pod.Status.ContainerStatuses) - 1; i >= 0; i-- {
+			container := pod.Status.ContainerStatuses[i]
+
+			restarts += int(container.RestartCount)
+			if container.State.Waiting != nil && container.State.Waiting.Reason != "" {
+				reason = container.State.Waiting.Reason
+			} else if container.State.Terminated != nil && container.State.Terminated.Reason != "" {
+				reason = container.State.Terminated.Reason
+			} else if container.State.Terminated != nil && container.State.Terminated.Reason == "" {
+				if container.State.Terminated.Signal != 0 {
+					reason = fmt.Sprintf("Signal:%d", container.State.Terminated.Signal)
+				} else {
+					reason = fmt.Sprintf("ExitCode:%d", container.State.Terminated.ExitCode)
+				}
+			} else if container.Ready && container.State.Running != nil {
+				readyContainers++
 			}
-		} else if container.Ready && container.State.Running != nil {
-			readyContainers++
 		}
 	}
 	if pod.DeletionTimestamp != nil {

pkg/kubelet/container/ref.go

@@ -57,5 +57,15 @@ func fieldPath(pod *api.Pod, container *api.Container) (string, error) {
 			}
 		}
 	}
+	for i := range pod.Spec.InitContainers {
+		here := &pod.Spec.InitContainers[i]
+		if here.Name == container.Name {
+			if here.Name == "" {
+				return fmt.Sprintf("spec.initContainers[%d]", i), nil
+			} else {
+				return fmt.Sprintf("spec.initContainers{%s}", here.Name), nil
+			}
+		}
+	}
 	return "", fmt.Errorf("container %#v not found in pod %#v", container, pod)
 }

pkg/kubelet/container/sync_result.go

@@ -50,9 +50,10 @@ var (
 )
 
 var (
-	ErrRunContainer  = errors.New("RunContainerError")
+	ErrRunContainer     = errors.New("RunContainerError")
-	ErrKillContainer = errors.New("KillContainerError")
+	ErrKillContainer    = errors.New("KillContainerError")
-	ErrVerifyNonRoot = errors.New("VerifyNonRootError")
+	ErrVerifyNonRoot    = errors.New("VerifyNonRootError")
+	ErrRunInitContainer = errors.New("RunInitContainerError")
 )
 
 var (
@@ -69,6 +70,7 @@ const (
 	KillContainer   SyncAction = "KillContainer"
 	SetupNetwork    SyncAction = "SetupNetwork"
 	TeardownNetwork SyncAction = "TeardownNetwork"
+	InitContainer   SyncAction = "InitContainer"
 )
 
 // SyncResult is the result of sync action.

pkg/kubelet/dockertools/manager.go

@@ -37,6 +37,7 @@ import (
 	dockernat "github.com/docker/go-connections/nat"
 	"github.com/golang/glog"
 	cadvisorapi "github.com/google/cadvisor/info/v1"
+
 	"k8s.io/kubernetes/pkg/api"
 	"k8s.io/kubernetes/pkg/api/unversioned"
 	"k8s.io/kubernetes/pkg/client/record"
@@ -57,6 +58,7 @@ import (
 	"k8s.io/kubernetes/pkg/util/oom"
 	"k8s.io/kubernetes/pkg/util/procfs"
 	utilruntime "k8s.io/kubernetes/pkg/util/runtime"
+	"k8s.io/kubernetes/pkg/util/sets"
 	utilstrings "k8s.io/kubernetes/pkg/util/strings"
 )
 
@@ -876,6 +878,9 @@ func (dm *DockerManager) podInfraContainerChanged(pod *api.Pod, podInfraContaine
 	} else if dm.networkPlugin.Name() != "cni" && dm.networkPlugin.Name() != "kubenet" {
 		// Docker only exports ports from the pod infra container. Let's
 		// collect all of the relevant ports and export them.
+		for _, container := range pod.Spec.InitContainers {
+			ports = append(ports, container.Ports...)
+		}
 		for _, container := range pod.Spec.Containers {
 			ports = append(ports, container.Ports...)
 		}
@@ -1179,6 +1184,14 @@ func (dm *DockerManager) killPodWithSyncResult(pod *api.Pod, runningPod kubecont
 						break
 					}
 				}
+				if containerSpec == nil {
+					for i, c := range pod.Spec.InitContainers {
+						if c.Name == container.Name {
+							containerSpec = &pod.Spec.InitContainers[i]
+							break
+						}
+					}
+				}
 			}
 
 			// TODO: Handle this without signaling the pod infra container to
@@ -1369,6 +1382,14 @@ func containerAndPodFromLabels(inspect *dockertypes.ContainerJSON) (pod *api.Pod
 					break
 				}
 			}
+			if container == nil {
+				for ix := range pod.Spec.InitContainers {
+					if pod.Spec.InitContainers[ix].Name == name {
+						container = &pod.Spec.InitContainers[ix]
+						break
+					}
+				}
+			}
 			if container == nil {
 				err = fmt.Errorf("unable to find container %s in pod %v", name, pod)
 			}
@@ -1425,6 +1446,7 @@ func (dm *DockerManager) runContainerInPod(pod *api.Pod, container *api.Containe
 	if err != nil {
 		glog.Errorf("Can't make a ref to pod %v, container %v: '%v'", pod.Name, container.Name, err)
 	}
+	glog.Infof("Generating ref for container %s: %#v", container.Name, ref)
 
 	opts, err := dm.runtimeHelper.GenerateRunContainerOptions(pod, container, podIP)
 	if err != nil {
@@ -1603,6 +1625,9 @@ func (dm *DockerManager) createPodInfraContainer(pod *api.Pod) (kubecontainer.Do
 	} else {
 		// Docker only exports ports from the pod infra container.  Let's
 		// collect all of the relevant ports and export them.
+		for _, container := range pod.Spec.InitContainers {
+			ports = append(ports, container.Ports...)
+		}
 		for _, container := range pod.Spec.Containers {
 			ports = append(ports, container.Ports...)
 		}
@@ -1640,13 +1665,16 @@ func (dm *DockerManager) createPodInfraContainer(pod *api.Pod) (kubecontainer.Do
 //   should be kept running. If startInfraContainer is false then it contains an entry for infraContainerId (mapped to -1).
 //   It shouldn't be the case where containersToStart is empty and containersToKeep contains only infraContainerId. In such case
 //   Infra Container should be killed, hence it's removed from this map.
-// - all running containers which are NOT contained in containersToKeep should be killed.
+// - all init containers are stored in initContainersToKeep
+// - all running containers which are NOT contained in containersToKeep and initContainersToKeep should be killed.
 type podContainerChangesSpec struct {
-	StartInfraContainer bool
+	StartInfraContainer  bool
-	InfraChanged        bool
+	InfraChanged         bool
-	InfraContainerId    kubecontainer.DockerID
+	InfraContainerId     kubecontainer.DockerID
-	ContainersToStart   map[int]string
+	InitFailed           bool
-	ContainersToKeep    map[kubecontainer.DockerID]int
+	InitContainersToKeep map[kubecontainer.DockerID]int
+	ContainersToStart    map[int]string
+	ContainersToKeep     map[kubecontainer.DockerID]int
 }
 
 func (dm *DockerManager) computePodContainerChanges(pod *api.Pod, podStatus *kubecontainer.PodStatus) (podContainerChangesSpec, error) {
@@ -1683,6 +1711,35 @@ func (dm *DockerManager) computePodContainerChanges(pod *api.Pod, podStatus *kub
 		containersToKeep[podInfraContainerID] = -1
 	}
 
+	// check the status of the init containers
+	initFailed := false
+	initContainersToKeep := make(map[kubecontainer.DockerID]int)
+	// always reset the init containers if the pod is reset
+	if !createPodInfraContainer {
+		// keep all successfully completed containers up to and including the first failing container
+	Containers:
+		for i, container := range pod.Spec.InitContainers {
+			containerStatus := podStatus.FindContainerStatusByName(container.Name)
+			if containerStatus == nil {
+				continue
+			}
+			switch {
+			case containerStatus == nil:
+				continue
+			case containerStatus.State == kubecontainer.ContainerStateRunning:
+				initContainersToKeep[kubecontainer.DockerID(containerStatus.ID.ID)] = i
+			case containerStatus.State == kubecontainer.ContainerStateExited:
+				initContainersToKeep[kubecontainer.DockerID(containerStatus.ID.ID)] = i
+				// TODO: should we abstract the "did the init container fail" check?
+				if containerStatus.ExitCode != 0 {
+					initFailed = true
+					break Containers
+				}
+			}
+		}
+	}
+
+	// check the status of the containers
 	for index, container := range pod.Spec.Containers {
 		expectedHash := kubecontainer.HashContainer(&container)
 
@@ -1716,6 +1773,19 @@ func (dm *DockerManager) computePodContainerChanges(pod *api.Pod, podStatus *kub
 			continue
 		}
 
+		if initFailed {
+			// initialization failed and Container exists
+			// If we have an initialization failure everything will be killed anyway
+			// If RestartPolicy is Always or OnFailure we restart containers that were running before we
+			// killed them when re-running initialization
+			if pod.Spec.RestartPolicy != api.RestartPolicyNever {
+				message := fmt.Sprintf("Failed to initialize pod. %q will be restarted.", container.Name)
+				glog.V(1).Info(message)
+				containersToStart[index] = message
+			}
+			continue
+		}
+
 		// At this point, the container is running and pod infra container is good.
 		// We will look for changes and check healthiness for the container.
 		containerChanged := hash != 0 && hash != expectedHash
@@ -1743,17 +1813,21 @@ func (dm *DockerManager) computePodContainerChanges(pod *api.Pod, podStatus *kub
 	// (In fact, when createPodInfraContainer is false, containersToKeep will not be touched).
 	// - createPodInfraContainer is false and containersToKeep contains at least ID of Infra Container
 
