// Handle processes deployment and either creates a deployer pod or responds
// to a terminal deployment status. Since this controller started using caches,
// the provided rc MUST be deep-copied beforehand (see work() in factory.go).
func (c *DeploymentController) Handle(deployment *kapi.ReplicationController) error {
currentStatus := deployutil.DeploymentStatusFor(deployment)
nextStatus := currentStatus
deploymentScaled := false
deployerPodName := deployutil.DeployerPodNameForDeployment(deployment.Name)
deployer, deployerErr := c.getPod(deployment.Namespace, deployerPodName)
switch currentStatus {
case deployapi.DeploymentStatusNew:
// If the deployment has been cancelled, don't create a deployer pod.
// Instead try to delete any deployer pods found and transition the
// deployment to Pending so that the deployment config controller
// continues to see the deployment as in-flight. Eventually the deletion
// of the deployer pod should cause a requeue of this deployment and
// then it can be transitioned to Failed by this controller.
if deployutil.IsDeploymentCancelled(deployment) {
nextStatus = deployapi.DeploymentStatusPending
if err := c.cleanupDeployerPods(deployment); err != nil {
return err
}
break
}
// If the pod already exists, it's possible that a previous CreatePod
// succeeded but the deployment state update failed and now we're re-
// entering. Ensure that the pod is the one we created by verifying the
// annotation on it, and throw a retryable error.
if deployerErr != nil && !kerrors.IsNotFound(deployerErr) {
return fmt.Errorf("couldn't fetch existing deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), deployerErr)
}
if deployerErr == nil && deployer != nil {
// Do a stronger check to validate that the existing deployer pod is
// actually for this deployment, and if not, fail this deployment.
//
// TODO: Investigate checking the container image of the running pod and
// comparing with the intended deployer pod image. If we do so, we'll need
// to ensure that changes to 'unrelated' pods don't result in updates to
// the deployment. So, the image check will have to be done in other areas
// of the code as well.
if deployutil.DeploymentNameFor(deployer) != deployment.Name {
nextStatus = deployapi.DeploymentStatusFailed
deployment.Annotations[deployapi.DeploymentStatusReasonAnnotation] = deployapi.DeploymentFailedUnrelatedDeploymentExists
c.emitDeploymentEvent(deployment, kapi.EventTypeWarning, "FailedCreate", fmt.Sprintf("Error creating deployer pod since another pod with the same name (%q) exists", deployer.Name))
glog.V(2).Infof("Couldn't create deployer pod for %s since an unrelated pod with the same name (%q) exists", deployutil.LabelForDeployment(deployment), deployer.Name)
} else {
// Update to pending relative to the existing validated deployer pod.
deployment.Annotations[deployapi.DeploymentPodAnnotation] = deployer.Name
nextStatus = deployapi.DeploymentStatusPending
glog.V(4).Infof("Detected existing deployer pod %s for deployment %s", deployer.Name, deployutil.LabelForDeployment(deployment))
}
// Don't try and re-create the deployer pod.
break
}
if _, ok := deployment.Annotations[deployapi.DeploymentIgnorePodAnnotation]; ok {
return nil
}
// Generate a deployer pod spec.
deployerPod, err := c.makeDeployerPod(deployment)
if err != nil {
return fatalError(fmt.Sprintf("couldn't make deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err))
}
// Create the deployer pod.
deploymentPod, err := c.pn.Pods(deployment.Namespace).Create(deployerPod)
// Retry on error.
if err != nil {
return actionableError(fmt.Sprintf("couldn't create deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err))
}
deployment.Annotations[deployapi.DeploymentPodAnnotation] = deploymentPod.Name
nextStatus = deployapi.DeploymentStatusPending
glog.V(4).Infof("Created deployer pod %s for deployment %s", deploymentPod.Name, deployutil.LabelForDeployment(deployment))
case deployapi.DeploymentStatusPending, deployapi.DeploymentStatusRunning:
switch {
case kerrors.IsNotFound(deployerErr):
nextStatus = deployapi.DeploymentStatusFailed
// If the deployment is cancelled here then we deleted the deployer in a previous
// resync of the deployment.
