// 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
switch currentStatus {
case deployapi.DeploymentStatusNew:
// If the deployment has been cancelled, don't create a deployer pod, and
// transition to failed immediately.
if deployutil.IsDeploymentCancelled(deployment) {
nextStatus = deployapi.DeploymentStatusFailed
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) {
c.recorder.Eventf(deployment, "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 {
c.recorder.Eventf(deployment, "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
c.recorder.Eventf(deployment, "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
c.recorder.Eventf(deployment, "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
// Also, it will scale down the failed deployment and scale back up
// the last successful completed deployment
if deployutil.IsDeploymentCancelled(deployment) {
deployerPods, err := c.podClient.getDeployerPodsFor(deployment.Namespace, deployment.Name)
if err != nil {
return fmt.Errorf("couldn't fetch deployer pods for %s while trying to cancel deployment: %v", deployutil.LabelForDeployment(deployment), err)
}
glog.V(4).Infof("Cancelling %d deployer pods for deployment %s", len(deployerPods), deployutil.LabelForDeployment(deployment))
zeroDelay := int64(1)
for _, deployerPod := range deployerPods {
// Set the ActiveDeadlineSeconds on the pod so it's terminated very soon.
if deployerPod.Spec.ActiveDeadlineSeconds == nil || *deployerPod.Spec.ActiveDeadlineSeconds != zeroDelay {
deployerPod.Spec.ActiveDeadlineSeconds = &zeroDelay
if _, err := c.podClient.updatePod(deployerPod.Namespace, &deployerPod); err != nil {
c.recorder.Eventf(deployment, "failedCancellation", "Error cancelling deployer pod %s for deployment %s: %v", deployerPod.Name, deployutil.LabelForDeployment(deployment), err)
return fmt.Errorf("couldn't cancel deployer pod %s for deployment %s: %v", deployutil.LabelForDeployment(deployment), err)
}
glog.V(4).Infof("Cancelled deployer pod %s for deployment %s", deployerPod.Name, deployutil.LabelForDeployment(deployment))
}
}
c.recorder.Eventf(deployment, "cancelled", "Cancelled deployment")
}
case deployapi.DeploymentStatusFailed:
// Nothing to do in this terminal state.
glog.V(4).Infof("Ignoring deployment %s (status %s)", deployutil.LabelForDeployment(deployment), currentStatus)
case deployapi.DeploymentStatusComplete:
// 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)
}
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
kutil.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 currentStatus != nextStatus {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
if _, err := c.deploymentClient.updateDeployment(deployment.Namespace, deployment); err != nil {
c.recorder.Eventf(deployment, "failedUpdate", "Error updating deployment %s status to %s", deployutil.LabelForDeployment(deployment), nextStatus)
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", deployutil.LabelForDeployment(deployment), currentStatus, nextStatus)
}
return nil
}
// 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, err := c.deploymentClient.getDeployment(pod.Namespace, 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 occured. 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).
deployers, err := c.deployerPodsFor(pod.Namespace, deploymentName)
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.deletePod(deployer.Namespace, deployer.Name)
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:
// Detect failure based on the container state
nextStatus = deployapi.DeploymentStatusComplete
for _, info := range pod.Status.ContainerStatuses {
if info.State.Termination != nil && info.State.Termination.ExitCode != 0 {
nextStatus = deployapi.DeploymentStatusFailed
}
}
case kapi.PodFailed:
// if the deployment is already marked Failed, do not attempt clean up again
if currentStatus != deployapi.DeploymentStatusFailed {
// clean up will also update the deployment status to Failed
// failure to clean up will result in retries and
// the deployment will not be marked Failed
// Note: this will prevent new deployments from being created for this config
err := c.cleanupFailedDeployment(deployment)
if err != nil {
return transientError(fmt.Sprintf("couldn't clean up failed deployment: %v", err))
}
}
}
if currentStatus != nextStatus {
deployment.Annotations[deployapi.DeploymentStatusAnnotation] = string(nextStatus)
if _, err := c.deploymentClient.updateDeployment(deployment.Namespace, 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", deployutil.LabelForDeployment(deployment), currentStatus, nextStatus)
}
return nil
}