-	// If Infra container is the last running one, we don't want to keep it.
+	// If Infra container is the last running one, we don't want to keep it, and we don't want to
+	// keep any init containers.
 	if !createPodInfraContainer && len(containersToStart) == 0 && len(containersToKeep) == 1 {
 		containersToKeep = make(map[kubecontainer.DockerID]int)
+		initContainersToKeep = make(map[kubecontainer.DockerID]int)
 	}
 
 	return podContainerChangesSpec{
-		StartInfraContainer: createPodInfraContainer,
+		StartInfraContainer:  createPodInfraContainer,
-		InfraChanged:        changed,
+		InfraChanged:         changed,
-		InfraContainerId:    podInfraContainerID,
+		InfraContainerId:     podInfraContainerID,
-		ContainersToStart:   containersToStart,
+		InitFailed:           initFailed,
-		ContainersToKeep:    containersToKeep,
+		InitContainersToKeep: initContainersToKeep,
+		ContainersToStart:    containersToStart,
+		ContainersToKeep:     containersToKeep,
 	}, nil
 }
 
@@ -1797,7 +1871,8 @@ func (dm *DockerManager) SyncPod(pod *api.Pod, _ api.PodStatus, podStatus *kubec
 		runningContainerStatues := podStatus.GetRunningContainerStatuses()
 		for _, containerStatus := range runningContainerStatues {
 			_, keep := containerChanges.ContainersToKeep[kubecontainer.DockerID(containerStatus.ID.ID)]
-			if !keep {
+			_, keepInit := containerChanges.InitContainersToKeep[kubecontainer.DockerID(containerStatus.ID.ID)]
+			if !keep && !keepInit {
 				glog.V(3).Infof("Killing unwanted container %q(id=%q) for pod %q", containerStatus.Name, containerStatus.ID, format.Pod(pod))
 				// attempt to find the appropriate container policy
 				var podContainer *api.Container
@@ -1820,6 +1895,9 @@ func (dm *DockerManager) SyncPod(pod *api.Pod, _ api.PodStatus, podStatus *kubec
 		}
 	}
 
+	// Keep terminated init containers fairly aggressively controlled
+	dm.pruneInitContainersBeforeStart(pod, podStatus, containerChanges.InitContainersToKeep)
+
 	// We pass the value of the podIP down to runContainerInPod, which in turn
 	// passes it to various other functions, in order to facilitate
 	// functionality that requires this value (hosts file and downward API)
@@ -1889,14 +1967,78 @@ func (dm *DockerManager) SyncPod(pod *api.Pod, _ api.PodStatus, podStatus *kubec
 		}
 	}
 
-	// Start everything
+	next, status, done := findActiveInitContainer(pod, podStatus)
+	if status != nil {
+		if status.ExitCode != 0 {
+			// container initialization has failed, flag the pod as failed
+			initContainerResult := kubecontainer.NewSyncResult(kubecontainer.InitContainer, status.Name)
+			initContainerResult.Fail(kubecontainer.ErrRunInitContainer, fmt.Sprintf("init container %q exited with %d", status.Name, status.ExitCode))
+			result.AddSyncResult(initContainerResult)
+			if pod.Spec.RestartPolicy == api.RestartPolicyNever {
+				utilruntime.HandleError(fmt.Errorf("error running pod %q init container %q, restart=Never: %+v", format.Pod(pod), status.Name, status))
+				return
+			}
+			utilruntime.HandleError(fmt.Errorf("Error running pod %q init container %q, restarting: %+v", format.Pod(pod), status.Name, status))
+		}
+	}
+
+	// Note: when configuring the pod's containers anything that can be configured by pointing
+	// to the namespace of the infra container should use namespaceMode.  This includes things like the net namespace
+	// and IPC namespace.  PID mode cannot point to another container right now.
+	// See createPodInfraContainer for infra container setup.
+	namespaceMode := fmt.Sprintf("container:%v", podInfraContainerID)
+	pidMode := getPidMode(pod)
+
+	if next != nil {
+		if len(containerChanges.ContainersToStart) == 0 {
+			glog.V(4).Infof("No containers to start, stopping at init container %+v in pod %v", next.Name, format.Pod(pod))
+			return
+		}
+
+		// If we need to start the next container, do so now then exit
+		container := next
+		startContainerResult := kubecontainer.NewSyncResult(kubecontainer.StartContainer, container.Name)
+		result.AddSyncResult(startContainerResult)
+
+		// containerChanges.StartInfraContainer causes the containers to be restarted for config reasons
+		if !containerChanges.StartInfraContainer {
+			isInBackOff, err, msg := dm.doBackOff(pod, container, podStatus, backOff)
+			if isInBackOff {
+				startContainerResult.Fail(err, msg)
+				glog.V(4).Infof("Backing Off restarting init container %+v in pod %v", container, format.Pod(pod))
+				return
+			}
+		}
+
+		glog.V(4).Infof("Creating init container %+v in pod %v", container, format.Pod(pod))
+		if err, msg := dm.tryContainerStart(container, pod, podStatus, pullSecrets, namespaceMode, pidMode, podIP); err != nil {
+			startContainerResult.Fail(err, msg)
+			utilruntime.HandleError(fmt.Errorf("container start failed: %v: %s", err, msg))
+			return
+		}
+
+		// Successfully started the container; clear the entry in the failure
+		glog.V(4).Infof("Completed init container %q for pod %q", container.Name, format.Pod(pod))
+		return
+	}
+	if !done {
+		// init container still running
+		glog.V(4).Infof("An init container is still running in pod %v", format.Pod(pod))
+		return
+	}
+	if containerChanges.InitFailed {
+		// init container still running
+		glog.V(4).Infof("Not all init containers have succeeded for pod %v", format.Pod(pod))
+		return
+	}
+
+	// Start regular containers
 	for idx := range containerChanges.ContainersToStart {
 		container := &pod.Spec.Containers[idx]
 		startContainerResult := kubecontainer.NewSyncResult(kubecontainer.StartContainer, container.Name)
 		result.AddSyncResult(startContainerResult)
 
 		// containerChanges.StartInfraContainer causes the containers to be restarted for config reasons
-		// ignore backoff
 		if !containerChanges.StartInfraContainer {
 			isInBackOff, err, msg := dm.doBackOff(pod, container, podStatus, backOff)
 			if isInBackOff {
@@ -1905,46 +2047,131 @@ func (dm *DockerManager) SyncPod(pod *api.Pod, _ api.PodStatus, podStatus *kubec
 				continue
 			}
 		}
+
 		glog.V(4).Infof("Creating container %+v in pod %v", container, format.Pod(pod))
-		err, msg := dm.imagePuller.PullImage(pod, container, pullSecrets)
+		if err, msg := dm.tryContainerStart(container, pod, podStatus, pullSecrets, namespaceMode, pidMode, podIP); err != nil {
-		if err != nil {
 			startContainerResult.Fail(err, msg)
+			utilruntime.HandleError(fmt.Errorf("container start failed: %v: %s", err, msg))
 			continue
 		}
-
-		if container.SecurityContext != nil && container.SecurityContext.RunAsNonRoot != nil && *container.SecurityContext.RunAsNonRoot {
-			err := dm.verifyNonRoot(container)
-			if err != nil {
-				startContainerResult.Fail(kubecontainer.ErrVerifyNonRoot, err.Error())
-				glog.Errorf("Error running pod %q container %q: %v", format.Pod(pod), container.Name, err)
-				continue
-			}
-		}
-		// For a new container, the RestartCount should be 0
-		restartCount := 0
-		containerStatus := podStatus.FindContainerStatusByName(container.Name)
-		if containerStatus != nil {
-			restartCount = containerStatus.RestartCount + 1
-		}
-
-		// TODO(dawnchen): Check RestartPolicy.DelaySeconds before restart a container
-		// Note: when configuring the pod's containers anything that can be configured by pointing
-		// to the namespace of the infra container should use namespaceMode.  This includes things like the net namespace
-		// and IPC namespace.  PID mode cannot point to another container right now.
-		// See createPodInfraContainer for infra container setup.
-		namespaceMode := fmt.Sprintf("container:%v", podInfraContainerID)
-		_, err = dm.runContainerInPod(pod, container, namespaceMode, namespaceMode, getPidMode(pod), podIP, restartCount)
-		if err != nil {
-			startContainerResult.Fail(kubecontainer.ErrRunContainer, err.Error())
-			// TODO(bburns) : Perhaps blacklist a container after N failures?
-			glog.Errorf("Error running pod %q container %q: %v", format.Pod(pod), container.Name, err)
-			continue
-		}
-		// Successfully started the container; clear the entry in the failure
 	}
 	return
 }
 