if !deployutil.IsDeploymentCancelled(deployment) {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
deployment.Annotations[deployapi.DeploymentStatusReasonAnnotation] = deployapi.DeploymentFailedDeployerPodNoLongerExists
c.emitDeploymentEvent(deployment, kapi.EventTypeWarning, "Failed", fmt.Sprintf("Deployer pod %q has gone missing", deployerPodName))
glog.V(4).Infof("Failing deployment %q because its deployer pod %q disappeared", deployutil.LabelForDeployment(deployment), deployerPodName)
}
case deployerErr != nil:
// We'll try again later on resync. Continue to process cancellations.
glog.V(4).Infof("Error getting deployer pod %s for deployment %s: %v", deployerPodName, deployutil.LabelForDeployment(deployment), deployerErr)
default: /* err == nil */
// If the deployment has been cancelled, delete any deployer pods
// found and transition the deployment to Pending so that the
// deployment config controller continues to see the deployment
// as in-flight. Eventually the deletion of the deployer pod should
// cause a requeue of this deployment and then it can be transitioned
// to Failed by this controller.
if deployutil.IsDeploymentCancelled(deployment) {
if err := c.cleanupDeployerPods(deployment); err != nil {
return err
}
}
}
case deployapi.DeploymentStatusFailed:
// Check for test deployment and ensure the deployment scale matches
if config, err := deployutil.DecodeDeploymentConfig(deployment, c.codec); err == nil && config.Spec.Test {
deploymentScaled = deployment.Spec.Replicas != 0
deployment.Spec.Replicas = 0
}
// Try to cleanup once more a cancelled deployment in case hook pods
// were created just after we issued the first cleanup request.
if deployutil.IsDeploymentCancelled(deployment) {
if err := c.cleanupDeployerPods(deployment); err != nil {
return err
}
}
case deployapi.DeploymentStatusComplete:
// Check for test deployment and ensure the deployment scale matches
if config, err := deployutil.DecodeDeploymentConfig(deployment, c.codec); err == nil && config.Spec.Test {
deploymentScaled = deployment.Spec.Replicas != 0
deployment.Spec.Replicas = 0
}
if err := c.cleanupDeployerPods(deployment); err != nil {
return err
}
}
if deployutil.CanTransitionPhase(currentStatus, nextStatus) || deploymentScaled {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
if _, err := c.rn.ReplicationControllers(deployment.Namespace).Update(deployment); err != nil {
return fmt.Errorf("couldn't update deployment %s to status %s: %v", deployutil.LabelForDeployment(deployment), nextStatus, err)
}
glog.V(4).Infof("Updated deployment %s status from %s to %s (scale: %d)", deployutil.LabelForDeployment(deployment), currentStatus, nextStatus, deployment.Spec.Replicas)
}
return nil
}
// Handle processes deployment and either creates a deployer pod or responds
// to a terminal deployment status.
func (c *DeploymentController) Handle(deployment *kapi.ReplicationController) error {
currentStatus := deployutil.DeploymentStatusFor(deployment)
nextStatus := currentStatus
deploymentScaled := false
switch currentStatus {
case deployapi.DeploymentStatusNew:
// If the deployment has been cancelled, don't create a deployer pod.
// Transition the deployment to Pending so that re-syncs will check
// up on the deployer pods and so that the deployment config controller
// continues to see the deployment as in-flight (which it is until we
// have deployer pod outcomes).
if deployutil.IsDeploymentCancelled(deployment) {
nextStatus = deployapi.DeploymentStatusPending
if err := c.cancelDeployerPods(deployment); err != nil {
return err
}
break
}
// Generate a deployer pod spec.
podTemplate, err := c.makeDeployerPod(deployment)
if err != nil {
return fatalError(fmt.Sprintf("couldn't make deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err))
}
// Create the deployer pod.
deploymentPod, err := c.podClient.createPod(deployment.Namespace, podTemplate)
if err == nil {
deployment.Annotations[deployapi.DeploymentPodAnnotation] = deploymentPod.Name
nextStatus = deployapi.DeploymentStatusPending
glog.V(4).Infof("Created pod %s for deployment %s", deploymentPod.Name, deployutil.LabelForDeployment(deployment))
break
}
// Retry on error.
if !kerrors.IsAlreadyExists(err) {
if config, decodeErr := c.decodeConfig(deployment); decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "FailedCreate", "Error creating deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
} else {
c.recorder.Eventf(deployment, kapi.EventTypeWarning, "FailedCreate", "Error creating deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
}
return fmt.Errorf("couldn't create deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
}
// If the pod already exists, it's possible that a previous CreatePod
// succeeded but the deployment state update failed and now we're re-
// entering. Ensure that the pod is the one we created by verifying the
// annotation on it, and throw a retryable error.