+// tryContainerStart attempts to pull and start the container, returning an error and a reason string if the start
+// was not successful.
+func (dm *DockerManager) tryContainerStart(container *api.Container, pod *api.Pod, podStatus *kubecontainer.PodStatus, pullSecrets []api.Secret, namespaceMode, pidMode, podIP string) (err error, reason string) {
+	err, msg := dm.imagePuller.PullImage(pod, container, pullSecrets)
+	if err != nil {
+		return err, msg
+	}
+
+	if container.SecurityContext != nil && container.SecurityContext.RunAsNonRoot != nil && *container.SecurityContext.RunAsNonRoot {
+		err := dm.verifyNonRoot(container)
+		if err != nil {
+			return kubecontainer.ErrVerifyNonRoot, err.Error()
+		}
+	}
+
+	// For a new container, the RestartCount should be 0
+	restartCount := 0
+	containerStatus := podStatus.FindContainerStatusByName(container.Name)
+	if containerStatus != nil {
+		restartCount = containerStatus.RestartCount + 1
+	}
+
+	// TODO(dawnchen): Check RestartPolicy.DelaySeconds before restart a container
+	_, err = dm.runContainerInPod(pod, container, namespaceMode, namespaceMode, pidMode, podIP, restartCount)
+	if err != nil {
+		// TODO(bburns) : Perhaps blacklist a container after N failures?
+		return kubecontainer.ErrRunContainer, err.Error()
+	}
+	return nil, ""
+}
+
+// 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 (dm *DockerManager) pruneInitContainersBeforeStart(pod *api.Pod, podStatus *kubecontainer.PodStatus, initContainersToKeep map[kubecontainer.DockerID]int) {
+	// only the last execution of an init container should be preserved, and only preserve it if it is in the
+	// list of init containers to keep.
+	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 !initContainerNames.Has(status.Name) || status.State != kubecontainer.ContainerStateExited {
+				continue
+			}
+			count++
+			// keep the first init container we see
+			if count == 1 {
+				continue
+			}
+			// if there is a reason to preserve the older container, do so
+			if _, ok := initContainersToKeep[kubecontainer.DockerID(status.ID.ID)]; ok {
+				continue
+			}
+
+			// prune all other init containers that match this container name
+			// TODO: we may not need aggressive pruning
+			glog.V(4).Infof("Removing init container %q instance %q %d", status.Name, status.ID.ID, count)
+			if err := dm.client.RemoveContainer(status.ID.ID, dockertypes.ContainerRemoveOptions{RemoveVolumes: true}); err != nil {
+				if _, ok := err.(containerNotFoundError); ok {
+					count--
+					continue
+				}
+				utilruntime.HandleError(fmt.Errorf("failed to remove pod init container %q: %v; Skipping pod %q", name, err, format.Pod(pod)))
+				// TODO: report serious errors
+				continue
+			}
+
+			// remove any references to this container
+			if _, ok := dm.containerRefManager.GetRef(status.ID); ok {
+				dm.containerRefManager.ClearRef(status.ID)
+			} else {
+				glog.Warningf("No ref for pod '%q'", pod.Name)
+			}
+		}
+	}
+}
+
+// findActiveInitContainer 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
+// failed, in which case next will point to the current container.
+func findActiveInitContainer(pod *api.Pod, podStatus *kubecontainer.PodStatus) (next *api.Container, status *kubecontainer.ContainerStatus, done bool) {
+	if len(pod.Spec.InitContainers) == 0 {
+		return nil, nil, true
+	}
+
+	for i := len(pod.Spec.InitContainers) - 1; i >= 0; i-- {
+		container := &pod.Spec.InitContainers[i]
+		status := podStatus.FindContainerStatusByName(container.Name)
+		switch {
+		case status == nil:
+			continue
+		case status.State == kubecontainer.ContainerStateRunning:
+			return nil, nil, false
+		case status.State == kubecontainer.ContainerStateExited:
+			switch {
+			// the container has failed, we'll have to retry
+			case status.ExitCode != 0:
+				return &pod.Spec.InitContainers[i], status, false
+			// all init containers successful
+			case i == (len(pod.Spec.InitContainers) - 1):
+				return nil, nil, true
+			// all containers up to i successful, go to i+1
+			default:
+				return &pod.Spec.InitContainers[i+1], nil, false
+			}
+		}
+	}
+
+	return &pod.Spec.InitContainers[0], nil, false
+}
+
 // verifyNonRoot returns an error if the container or image will run as the root user.
 func (dm *DockerManager) verifyNonRoot(container *api.Container) error {
 	if securitycontext.HasRunAsUser(container) {
@@ -2018,6 +2245,7 @@ func (dm *DockerManager) doBackOff(pod *api.Pod, container *api.Container, podSt
 		}
 	}
 	if cStatus != nil {
+		glog.Infof("checking backoff for container %q in pod %q", container.Name, pod.Name)
 		ts := cStatus.FinishedAt
 		// found a container that requires backoff
 		dockerName := KubeletContainerName{

pkg/kubelet/kubelet.go

@@ -1784,6 +1784,7 @@ func (kl *Kubelet) syncPod(o syncPodOptions) error {
 		!firstSeenTime.IsZero() {
 		metrics.PodStartLatency.Observe(metrics.SinceInMicroseconds(firstSeenTime))
 	}
+
 	// Update status in the status manager
 	kl.statusManager.SetPodStatus(pod, apiPodStatus)
 
@@ -2370,6 +2371,10 @@ func hasHostPortConflicts(pods []*api.Pod) bool {
 			glog.Errorf("Pod %q: HostPort is already allocated, ignoring: %v", format.Pod(pod), errs)
 			return true
 		}
+		if errs := validation.AccumulateUniqueHostPorts(pod.Spec.InitContainers, &ports, field.NewPath("spec", "initContainers")); len(errs) > 0 {
+			glog.Errorf("Pod %q: HostPort is already allocated, ignoring: %v", format.Pod(pod), errs)
+			return true
+		}
 	}
 	return false
 }
@@ -3392,12 +3397,46 @@ func (kl *Kubelet) tryUpdateNodeStatus() error {
 // This func is exported to simplify integration with 3rd party kubelet
 // integrations like kubernetes-mesos.
 func GetPhase(spec *api.PodSpec, info []api.ContainerStatus) api.PodPhase {
+	initialized := 0
+	pendingInitialization := 0
+	failedInitialization := 0
+	for _, container := range spec.InitContainers {
+		containerStatus, ok := api.GetContainerStatus(info, container.Name)
+		if !ok {
+			pendingInitialization++
+			continue
+		}
+
+		switch {
+		case containerStatus.State.Running != nil:
+			pendingInitialization++
+		case containerStatus.State.Terminated != nil:
+			if containerStatus.State.Terminated.ExitCode == 0 {
+				initialized++
+			} else {
+				failedInitialization++
+			}
+		case containerStatus.State.Waiting != nil:
+			if containerStatus.LastTerminationState.Terminated != nil {
+				if containerStatus.LastTerminationState.Terminated.ExitCode == 0 {
+					initialized++
+				} else {
+					failedInitialization++
+				}
+			} else {
+				pendingInitialization++
+			}
+		default:
+			pendingInitialization++
+		}
+	}
+
+	unknown := 0
 	running := 0
 	waiting := 0
 	stopped := 0
 	failed := 0
 	succeeded := 0
-	unknown := 0
 	for _, container := range spec.Containers {
 		containerStatus, ok := api.GetContainerStatus(info, container.Name)
 		if !ok {
@@ -3426,7 +3465,13 @@ func GetPhase(spec *api.PodSpec, info []api.ContainerStatus) api.PodPhase {
 		}
 	}
 
+	if failedInitialization > 0 && spec.RestartPolicy == api.RestartPolicyNever {
+		return api.PodFailed
+	}
+
 	switch {
+	case pendingInitialization > 0:
+		fallthrough
 	case waiting > 0:
 		glog.V(5).Infof("pod waiting > 0, pending")
 		// One or more containers has not been started
@@ -3491,8 +3536,10 @@ func (kl *Kubelet) generateAPIPodStatus(pod *api.Pod, podStatus *kubecontainer.P
 
 	// Assume info is ready to process
 	spec := &pod.Spec
-	s.Phase = GetPhase(spec, s.ContainerStatuses)
+	allStatus := append(append([]api.ContainerStatus{}, s.ContainerStatuses...), s.InitContainerStatuses...)
+	s.Phase = GetPhase(spec, allStatus)
 	kl.probeManager.UpdatePodStatus(pod.UID, s)
+	s.Conditions = append(s.Conditions, status.GeneratePodInitializedCondition(spec, s.InitContainerStatuses, s.Phase))
 	s.Conditions = append(s.Conditions, status.GeneratePodReadyCondition(spec, s.ContainerStatuses, s.Phase))
 	// s (the PodStatus we are creating) will not have a PodScheduled condition yet, because converStatusToAPIStatus()
 	// does not create one. If the existing PodStatus has a PodScheduled condition, then copy it into s and make sure
@@ -3525,9 +3572,27 @@ func (kl *Kubelet) generateAPIPodStatus(pod *api.Pod, podStatus *kubecontainer.P
 // alter the kubelet state at all.
 func (kl *Kubelet) convertStatusToAPIStatus(pod *api.Pod, podStatus *kubecontainer.PodStatus) *api.PodStatus {
 	var apiPodStatus api.PodStatus
-	uid := pod.UID
 	apiPodStatus.PodIP = podStatus.IP
 
+	apiPodStatus.ContainerStatuses = kl.convertToAPIContainerStatuses(
+		pod, podStatus,
+		pod.Status.ContainerStatuses,
+		pod.Spec.Containers,
+		len(pod.Spec.InitContainers) > 0,
+		false,
+	)
+	apiPodStatus.InitContainerStatuses = kl.convertToAPIContainerStatuses(
+		pod, podStatus,
+		pod.Status.InitContainerStatuses,
+		pod.Spec.InitContainers,
+		len(pod.Spec.InitContainers) > 0,
+		true,
+	)
+
+	return &apiPodStatus
+}
+
+func (kl *Kubelet) convertToAPIContainerStatuses(pod *api.Pod, podStatus *kubecontainer.PodStatus, previousStatus []api.ContainerStatus, containers []api.Container, hasInitContainers, isInitContainer bool) []api.ContainerStatus {
 	convertContainerStatus := func(cs *kubecontainer.ContainerStatus) *api.ContainerStatus {
 		cid := cs.ID.String()
 		status := &api.ContainerStatus{
@@ -3556,15 +3621,19 @@ func (kl *Kubelet) convertStatusToAPIStatus(pod *api.Pod, podStatus *kubecontain
 	}
 
 	// Fetch old containers statuses from old pod status.
-	oldStatuses := make(map[string]api.ContainerStatus, len(pod.Spec.Containers))
+	oldStatuses := make(map[string]api.ContainerStatus, len(containers))
-	for _, status := range pod.Status.ContainerStatuses {
+	for _, status := range previousStatus {
 		oldStatuses[status.Name] = status
 	}
 
 	// Set all container statuses to default waiting state
-	statuses := make(map[string]*api.ContainerStatus, len(pod.Spec.Containers))
+	statuses := make(map[string]*api.ContainerStatus, len(containers))
 	defaultWaitingState := api.ContainerState{Waiting: &api.ContainerStateWaiting{Reason: "ContainerCreating"}}
-	for _, container := range pod.Spec.Containers {
+	if hasInitContainers {
+		defaultWaitingState = api.ContainerState{Waiting: &api.ContainerStateWaiting{Reason: "PodInitializing"}}
+	}
+
+	for _, container := range containers {
 		status := &api.ContainerStatus{
 			Name:  container.Name,
 			Image: container.Image,
@@ -3580,7 +3649,6 @@ func (kl *Kubelet) convertStatusToAPIStatus(pod *api.Pod, podStatus *kubecontain
 