existingPod, err := c.podClient.getPod(deployment.Namespace, deployutil.DeployerPodNameForDeployment(deployment.Name))
if err != nil {
if config, decodeErr := c.decodeConfig(deployment); decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "FailedCreate", "Error getting existing deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
} else {
c.recorder.Eventf(deployment, kapi.EventTypeWarning, "FailedCreate", "Error getting existing deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
}
return fmt.Errorf("couldn't fetch existing deployer pod for %s: %v", deployutil.LabelForDeployment(deployment), err)
}
// Do a stronger check to validate that the existing deployer pod is
// actually for this deployment, and if not, fail this deployment.
//
// TODO: Investigate checking the container image of the running pod and
// comparing with the intended deployer pod image. If we do so, we'll need
// to ensure that changes to 'unrelated' pods don't result in updates to
// the deployment. So, the image check will have to be done in other areas
// of the code as well.
if deployutil.DeploymentNameFor(existingPod) != deployment.Name {
nextStatus = deployapi.DeploymentStatusFailed
deployment.Annotations[deployapi.DeploymentStatusReasonAnnotation] = deployapi.DeploymentFailedUnrelatedDeploymentExists
if config, decodeErr := c.decodeConfig(deployment); decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "FailedCreate", "Error creating deployer pod for %s since another pod with the same name (%q) exists", deployutil.LabelForDeployment(deployment), existingPod.Name)
} else {
c.recorder.Eventf(deployment, kapi.EventTypeWarning, "FailedCreate", "Error creating deployer pod for %s since another pod with the same name (%q) exists", deployutil.LabelForDeployment(deployment), existingPod.Name)
}
glog.V(2).Infof("Couldn't create deployer pod for %s since an unrelated pod with the same name (%q) exists", deployutil.LabelForDeployment(deployment), existingPod.Name)
break
}
// Update to pending relative to the existing validated deployer pod.
deployment.Annotations[deployapi.DeploymentPodAnnotation] = existingPod.Name
nextStatus = deployapi.DeploymentStatusPending
glog.V(4).Infof("Detected existing deployer pod %s for deployment %s", existingPod.Name, deployutil.LabelForDeployment(deployment))
case deployapi.DeploymentStatusPending, deployapi.DeploymentStatusRunning:
// If the deployer pod has vanished, consider the deployment a failure.
deployerPodName := deployutil.DeployerPodNameForDeployment(deployment.Name)
if _, err := c.podClient.getPod(deployment.Namespace, deployerPodName); err != nil {
if kerrors.IsNotFound(err) {
nextStatus = deployapi.DeploymentStatusFailed
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
deployment.Annotations[deployapi.DeploymentStatusReasonAnnotation] = deployapi.DeploymentFailedDeployerPodNoLongerExists
if config, decodeErr := c.decodeConfig(deployment); decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "Failed", "Deployer pod %q has gone missing", deployerPodName)
} else {
c.recorder.Eventf(deployment, kapi.EventTypeWarning, "Failed", "Deployer pod %q has gone missing", deployerPodName)
}
glog.V(4).Infof("Failing deployment %q because its deployer pod %q disappeared", deployutil.LabelForDeployment(deployment), deployerPodName)
break
} else {
// We'll try again later on resync. Continue to process cancellations.
glog.V(2).Infof("Error getting deployer pod %s for deployment %s: %#v", deployerPodName, deployutil.LabelForDeployment(deployment), err)
}
}
// If the deployment is cancelled, terminate any deployer/hook pods.
// NOTE: Do not mark the deployment as Failed just yet.
// The deployment will be marked as Failed by the deployer pod controller
// when the deployer pod failure state is picked up.