 	// Make the latest container status comes first.
 	sort.Sort(sort.Reverse(kubecontainer.SortContainerStatusesByCreationTime(podStatus.ContainerStatuses)))
-
 	// Set container statuses according to the statuses seen in pod status
 	containerSeen := map[string]int{}
 	for _, cStatus := range podStatus.ContainerStatuses {
@@ -3602,13 +3670,13 @@ func (kl *Kubelet) convertStatusToAPIStatus(pod *api.Pod, podStatus *kubecontain
 	}
 
 	// Handle the containers failed to be started, which should be in Waiting state.
-	for _, container := range pod.Spec.Containers {
+	for _, container := range containers {
 		// If a container should be restarted in next syncpod, it is *Waiting*.
 		if !kubecontainer.ShouldContainerBeRestarted(&container, pod, podStatus) {
 			continue
 		}
 		status := statuses[container.Name]
-		reason, message, ok := kl.reasonCache.Get(uid, container.Name)
+		reason, message, ok := kl.reasonCache.Get(pod.UID, container.Name)
 		if !ok {
 			// In fact, we could also apply Waiting state here, but it is less informative,
 			// and the container will be restarted soon, so we prefer the original state here.
@@ -3630,15 +3698,15 @@ func (kl *Kubelet) convertStatusToAPIStatus(pod *api.Pod, podStatus *kubecontain
 		statuses[container.Name] = status
 	}
 
-	apiPodStatus.ContainerStatuses = make([]api.ContainerStatus, 0)
+	var containerStatuses []api.ContainerStatus
 	for _, status := range statuses {
-		apiPodStatus.ContainerStatuses = append(apiPodStatus.ContainerStatuses, *status)
+		containerStatuses = append(containerStatuses, *status)
 	}
 
 	// Sort the container statuses since clients of this interface expect the list
 	// of containers in a pod has a deterministic order.
-	sort.Sort(kubetypes.SortedContainerStatuses(apiPodStatus.ContainerStatuses))
+	sort.Sort(kubetypes.SortedContainerStatuses(containerStatuses))
-	return &apiPodStatus
+	return containerStatuses
 }
 
 // Returns logs of current machine.

pkg/kubelet/prober/manager.go

@@ -207,6 +207,15 @@ func (m *manager) UpdatePodStatus(podUID types.UID, podStatus *api.PodStatus) {
 		}
 		podStatus.ContainerStatuses[i].Ready = ready
 	}
+	// init containers are ready if they have exited with success or if a readiness probe has
+	// succeeded.
+	for i, c := range podStatus.InitContainerStatuses {
+		var ready bool
+		if c.State.Terminated != nil && c.State.Terminated.ExitCode == 0 {
+			ready = true
+		}
+		podStatus.InitContainerStatuses[i].Ready = ready
+	}
 }
 
 func (m *manager) getWorker(podUID types.UID, containerName string, probeType probeType) (*worker, bool) {

pkg/kubelet/server/server.go

@@ -465,6 +465,13 @@ func (s *Server) getContainerLogs(request *restful.Request, response *restful.Re
 			containerExists = true
 		}
 	}
+	if !containerExists {
+		for _, container := range pod.Spec.InitContainers {
+			if container.Name == containerName {
+				containerExists = true
+			}
+		}
+	}
 	if !containerExists {
 		response.WriteError(http.StatusNotFound, fmt.Errorf("container %q not found in pod %q\n", containerName, podID))
 		return

pkg/kubelet/status/generate.go

@@ -77,3 +77,58 @@ func GeneratePodReadyCondition(spec *api.PodSpec, containerStatuses []api.Contai
 		Status: api.ConditionTrue,
 	}
 }
+
+// GeneratePodInitializedCondition returns initialized condition if all init containers in a pod are ready, else it
+// returns an uninitialized condition.
+func GeneratePodInitializedCondition(spec *api.PodSpec, containerStatuses []api.ContainerStatus, podPhase api.PodPhase) api.PodCondition {
+	// Find if all containers are ready or not.
+	if containerStatuses == nil && len(spec.InitContainers) > 0 {
+		return api.PodCondition{
+			Type:   api.PodInitialized,
+			Status: api.ConditionFalse,
+			Reason: "UnknownContainerStatuses",
+		}
+	}
+	unknownContainers := []string{}
+	unreadyContainers := []string{}
+	for _, container := range spec.InitContainers {
+		if containerStatus, ok := api.GetContainerStatus(containerStatuses, container.Name); ok {
+			if !containerStatus.Ready {
+				unreadyContainers = append(unreadyContainers, container.Name)
+			}
+		} else {
+			unknownContainers = append(unknownContainers, container.Name)
+		}
+	}
+
+	// If all init containers are known and succeeded, just return PodCompleted.
+	if podPhase == api.PodSucceeded && len(unknownContainers) == 0 {
+		return api.PodCondition{
+			Type:   api.PodInitialized,
+			Status: api.ConditionTrue,
+			Reason: "PodCompleted",
+		}
+	}
+
+	unreadyMessages := []string{}
+	if len(unknownContainers) > 0 {
+		unreadyMessages = append(unreadyMessages, fmt.Sprintf("containers with unknown status: %s", unknownContainers))
+	}
+	if len(unreadyContainers) > 0 {
+		unreadyMessages = append(unreadyMessages, fmt.Sprintf("containers with incomplete status: %s", unreadyContainers))
+	}
+	unreadyMessage := strings.Join(unreadyMessages, ", ")
+	if unreadyMessage != "" {
+		return api.PodCondition{
+			Type:    api.PodInitialized,
+			Status:  api.ConditionFalse,
+			Reason:  "ContainersNotInitialized",
+			Message: unreadyMessage,
+		}
+	}
+
+	return api.PodCondition{
+		Type:   api.PodInitialized,
+		Status: api.ConditionTrue,
+	}
+}

pkg/kubelet/status/manager.go

@@ -172,20 +172,14 @@ func (m *manager) SetContainerReadiness(podUID types.UID, containerID kubecontai
 	}
 
 	// Find the container to update.
-	containerIndex := -1
+	containerStatus, _, ok := findContainerStatus(&oldStatus.status, containerID.String())
-	for i, c := range oldStatus.status.ContainerStatuses {
+	if !ok {
-		if c.ContainerID == containerID.String() {
-			containerIndex = i
-			break
-		}
-	}
-	if containerIndex == -1 {
 		glog.Warningf("Container readiness changed for unknown container: %q - %q",
 			format.Pod(pod), containerID.String())
 		return
 	}
 
-	if oldStatus.status.ContainerStatuses[containerIndex].Ready == ready {
+	if containerStatus.Ready == ready {
 		glog.V(4).Infof("Container readiness unchanged (%v): %q - %q", ready,
 			format.Pod(pod), containerID.String())
 		return
@@ -196,7 +190,8 @@ func (m *manager) SetContainerReadiness(podUID types.UID, containerID kubecontai
 	if err != nil {
 		return
 	}
-	status.ContainerStatuses[containerIndex].Ready = ready
+	containerStatus, _, _ = findContainerStatus(&status, containerID.String())
+	containerStatus.Ready = ready
 
 	// Update pod condition.
 	readyConditionIndex := -1
@@ -217,6 +212,31 @@ func (m *manager) SetContainerReadiness(podUID types.UID, containerID kubecontai
 	m.updateStatusInternal(pod, status, false)
 }
 
+func findContainerStatus(status *api.PodStatus, containerID string) (containerStatus *api.ContainerStatus, init bool, ok bool) {
+	// Find the container to update.
+	containerIndex := -1
+	for i, c := range status.ContainerStatuses {
+		if c.ContainerID == containerID {
+			containerIndex = i
+			break
+		}
+	}
+	if containerIndex != -1 {
+		return &status.ContainerStatuses[containerIndex], false, true
+	}
+
+	for i, c := range status.InitContainerStatuses {
+		if c.ContainerID == containerID {
+			containerIndex = i
+			break
+		}
+	}
+	if containerIndex != -1 {
+		return &status.InitContainerStatuses[containerIndex], true, true
+	}
+	return nil, false, false
+}
+
 func (m *manager) TerminatePod(pod *api.Pod) {
 	m.podStatusesLock.Lock()
 	defer m.podStatusesLock.Unlock()
@@ -233,6 +253,11 @@ func (m *manager) TerminatePod(pod *api.Pod) {
 			Terminated: &api.ContainerStateTerminated{},
 		}
 	}
+	for i := range status.InitContainerStatuses {
+		status.InitContainerStatuses[i].State = api.ContainerState{
+			Terminated: &api.ContainerStateTerminated{},
+		}
+	}
 	m.updateStatusInternal(pod, pod.Status, true)
 }
 
@@ -251,16 +276,27 @@ func (m *manager) updateStatusInternal(pod *api.Pod, status api.PodStatus, force
 	}
 
 	// Set ReadyCondition.LastTransitionTime.
-	if readyCondition := api.GetPodReadyCondition(status); readyCondition != nil {
+	if _, readyCondition := api.GetPodCondition(&status, api.PodReady); readyCondition != nil {
 		// Need to set LastTransitionTime.
 		lastTransitionTime := unversioned.Now()
-		oldReadyCondition := api.GetPodReadyCondition(oldStatus)
+		_, oldReadyCondition := api.GetPodCondition(&oldStatus, api.PodReady)
 		if oldReadyCondition != nil && readyCondition.Status == oldReadyCondition.Status {
 			lastTransitionTime = oldReadyCondition.LastTransitionTime
 		}
 		readyCondition.LastTransitionTime = lastTransitionTime
 	}
 
+	// Set InitializedCondition.LastTransitionTime.
+	if _, initCondition := api.GetPodCondition(&status, api.PodInitialized); initCondition != nil {
+		// Need to set LastTransitionTime.
+		lastTransitionTime := unversioned.Now()
+		_, oldInitCondition := api.GetPodCondition(&oldStatus, api.PodInitialized)
+		if oldInitCondition != nil && initCondition.Status == oldInitCondition.Status {
+			lastTransitionTime = oldInitCondition.LastTransitionTime
+		}
+		initCondition.LastTransitionTime = lastTransitionTime
+	}
+
 	// ensure that the start time does not change across updates.
 	if oldStatus.StartTime != nil && !oldStatus.StartTime.IsZero() {
 		status.StartTime = oldStatus.StartTime
@@ -490,6 +526,8 @@ func normalizeStatus(status *api.PodStatus) *api.PodStatus {
 		normalizeTimeStamp(&condition.LastProbeTime)
 		normalizeTimeStamp(&condition.LastTransitionTime)
 	}
+
+	// update container statuses
 	for i := range status.ContainerStatuses {
 		cstatus := &status.ContainerStatuses[i]
 		normalizeContainerState(&cstatus.State)
@@ -497,6 +535,15 @@ func normalizeStatus(status *api.PodStatus) *api.PodStatus {
 	}
 	// Sort the container statuses, so that the order won't affect the result of comparison
 	sort.Sort(kubetypes.SortedContainerStatuses(status.ContainerStatuses))
+
+	// update init container statuses
+	for i := range status.InitContainerStatuses {
+		cstatus := &status.InitContainerStatuses[i]
+		normalizeContainerState(&cstatus.State)
+		normalizeContainerState(&cstatus.LastTerminationState)
+	}
+	// Sort the container statuses, so that the order won't affect the result of comparison
+	sort.Sort(kubetypes.SortedContainerStatuses(status.InitContainerStatuses))
 	return status
 }
 

pkg/kubelet/util.go

@@ -63,6 +63,11 @@ func canRunPod(pod *api.Pod) error {
 				return fmt.Errorf("pod with UID %q specified privileged container, but is disallowed", pod.UID)
 			}
 		}
+		for _, container := range pod.Spec.InitContainers {
+			if securitycontext.HasPrivilegedRequest(&container) {
+				return fmt.Errorf("pod with UID %q specified privileged container, but is disallowed", pod.UID)
+			}
+		}
 	}
 	return nil
 }

plugin/pkg/admission/limitranger/admission.go

@@ -208,45 +208,54 @@ func defaultContainerResourceRequirements(limitRange *api.LimitRange) api.Resour
 	return requirements
 }
 
+// mergeContainerResources handles defaulting all of the resources on a container.
+func mergeContainerResources(container *api.Container, defaultRequirements *api.ResourceRequirements, annotationPrefix string, annotations []string) []string {
+	setRequests := []string{}
+	setLimits := []string{}
+	if container.Resources.Limits == nil {
+		container.Resources.Limits = api.ResourceList{}
+	}
+	if container.Resources.Requests == nil {
+		container.Resources.Requests = api.ResourceList{}
+	}
+	for k, v := range defaultRequirements.Limits {
+		_, found := container.Resources.Limits[k]
+		if !found {
+			container.Resources.Limits[k] = *v.Copy()
+			setLimits = append(setLimits, string(k))
+		}
+	}
+	for k, v := range defaultRequirements.Requests {
+		_, found := container.Resources.Requests[k]
+		if !found {
+			container.Resources.Requests[k] = *v.Copy()
+			setRequests = append(setRequests, string(k))
+		}
+	}
+	if len(setRequests) > 0 {
+		sort.Strings(setRequests)
+		a := strings.Join(setRequests, ", ") + fmt.Sprintf(" request for %s %s", annotationPrefix, container.Name)
+		annotations = append(annotations, a)
+	}
+	if len(setLimits) > 0 {
+		sort.Strings(setLimits)
+		a := strings.Join(setLimits, ", ") + fmt.Sprintf(" limit for %s %s", annotationPrefix, container.Name)
+		annotations = append(annotations, a)
+	}
+	return annotations
+}
+
 // mergePodResourceRequirements merges enumerated requirements with default requirements
 // it annotates the pod with information about what requirements were modified
 func mergePodResourceRequirements(pod *api.Pod, defaultRequirements *api.ResourceRequirements) {
 	annotations := []string{}
 
 	for i := range pod.Spec.Containers {
-		container := &pod.Spec.Containers[i]
+		annotations = mergeContainerResources(&pod.Spec.Containers[i], defaultRequirements, "container", annotations)
-		setRequests := []string{}
+	}
-		setLimits := []string{}
+
-		if container.Resources.Limits == nil {
+	for i := range pod.Spec.InitContainers {
-			container.Resources.Limits = api.ResourceList{}
+		annotations = mergeContainerResources(&pod.Spec.InitContainers[i], defaultRequirements, "init container", annotations)
-		}
-		if container.Resources.Requests == nil {
-			container.Resources.Requests = api.ResourceList{}
-		}
-		for k, v := range defaultRequirements.Limits {
-			_, found := container.Resources.Limits[k]
-			if !found {
-				container.Resources.Limits[k] = *v.Copy()
-				setLimits = append(setLimits, string(k))
-			}
-		}
-		for k, v := range defaultRequirements.Requests {
-			_, found := container.Resources.Requests[k]
-			if !found {
-				container.Resources.Requests[k] = *v.Copy()
-				setRequests = append(setRequests, string(k))
-			}
-		}
-		if len(setRequests) > 0 {
-			sort.Strings(setRequests)
-			a := strings.Join(setRequests, ", ") + " request for container " + container.Name
-			annotations = append(annotations, a)
-		}
-		if len(setLimits) > 0 {
-			sort.Strings(setLimits)
-			a := strings.Join(setLimits, ", ") + " limit for container " + container.Name
-			annotations = append(annotations, a)
-		}
 	}
 
 	if len(annotations) > 0 {
@@ -441,7 +450,7 @@ func PodLimitFunc(limitRange *api.LimitRange, pod *api.Pod) error {
 			}
 		}
 
-		// enforce pod limits
+		// enforce pod limits on init containers
 		if limitType == api.LimitTypePod {
 			containerRequests, containerLimits := []api.ResourceList{}, []api.ResourceList{}
 			for j := range pod.Spec.Containers {
@@ -451,6 +460,28 @@ func PodLimitFunc(limitRange *api.LimitRange, pod *api.Pod) error {
 			}
 			podRequests := sum(containerRequests)
 			podLimits := sum(containerLimits)
+			for j := range pod.Spec.InitContainers {
+				container := &pod.Spec.InitContainers[j]
+				// take max(sum_containers, any_init_container)
+				for k, v := range container.Resources.Requests {
+					if v2, ok := podRequests[k]; ok {
+						if v.Cmp(v2) > 0 {
+							podRequests[k] = v
+						}
+					} else {
+						podRequests[k] = v
+					}
+				}
+				for k, v := range container.Resources.Limits {
+					if v2, ok := podLimits[k]; ok {
+						if v.Cmp(v2) > 0 {
+							podLimits[k] = v
+						}
+					} else {
+						podLimits[k] = v
+					}
+				}
+			}
 			for k, v := range limit.Min {
 				if err := minConstraint(limitType, k, v, podRequests, podLimits); err != nil {
 					errs = append(errs, err)

plugin/pkg/admission/limitranger/admission_test.go

@@ -134,6 +134,17 @@ func validPod(name string, numContainers int, resources api.ResourceRequirements
 	return pod
 }
 
+func validPodInit(pod api.Pod, resources ...api.ResourceRequirements) api.Pod {
+	for i := 0; i < len(resources); i++ {
+		pod.Spec.InitContainers = append(pod.Spec.InitContainers, api.Container{
+			Image:     "foo:V" + strconv.Itoa(i),
+			Resources: resources[i],
+			Name:      "foo-" + strconv.Itoa(i),
+		})
+	}
+	return pod
+}
+
 func TestDefaultContainerResourceRequirements(t *testing.T) {
 	limitRange := validLimitRange()
 	expected := api.ResourceRequirements{
@@ -183,7 +194,7 @@ func TestMergePodResourceRequirements(t *testing.T) {
 
 	// pod with some resources enumerated should only merge empty
 	input := getResourceRequirements(getResourceList("", "512Mi"), getResourceList("", ""))
-	pod = validPod("limit-memory", 1, input)
+	pod = validPodInit(validPod("limit-memory", 1, input), input)
 	expected = api.ResourceRequirements{
 		Requests: api.ResourceList{
 			api.ResourceCPU:    defaultRequirements.Requests[api.ResourceCPU],
@@ -198,11 +209,18 @@ func TestMergePodResourceRequirements(t *testing.T) {
 			t.Errorf("pod %v, expected != actual; %v != %v", pod.Name, expected, actual)
 		}
 	}
+	for i := range pod.Spec.InitContainers {
+		actual := pod.Spec.InitContainers[i].Resources
+		if !api.Semantic.DeepEqual(expected, actual) {
+			t.Errorf("pod %v, expected != actual; %v != %v", pod.Name, expected, actual)
+		}
+	}
 	verifyAnnotation(t, &pod, "LimitRanger plugin set: cpu request for container foo-0; cpu, memory limit for container foo-0")
 
 	// pod with all resources enumerated should not merge anything
 	input = getResourceRequirements(getResourceList("100m", "512Mi"), getResourceList("200m", "1G"))
-	pod = validPod("limit-memory", 1, input)
+	initInputs := []api.ResourceRequirements{getResourceRequirements(getResourceList("200m", "1G"), getResourceList("400m", "2G"))}
+	pod = validPodInit(validPod("limit-memory", 1, input), initInputs...)
 	expected = input
 	mergePodResourceRequirements(&pod, &defaultRequirements)
 	for i := range pod.Spec.Containers {
@@ -211,6 +229,12 @@ func TestMergePodResourceRequirements(t *testing.T) {
 			t.Errorf("pod %v, expected != actual; %v != %v", pod.Name, expected, actual)
 		}
 	}
+	for i := range pod.Spec.InitContainers {
+		actual := pod.Spec.InitContainers[i].Resources
+		if !api.Semantic.DeepEqual(initInputs[i], actual) {
+			t.Errorf("pod %v, expected != actual; %v != %v", pod.Name, initInputs[i], actual)
+		}
+	}
 	expectNoAnnotation(t, &pod)
 }
 