// Then, the deployment config controller will scale down the failed deployment
// and scale back up the last successful completed deployment.
if deployutil.IsDeploymentCancelled(deployment) {
if err := c.cancelDeployerPods(deployment); err != nil {
return err
}
}
case deployapi.DeploymentStatusFailed:
// Check for test deployment and ensure the deployment scale matches
if config, err := c.decodeConfig(deployment); err == nil && config.Spec.Test {
deploymentScaled = deployment.Spec.Replicas != 0
deployment.Spec.Replicas = 0
}
case deployapi.DeploymentStatusComplete:
// Check for test deployment and ensure the deployment scale matches
if config, err := c.decodeConfig(deployment); err == nil && config.Spec.Test {
deploymentScaled = deployment.Spec.Replicas != 0
deployment.Spec.Replicas = 0
}
// now list any pods in the namespace that have the specified label
deployerPods, err := c.podClient.getDeployerPodsFor(deployment.Namespace, deployment.Name)
if err != nil {
return fmt.Errorf("couldn't fetch deployer pods for %s after successful completion: %v", deployutil.LabelForDeployment(deployment), err)
}
if len(deployerPods) > 0 {
glog.V(4).Infof("Deleting %d deployer pods for deployment %s", len(deployerPods), deployutil.LabelForDeployment(deployment))
}
cleanedAll := true
for _, deployerPod := range deployerPods {
if err := c.podClient.deletePod(deployerPod.Namespace, deployerPod.Name); err != nil {
if !kerrors.IsNotFound(err) {
// if the pod deletion failed, then log the error and continue
// we will try to delete any remaining deployer pods and return an error later
utilruntime.HandleError(fmt.Errorf("couldn't delete completed deployer pod %s/%s for deployment %s: %v", deployment.Namespace, deployerPod.Name, deployutil.LabelForDeployment(deployment), err))
cleanedAll = false
}
// Already deleted
} else {
glog.V(4).Infof("Deleted completed deployer pod %s/%s for deployment %s", deployment.Namespace, deployerPod.Name, deployutil.LabelForDeployment(deployment))
}
}
if !cleanedAll {
return fmt.Errorf("couldn't clean up all deployer pods for %s", deployutil.LabelForDeployment(deployment))
}
}
if deployutil.CanTransitionPhase(currentStatus, nextStatus) || deploymentScaled {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
if _, err := c.deploymentClient.updateDeployment(deployment.Namespace, deployment); err != nil {
if config, decodeErr := c.decodeConfig(deployment); decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "FailedUpdate", "Cannot update deployment %s status to %s: %v", deployutil.LabelForDeployment(deployment), nextStatus, err)
} else {
c.recorder.Eventf(deployment, kapi.EventTypeWarning, "FailedUpdate", "Cannot update deployment %s status to %s: %v", deployutil.LabelForDeployment(deployment), nextStatus, err)
}
return fmt.Errorf("couldn't update deployment %s to status %s: %v", deployutil.LabelForDeployment(deployment), nextStatus, err)
}
glog.V(4).Infof("Updated deployment %s status from %s to %s (scale: %d)", deployutil.LabelForDeployment(deployment), currentStatus, nextStatus, deployment.Spec.Replicas)
}
return nil
}
func nextStatusComp(fromDeployer, fromPath deployapi.DeploymentStatus) deployapi.DeploymentStatus {
if deployutil.CanTransitionPhase(fromPath, fromDeployer) {
return fromDeployer
}
return fromPath
}
// Handle syncs pod's status with any associated deployment.
func (c *DeployerPodController) Handle(pod *kapi.Pod) error {
// Find the deployment associated with the deployer pod.
deploymentName := deployutil.DeploymentNameFor(pod)
if len(deploymentName) == 0 {
return nil
}
// Reject updates to anything but the main deployer pod
// TODO: Find a way to filter this on the watch side.
if pod.Name != deployutil.DeployerPodNameForDeployment(deploymentName) {
return nil
}
deployment := &kapi.ReplicationController{ObjectMeta: kapi.ObjectMeta{Namespace: pod.Namespace, Name: deploymentName}}
cached, exists, err := c.store.Get(deployment)
if err == nil && exists {
// Try to use the cache first. Trust hits and return them.
deployment = cached.(*kapi.ReplicationController)
} else {
// Double-check with the master for cache misses/errors, since those
// are rare and API calls are expensive but more reliable.
deployment, err = c.kClient.ReplicationControllers(pod.Namespace).Get(deploymentName)
}
// If the deployment for this pod has disappeared, we should clean up this
// and any other deployer pods, then bail out.
if err != nil {
// Some retrieval error occurred. Retry.
if !kerrors.IsNotFound(err) {
return fmt.Errorf("couldn't get deployment %s/%s which owns deployer pod %s/%s", pod.Namespace, deploymentName, pod.Name, pod.Namespace)
}
// Find all the deployer pods for the deployment (including this one).