@@ -273,6 +297,20 @@ func TestPodLimitFunc(t *testing.T) {
 			pod:        validPod("pod-min-memory-request-limit", 2, getResourceRequirements(getResourceList("", "60Mi"), getResourceList("", "100Mi"))),
 			limitRange: createLimitRange(api.LimitTypePod, getResourceList("", "100Mi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
 		},
+		{
+			pod: validPodInit(
+				validPod("pod-init-min-memory-request", 2, getResourceRequirements(getResourceList("", "60Mi"), getResourceList("", ""))),
+				getResourceRequirements(getResourceList("", "100Mi"), getResourceList("", "")),
+			),
+			limitRange: createLimitRange(api.LimitTypePod, getResourceList("", "100Mi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
+		},
+		{
+			pod: validPodInit(
+				validPod("pod-init-min-memory-request-limit", 2, getResourceRequirements(getResourceList("", "60Mi"), getResourceList("", "100Mi"))),
+				getResourceRequirements(getResourceList("", "80Mi"), getResourceList("", "100Mi")),
+			),
+			limitRange: createLimitRange(api.LimitTypePod, getResourceList("", "100Mi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
+		},
 		{
 			pod:        validPod("pod-max-cpu-request-limit", 2, getResourceRequirements(getResourceList("500m", ""), getResourceList("1", ""))),
 			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("2", ""), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
@@ -281,6 +319,22 @@ func TestPodLimitFunc(t *testing.T) {
 			pod:        validPod("pod-max-cpu-limit", 2, getResourceRequirements(getResourceList("", ""), getResourceList("1", ""))),
 			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("2", ""), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
 		},
+		{
+			pod: validPodInit(
+				validPod("pod-init-max-cpu-request-limit", 2, getResourceRequirements(getResourceList("500m", ""), getResourceList("1", ""))),
+				getResourceRequirements(getResourceList("1", ""), getResourceList("2", "")),
+				getResourceRequirements(getResourceList("1", ""), getResourceList("1", "")),
+			),
+			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("2", ""), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
+		},
+		{
+			pod: validPodInit(
+				validPod("pod-init-max-cpu-limit", 2, getResourceRequirements(getResourceList("", ""), getResourceList("1", ""))),
+				getResourceRequirements(getResourceList("", ""), getResourceList("2", "")),
+				getResourceRequirements(getResourceList("", ""), getResourceList("2", "")),
+			),
+			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("2", ""), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
+		},
 		{
 			pod:        validPod("pod-max-mem-request-limit", 2, getResourceRequirements(getResourceList("", "250Mi"), getResourceList("", "500Mi"))),
 			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("", "1Gi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
@@ -387,6 +441,13 @@ func TestPodLimitFunc(t *testing.T) {
 			pod:        validPod("pod-max-mem-limit", 3, getResourceRequirements(getResourceList("", ""), getResourceList("", "500Mi"))),
 			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("", "1Gi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
 		},
+		{
+			pod: validPodInit(
+				validPod("pod-init-max-mem-limit", 1, getResourceRequirements(getResourceList("", ""), getResourceList("", "500Mi"))),
+				getResourceRequirements(getResourceList("", ""), getResourceList("", "1.5Gi")),
+			),
+			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("", "1Gi"), api.ResourceList{}, api.ResourceList{}, api.ResourceList{}),
+		},
 		{
 			pod:        validPod("pod-max-mem-ratio", 3, getResourceRequirements(getResourceList("", "250Mi"), getResourceList("", "500Mi"))),
 			limitRange: createLimitRange(api.LimitTypePod, api.ResourceList{}, getResourceList("", "2Gi"), api.ResourceList{}, api.ResourceList{}, getResourceList("", "1.5")),
@@ -403,7 +464,7 @@ func TestPodLimitFunc(t *testing.T) {
 
 func TestPodLimitFuncApplyDefault(t *testing.T) {
 	limitRange := validLimitRange()
-	testPod := validPod("foo", 1, getResourceRequirements(api.ResourceList{}, api.ResourceList{}))
+	testPod := validPodInit(validPod("foo", 1, getResourceRequirements(api.ResourceList{}, api.ResourceList{})), getResourceRequirements(api.ResourceList{}, api.ResourceList{}))
 	err := PodLimitFunc(&limitRange, &testPod)
 	if err != nil {
 		t.Errorf("Unexpected error for valid pod: %v, %v", testPod.Name, err)
@@ -429,6 +490,27 @@ func TestPodLimitFuncApplyDefault(t *testing.T) {
 			t.Errorf("Unexpected cpu value %s", requestCpu)
 		}
 	}
+
+	for i := range testPod.Spec.InitContainers {
+		container := testPod.Spec.InitContainers[i]
+		limitMemory := container.Resources.Limits.Memory().String()
+		limitCpu := container.Resources.Limits.Cpu().String()
+		requestMemory := container.Resources.Requests.Memory().String()
+		requestCpu := container.Resources.Requests.Cpu().String()
+
+		if limitMemory != "10Mi" {
+			t.Errorf("Unexpected memory value %s", limitMemory)
+		}
+		if limitCpu != "75m" {
+			t.Errorf("Unexpected cpu value %s", limitCpu)
+		}
+		if requestMemory != "5Mi" {
+			t.Errorf("Unexpected memory value %s", requestMemory)
+		}
+		if requestCpu != "50m" {
+			t.Errorf("Unexpected cpu value %s", requestCpu)
+		}
+	}
 }
 
 func TestLimitRangerIgnoresSubresource(t *testing.T) {

plugin/pkg/scheduler/algorithm/predicates/error.go

@@ -21,7 +21,7 @@ import "fmt"
 const (
 	podCountResourceName  string = "PodCount"
 	cpuResourceName       string = "CPU"
-	memoryResoureceName   string = "Memory"
+	memoryResourceName    string = "Memory"
 	nvidiaGpuResourceName string = "NvidiaGpu"
 )
 

plugin/pkg/scheduler/algorithm/predicates/predicates.go

@@ -359,6 +359,16 @@ func getResourceRequest(pod *api.Pod) resourceRequest {
 		result.milliCPU += requests.Cpu().MilliValue()
 		result.nvidiaGPU += requests.NvidiaGPU().Value()
 	}
+	// take max_resource(sum_pod, any_init_container)
+	for _, container := range pod.Spec.InitContainers {
+		requests := container.Resources.Requests
+		if mem := requests.Memory().Value(); mem > result.memory {
+			result.memory = mem
+		}
+		if cpu := requests.Cpu().MilliValue(); cpu > result.milliCPU {
+			result.milliCPU = cpu
+		}
+	}
 	return result
 }
 
@@ -428,7 +438,7 @@ func PodFitsResources(pod *api.Pod, nodeInfo *schedulercache.NodeInfo) (bool, er
 	}
 	if totalMemory < podRequest.memory+nodeInfo.RequestedResource().Memory {
 		return false,
-			newInsufficientResourceError(memoryResoureceName, podRequest.memory, nodeInfo.RequestedResource().Memory, totalMemory)
+			newInsufficientResourceError(memoryResourceName, podRequest.memory, nodeInfo.RequestedResource().Memory, totalMemory)
 	}
 	if totalNvidiaGPU < podRequest.nvidiaGPU+nodeInfo.RequestedResource().NvidiaGPU {
 		return false,

plugin/pkg/scheduler/algorithm/predicates/predicates_test.go

@@ -111,6 +111,11 @@ func newResourcePod(usage ...resourceRequest) *api.Pod {
 	}
 }
 
+func newResourceInitPod(pod *api.Pod, usage ...resourceRequest) *api.Pod {
+	pod.Spec.InitContainers = newResourcePod(usage...).Spec.Containers
+	return pod
+}
+
 func TestPodFitsResources(t *testing.T) {
 	enoughPodsTests := []struct {
 		pod      *api.Pod
@@ -135,6 +140,54 @@ func TestPodFitsResources(t *testing.T) {
 			test: "too many resources fails",
 			wErr: newInsufficientResourceError(cpuResourceName, 1, 10, 10),
 		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 3, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 8, memory: 19})),
+			fits: false,
+			test: "too many resources fails due to init container cpu",
+			wErr: newInsufficientResourceError(cpuResourceName, 3, 8, 10),
+		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 3, memory: 1}, resourceRequest{milliCPU: 2, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 8, memory: 19})),
+			fits: false,
+			test: "too many resources fails due to highest init container cpu",
+			wErr: newInsufficientResourceError(cpuResourceName, 3, 8, 10),
+		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 1, memory: 3}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
+			fits: false,
+			test: "too many resources fails due to init container memory",
+			wErr: newInsufficientResourceError(memoryResourceName, 3, 19, 20),
+		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 1, memory: 3}, resourceRequest{milliCPU: 1, memory: 2}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
+			fits: false,
+			test: "too many resources fails due to highest init container memory",
+			wErr: newInsufficientResourceError(memoryResourceName, 3, 19, 20),
+		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 1, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
+			fits: true,
+			test: "init container fits because it's the max, not sum, of containers and init containers",
+			wErr: nil,
+		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 1, memory: 1}), resourceRequest{milliCPU: 1, memory: 1}, resourceRequest{milliCPU: 1, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
+			fits: true,
+			test: "multiple init containers fit because it's the max, not sum, of containers and init containers",
+			wErr: nil,
+		},
 		{
 			pod: newResourcePod(resourceRequest{milliCPU: 1, memory: 1}),
 			nodeInfo: schedulercache.NewNodeInfo(
@@ -149,7 +202,7 @@ func TestPodFitsResources(t *testing.T) {
 				newResourcePod(resourceRequest{milliCPU: 5, memory: 19})),
 			fits: false,
 			test: "one resources fits",
-			wErr: newInsufficientResourceError(memoryResoureceName, 2, 19, 20),
+			wErr: newInsufficientResourceError(memoryResourceName, 2, 19, 20),
 		},
 		{
 			pod: newResourcePod(resourceRequest{milliCPU: 5, memory: 1}),
@@ -159,6 +212,14 @@ func TestPodFitsResources(t *testing.T) {
 			test: "equal edge case",
 			wErr: nil,
 		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 4, memory: 1}), resourceRequest{milliCPU: 5, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 5, memory: 19})),
+			fits: true,
+			test: "equal edge case for init container",
+			wErr: nil,
+		},
 	}
 