opts := kapi.ListOptions{LabelSelector: deployutil.DeployerPodSelector(deploymentName)}
deployers, err := c.kClient.Pods(pod.Namespace).List(opts)
if err != nil {
// Retry.
return fmt.Errorf("couldn't get deployer pods for %s: %v", deployutil.LabelForDeployment(deployment), err)
}
// Delete all deployers.
for _, deployer := range deployers.Items {
err := c.kClient.Pods(deployer.Namespace).Delete(deployer.Name, kapi.NewDeleteOptions(0))
if err != nil {
if !kerrors.IsNotFound(err) {
// TODO: Should this fire an event?
glog.V(2).Infof("Couldn't delete orphaned deployer pod %s/%s: %v", deployer.Namespace, deployer.Name, err)
}
} else {
// TODO: Should this fire an event?
glog.V(2).Infof("Deleted orphaned deployer pod %s/%s", deployer.Namespace, deployer.Name)
}
}
return nil
}
currentStatus := deployutil.DeploymentStatusFor(deployment)
nextStatus := currentStatus
switch pod.Status.Phase {
case kapi.PodRunning:
nextStatus = deployapi.DeploymentStatusRunning
case kapi.PodSucceeded:
nextStatus = deployapi.DeploymentStatusComplete
config, decodeErr := c.decodeConfig(deployment)
// If the deployment was cancelled just prior to the deployer pod succeeding
// then we need to remove the cancel annotations from the complete deployment
// and emit an event letting users know their cancellation failed.
if deployutil.IsDeploymentCancelled(deployment) {
delete(deployment.Annotations, deployapi.DeploymentCancelledAnnotation)
delete(deployment.Annotations, deployapi.DeploymentStatusReasonAnnotation)
if decodeErr == nil {
c.recorder.Eventf(config, kapi.EventTypeWarning, "FailedCancellation", "Deployment %q succeeded before cancel recorded", deployutil.LabelForDeployment(deployment))
} else {
c.recorder.Event(deployment, kapi.EventTypeWarning, "FailedCancellation", "Succeeded before cancel recorded")
}
}
// Sync the internal replica annotation with the target so that we can
// distinguish deployer updates from other scaling events.
deployment.Annotations[deployapi.DeploymentReplicasAnnotation] = deployment.Annotations[deployapi.DesiredReplicasAnnotation]
if nextStatus == deployapi.DeploymentStatusComplete {
delete(deployment.Annotations, deployapi.DesiredReplicasAnnotation)
}
// reset the size of any test container, since we are the ones updating the RC
if decodeErr == nil && config.Spec.Test {
deployment.Spec.Replicas = 0
}
case kapi.PodFailed:
nextStatus = deployapi.DeploymentStatusFailed
// reset the size of any test container, since we are the ones updating the RC
if config, err := c.decodeConfig(deployment); err == nil && config.Spec.Test {
deployment.Spec.Replicas = 0
}
}
if deployutil.CanTransitionPhase(currentStatus, nextStatus) {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
if _, err := c.kClient.ReplicationControllers(deployment.Namespace).Update(deployment); err != nil {
if kerrors.IsNotFound(err) {
return nil
}
return fmt.Errorf("couldn't update Deployment %s to status %s: %v", deployutil.LabelForDeployment(deployment), nextStatus, err)
}
glog.V(4).Infof("Updated deployment %s status from %s to %s (scale: %d)", deployutil.LabelForDeployment(deployment), currentStatus, nextStatus, deployment.Spec.Replicas)
// If the deployment was canceled, trigger a reconcilation of its deployment config
// so that the latest complete deployment can immediately rollback in place of the
// canceled deployment.
if nextStatus == deployapi.DeploymentStatusFailed && deployutil.IsDeploymentCancelled(deployment) {
// If we are unable to get the deployment config, then the deploymentconfig controller will
// perform its duties once the resync interval forces the deploymentconfig to be reconciled.
name := deployutil.DeploymentConfigNameFor(deployment)
kclient.RetryOnConflict(kclient.DefaultRetry, func() error {
config, err := c.client.DeploymentConfigs(deployment.Namespace).Get(name)
if err != nil {
return err
}
if config.Annotations == nil {
config.Annotations = make(map[string]string)
}
config.Annotations[deployapi.DeploymentCancelledAnnotation] = strconv.Itoa(config.Status.LatestVersion)
_, err = c.client.DeploymentConfigs(config.Namespace).Update(config)
return err
})
}
}
return nil
}