 	for _, test := range enoughPodsTests {
@@ -205,6 +266,14 @@ func TestPodFitsResources(t *testing.T) {
 			test: "even for equal edge case predicate fails when there's no space for additional pod",
 			wErr: newInsufficientResourceError(podCountResourceName, 1, 1, 1),
 		},
+		{
+			pod: newResourceInitPod(newResourcePod(resourceRequest{milliCPU: 5, memory: 1}), resourceRequest{milliCPU: 5, memory: 1}),
+			nodeInfo: schedulercache.NewNodeInfo(
+				newResourcePod(resourceRequest{milliCPU: 5, memory: 19})),
+			fits: false,
+			test: "even for equal edge case predicate fails when there's no space for additional pod due to init container",
+			wErr: newInsufficientResourceError(podCountResourceName, 1, 1, 1),
+		},
 	}
 	for _, test := range notEnoughPodsTests {
 		node := api.Node{Status: api.NodeStatus{Capacity: api.ResourceList{}, Allocatable: makeAllocatableResources(10, 20, 0, 1)}}

test/e2e/framework/util.go

@@ -879,6 +879,109 @@ func deleteNS(c *client.Client, namespace string, timeout time.Duration) error {
 	return nil
 }
 
+func ContainerInitInvariant(older, newer runtime.Object) error {
+	oldPod := older.(*api.Pod)
+	newPod := newer.(*api.Pod)
+	if len(oldPod.Spec.InitContainers) == 0 {
+		return nil
+	}
+	if len(oldPod.Spec.InitContainers) != len(newPod.Spec.InitContainers) {
+		return fmt.Errorf("init container list changed")
+	}
+	if oldPod.UID != newPod.UID {
+		return fmt.Errorf("two different pods exist in the condition: %s vs %s", oldPod.UID, newPod.UID)
+	}
+	if err := initContainersInvariants(oldPod); err != nil {
+		return err
+	}
+	if err := initContainersInvariants(newPod); err != nil {
+		return err
+	}
+	oldInit, _, _ := podInitialized(oldPod)
+	newInit, _, _ := podInitialized(newPod)
+	if oldInit && !newInit {
+		// TODO: we may in the future enable resetting PodInitialized = false if the kubelet needs to restart it
+		// from scratch
+		return fmt.Errorf("pod cannot be initialized and then regress to not being initialized")
+	}
+	return nil
+}
+
+func podInitialized(pod *api.Pod) (ok bool, failed bool, err error) {
+	allInit := true
+	initFailed := false
+	for _, s := range pod.Status.InitContainerStatuses {
+		switch {
+		case initFailed && s.State.Waiting == nil:
+			return allInit, initFailed, fmt.Errorf("container %s is after a failed container but isn't waiting", s.Name)
+		case allInit && s.State.Waiting == nil:
+			return allInit, initFailed, fmt.Errorf("container %s is after an initializing container but isn't waiting", s.Name)
+		case s.State.Terminated == nil:
+			allInit = false
+		case s.State.Terminated.ExitCode != 0:
+			allInit = false
+			initFailed = true
+		case !s.Ready:
+			return allInit, initFailed, fmt.Errorf("container %s initialized but isn't marked as ready", s.Name)
+		}
+	}
+	return allInit, initFailed, nil
+}
+
+func initContainersInvariants(pod *api.Pod) error {
+	allInit, initFailed, err := podInitialized(pod)
+	if err != nil {
+		return err
+	}
+	if !allInit || initFailed {
+		for _, s := range pod.Status.ContainerStatuses {
+			if s.State.Waiting == nil || s.RestartCount != 0 {
+				return fmt.Errorf("container %s is not waiting but initialization not complete", s.Name)
+			}
+			if s.State.Waiting.Reason != "PodInitializing" {
+				return fmt.Errorf("container %s should have reason PodInitializing: %s", s.Name, s.State.Waiting.Reason)
+			}
+		}
+	}
+	_, c := api.GetPodCondition(&pod.Status, api.PodInitialized)
+	if c == nil {
+		return fmt.Errorf("pod does not have initialized condition")
+	}
+	if c.LastTransitionTime.IsZero() {
+		return fmt.Errorf("PodInitialized condition should always have a transition time")
+	}
+	switch {
+	case c.Status == api.ConditionUnknown:
+		return fmt.Errorf("PodInitialized condition should never be Unknown")
+	case c.Status == api.ConditionTrue && (initFailed || !allInit):
+		return fmt.Errorf("PodInitialized condition was True but all not all containers initialized")
+	case c.Status == api.ConditionFalse && (!initFailed && allInit):
+		return fmt.Errorf("PodInitialized condition was False but all containers initialized")
+	}
+	return nil
+}
+
+type InvariantFunc func(older, newer runtime.Object) error
+
+func CheckInvariants(events []watch.Event, fns ...InvariantFunc) error {
+	errs := sets.NewString()
+	for i := range events {
+		j := i + 1
+		if j >= len(events) {
+			continue
+		}
+		for _, fn := range fns {
+			if err := fn(events[i].Object, events[j].Object); err != nil {
+				errs.Insert(err.Error())
+			}
+		}
+	}
+	if errs.Len() > 0 {
+		return fmt.Errorf("invariants violated:\n* %s", strings.Join(errs.List(), "\n* "))
+	}
+	return nil
+}
+
 // Waits default amount of time (PodStartTimeout) for the specified pod to become running.
 // Returns an error if timeout occurs first, or pod goes in to failed state.
 func WaitForPodRunningInNamespace(c *client.Client, podName string, namespace string) error {
@@ -2218,7 +2321,11 @@ func DumpNodeDebugInfo(c *client.Client, nodeNames []string) {
 			continue
 		}
 		for _, p := range podList.Items {
-			Logf("%v started at %v (%d container statuses recorded)", p.Name, p.Status.StartTime, len(p.Status.ContainerStatuses))
+			Logf("%v started at %v (%d+%d container statuses recorded)", p.Name, p.Status.StartTime, len(p.Status.InitContainerStatuses), len(p.Status.ContainerStatuses))
+			for _, c := range p.Status.InitContainerStatuses {
+				Logf("\tInit container %v ready: %v, restart count %v",
+					c.Name, c.Ready, c.RestartCount)
+			}
 			for _, c := range p.Status.ContainerStatuses {
 				Logf("\tContainer %v ready: %v, restart count %v",
 					c.Name, c.Ready, c.RestartCount)

test/e2e/pods.go

@@ -659,6 +659,364 @@ var _ = framework.KubeDescribe("Pods", func() {
 		})
 	})
 
+	It("should invoke init containers on a RestartNever pod", func() {
+		podClient := f.Client.Pods(f.Namespace.Name)
+
+		By("creating the pod")
+		name := "pod-init-" + string(util.NewUUID())
+		value := strconv.Itoa(time.Now().Nanosecond())
+		pod := &api.Pod{
+			ObjectMeta: api.ObjectMeta{
+				Name: name,
+				Labels: map[string]string{
+					"name": "foo",
+					"time": value,
+				},
+			},
+			Spec: api.PodSpec{
+				RestartPolicy: api.RestartPolicyNever,
+				InitContainers: []api.Container{
+					{
+						Name:    "init1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+					{
+						Name:    "init2",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+				},
+				Containers: []api.Container{
+					{
+						Name:    "run1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+				},
+			},
+		}
+		defer podClient.Delete(pod.Name, nil)
+		startedPod, err := podClient.Create(pod)
+		if err != nil {
+			framework.Failf("Error creating a pod: %v", err)
+		}
+		w, err := podClient.Watch(api.SingleObject(startedPod.ObjectMeta))
+		if err != nil {
+			framework.Failf("Error watching a pod: %v", err)
+		}
+		wr := watch.NewRecorder(w)
+		event, err := watch.Until(framework.PodStartTimeout, wr, client.PodCompleted)
+		Expect(err).To(BeNil())
+		framework.CheckInvariants(wr.Events(), framework.ContainerInitInvariant)
+		endPod := event.Object.(*api.Pod)
+
+		Expect(endPod.Status.Phase).To(Equal(api.PodSucceeded))
+		_, init := api.GetPodCondition(&endPod.Status, api.PodInitialized)
+		Expect(init).NotTo(BeNil())
+		Expect(init.Status).To(Equal(api.ConditionTrue))
+
+		Expect(len(endPod.Status.InitContainerStatuses)).To(Equal(2))
+		for _, status := range endPod.Status.InitContainerStatuses {
+			Expect(status.Ready).To(BeTrue())
+			Expect(status.State.Terminated).NotTo(BeNil())
+			Expect(status.State.Terminated.ExitCode).To(BeZero())
+		}
+	})
+
+	It("should invoke init containers on a RestartAlways pod", func() {
+		podClient := f.Client.Pods(f.Namespace.Name)
+
+		By("creating the pod")
+		name := "pod-init-" + string(util.NewUUID())
+		value := strconv.Itoa(time.Now().Nanosecond())
+		pod := &api.Pod{
+			ObjectMeta: api.ObjectMeta{
+				Name: name,
+				Labels: map[string]string{
+					"name": "foo",
+					"time": value,
+				},
+			},
+			Spec: api.PodSpec{
+				InitContainers: []api.Container{
+					{
+						Name:    "init1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+					{
+						Name:    "init2",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+				},
+				Containers: []api.Container{
+					{
+						Name:  "run1",
+						Image: "gcr.io/google_containers/pause:2.0",
+						Resources: api.ResourceRequirements{
+							Limits: api.ResourceList{
+								api.ResourceCPU:    *resource.NewMilliQuantity(100, resource.DecimalSI),
+								api.ResourceMemory: *resource.NewQuantity(10*1024*1024, resource.DecimalSI),
+							},
+						},
+					},
+				},
+			},
+		}
+		defer podClient.Delete(pod.Name, nil)
+		startedPod, err := podClient.Create(pod)
+		if err != nil {
+			framework.Failf("Error creating a pod: %v", err)
+		}
+		w, err := podClient.Watch(api.SingleObject(startedPod.ObjectMeta))
+		if err != nil {
+			framework.Failf("Error watching a pod: %v", err)
+		}
+		wr := watch.NewRecorder(w)
+		event, err := watch.Until(framework.PodStartTimeout, wr, client.PodRunning)
+		Expect(err).To(BeNil())
+		framework.CheckInvariants(wr.Events(), framework.ContainerInitInvariant)
+		endPod := event.Object.(*api.Pod)
+
+		Expect(endPod.Status.Phase).To(Equal(api.PodRunning))
+		_, init := api.GetPodCondition(&endPod.Status, api.PodInitialized)
+		Expect(init).NotTo(BeNil())
+		Expect(init.Status).To(Equal(api.ConditionTrue))
+
+		Expect(len(endPod.Status.InitContainerStatuses)).To(Equal(2))
+		for _, status := range endPod.Status.InitContainerStatuses {
+			Expect(status.Ready).To(BeTrue())
+			Expect(status.State.Terminated).NotTo(BeNil())
+			Expect(status.State.Terminated.ExitCode).To(BeZero())
+		}
+	})
+
+	It("should not start app containers if init containers fail on a RestartAlways pod", func() {
+		podClient := f.Client.Pods(f.Namespace.Name)
+
+		By("creating the pod")
+		name := "pod-init-" + string(util.NewUUID())
+		value := strconv.Itoa(time.Now().Nanosecond())
+		pod := &api.Pod{
+			ObjectMeta: api.ObjectMeta{
+				Name: name,
+				Labels: map[string]string{
+					"name": "foo",
+					"time": value,
+				},
+			},
+			Spec: api.PodSpec{
+				InitContainers: []api.Container{
+					{
+						Name:    "init1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/false"},
+					},
+					{
+						Name:    "init2",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+				},
+				Containers: []api.Container{
+					{
+						Name:  "run1",
+						Image: "gcr.io/google_containers/pause:2.0",
+						Resources: api.ResourceRequirements{
+							Limits: api.ResourceList{
+								api.ResourceCPU:    *resource.NewMilliQuantity(100, resource.DecimalSI),
+								api.ResourceMemory: *resource.NewQuantity(10*1024*1024, resource.DecimalSI),
+							},
+						},
+					},
+				},
+			},
+		}
+		defer podClient.Delete(pod.Name, nil)
+		startedPod, err := podClient.Create(pod)
+		if err != nil {
+			framework.Failf("Error creating a pod: %v", err)
+		}
+		w, err := podClient.Watch(api.SingleObject(startedPod.ObjectMeta))
+		if err != nil {
+			framework.Failf("Error watching a pod: %v", err)
+		}
+
+		wr := watch.NewRecorder(w)
+		event, err := watch.Until(
+			framework.PodStartTimeout, wr,
+			// check for the first container to fail at least once
+			func(evt watch.Event) (bool, error) {
+				switch t := evt.Object.(type) {
+				case *api.Pod:
+					for _, status := range t.Status.ContainerStatuses {
+						if status.State.Waiting == nil {
+							return false, fmt.Errorf("container %q should not be out of waiting: %#v", status.Name, status)
+						}
+						if status.State.Waiting.Reason != "PodInitializing" {
+							return false, fmt.Errorf("container %q should have reason PodInitializing: %#v", status.Name, status)
+						}
+					}
+					if len(t.Status.InitContainerStatuses) != 2 {
+						return false, nil
+					}
+					status := t.Status.InitContainerStatuses[1]
+					if status.State.Waiting == nil {
+						return false, fmt.Errorf("second init container should not be out of waiting: %#v", status)
+					}
+					if status.State.Waiting.Reason != "PodInitializing" {
+						return false, fmt.Errorf("second init container should have reason PodInitializing: %#v", status)
+					}
+					status = t.Status.InitContainerStatuses[0]
+					if status.State.Terminated != nil && status.State.Terminated.ExitCode == 0 {
+						return false, fmt.Errorf("first init container should have exitCode != 0: %#v", status)
+					}
+					// continue until we see an attempt to restart the pod
+					return status.LastTerminationState.Terminated != nil, nil
+				default:
+					return false, fmt.Errorf("unexpected object: %#v", t)
+				}
+			},
+			// verify we get two restarts
+			func(evt watch.Event) (bool, error) {
+				switch t := evt.Object.(type) {
+				case *api.Pod:
+					status := t.Status.InitContainerStatuses[0]
+					if status.RestartCount < 3 {
+						return false, nil
+					}
+					framework.Logf("init container has failed twice: %#v", t)
+					// TODO: more conditions
+					return true, nil
+				default:
+					return false, fmt.Errorf("unexpected object: %#v", t)
+				}
+			},
+		)
+		Expect(err).To(BeNil())
+		framework.CheckInvariants(wr.Events(), framework.ContainerInitInvariant)
+		endPod := event.Object.(*api.Pod)
+
+		Expect(endPod.Status.Phase).To(Equal(api.PodPending))
+		_, init := api.GetPodCondition(&endPod.Status, api.PodInitialized)
+		Expect(init).NotTo(BeNil())
+		Expect(init.Status).To(Equal(api.ConditionFalse))
+		Expect(init.Reason).To(Equal("ContainersNotInitialized"))
+		Expect(init.Message).To(Equal("containers with incomplete status: [init1 init2]"))
+		Expect(len(endPod.Status.InitContainerStatuses)).To(Equal(2))
+	})
+
+	It("should not start app containers and fail the pod if init containers fail on a RestartNever pod", func() {
+		podClient := f.Client.Pods(f.Namespace.Name)
+
+		By("creating the pod")
+		name := "pod-init-" + string(util.NewUUID())
+		value := strconv.Itoa(time.Now().Nanosecond())
+		pod := &api.Pod{
+			ObjectMeta: api.ObjectMeta{
+				Name: name,
+				Labels: map[string]string{
+					"name": "foo",
+					"time": value,
+				},
+			},
+			Spec: api.PodSpec{
+				RestartPolicy: api.RestartPolicyNever,
+				InitContainers: []api.Container{
+					{
+						Name:    "init1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+					},
+					{
+						Name:    "init2",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/false"},
+					},
+				},
+				Containers: []api.Container{
+					{
+						Name:    "run1",
+						Image:   "gcr.io/google_containers/busybox:1.24",
+						Command: []string{"/bin/true"},
+						Resources: api.ResourceRequirements{
+							Limits: api.ResourceList{
+								api.ResourceCPU:    *resource.NewMilliQuantity(100, resource.DecimalSI),
+								api.ResourceMemory: *resource.NewQuantity(10*1024*1024, resource.DecimalSI),
+							},
+						},
+					},
+				},
+			},
+		}
+		defer podClient.Delete(pod.Name, nil)
+		startedPod, err := podClient.Create(pod)
+		if err != nil {
+			framework.Failf("Error creating a pod: %v", err)
+		}
+		w, err := podClient.Watch(api.SingleObject(startedPod.ObjectMeta))
+		if err != nil {
+			framework.Failf("Error watching a pod: %v", err)
+		}
+
+		wr := watch.NewRecorder(w)
+		event, err := watch.Until(
+			framework.PodStartTimeout, wr,
+			// check for the second container to fail at least once
+			func(evt watch.Event) (bool, error) {
+				switch t := evt.Object.(type) {
+				case *api.Pod:
+					for _, status := range t.Status.ContainerStatuses {
+						if status.State.Waiting == nil {
+							return false, fmt.Errorf("container %q should not be out of waiting: %#v", status.Name, status)
+						}
+						if status.State.Waiting.Reason != "PodInitializing" {
+							return false, fmt.Errorf("container %q should have reason PodInitializing: %#v", status.Name, status)
+						}
+					}
+					if len(t.Status.InitContainerStatuses) != 2 {
+						return false, nil
+					}
+					status := t.Status.InitContainerStatuses[0]
+					if status.State.Terminated == nil {
+						if status.State.Waiting != nil && status.State.Waiting.Reason != "PodInitializing" {
+							return false, fmt.Errorf("second init container should have reason PodInitializing: %#v", status)
+						}
+						return false, nil
+					}
+					if status.State.Terminated != nil && status.State.Terminated.ExitCode != 0 {
+						return false, fmt.Errorf("first init container should have exitCode != 0: %#v", status)
+					}
+					status = t.Status.InitContainerStatuses[1]
+					if status.State.Terminated == nil {
+						return false, nil
+					}
+					if status.State.Terminated.ExitCode == 0 {
+						return false, fmt.Errorf("second init container should have failed: %#v", status)
+					}
+					return true, nil
+				default:
+					return false, fmt.Errorf("unexpected object: %#v", t)
+				}
+			},
+			client.PodCompleted,
+		)
+		Expect(err).To(BeNil())
+		framework.CheckInvariants(wr.Events(), framework.ContainerInitInvariant)
+		endPod := event.Object.(*api.Pod)
+
+		Expect(endPod.Status.Phase).To(Equal(api.PodFailed))
+		_, init := api.GetPodCondition(&endPod.Status, api.PodInitialized)
+		Expect(init).NotTo(BeNil())
+		Expect(init.Status).To(Equal(api.ConditionFalse))
+		Expect(init.Reason).To(Equal("ContainersNotInitialized"))
+		Expect(init.Message).To(Equal("containers with incomplete status: [init2]"))
+		Expect(len(endPod.Status.InitContainerStatuses)).To(Equal(2))
+		Expect(endPod.Status.ContainerStatuses[0].State.Waiting).ToNot(BeNil())
+	})
+
 	It("should be restarted with a docker exec \"cat /tmp/health\" liveness probe [Conformance]", func() {
 		runLivenessTest(f.Client, f.Namespace.Name, &api.Pod{
 			ObjectMeta: api.ObjectMeta{