// scaleDownOldReplicaSetsForRollingUpdate scales down old replica sets when deployment strategy is "RollingUpdate".
// Need check maxUnavailable to ensure availability
func (dc *DeploymentController) scaleDownOldReplicaSetsForRollingUpdate(allRSs []*extensions.ReplicaSet, oldRSs []*extensions.ReplicaSet, deployment *extensions.Deployment) (int32, error) {
maxUnavailable := deploymentutil.MaxUnavailable(*deployment)
// Check if we can scale down.
minAvailable := deployment.Spec.Replicas - maxUnavailable
minReadySeconds := deployment.Spec.MinReadySeconds
// Find the number of ready pods.
// TODO: use dc.getAvailablePodsForReplicaSets instead
availablePodCount, err := deploymentutil.GetAvailablePodsForReplicaSets(dc.client, deployment, allRSs, minReadySeconds)
if err != nil {
return 0, fmt.Errorf("could not find available pods: %v", err)
}
if availablePodCount <= minAvailable {
// Cannot scale down.
return 0, nil
}
glog.V(4).Infof("Found %d available pods in deployment %s, scaling down old RSes", availablePodCount, deployment.Name)
sort.Sort(controller.ReplicaSetsByCreationTimestamp(oldRSs))
totalScaledDown := int32(0)
totalScaleDownCount := availablePodCount - minAvailable
for _, targetRS := range oldRSs {
if totalScaledDown >= totalScaleDownCount {
// No further scaling required.
break
}
if targetRS.Spec.Replicas == 0 {
// cannot scale down this ReplicaSet.
continue
}
// Scale down.
scaleDownCount := int32(integer.IntMin(int(targetRS.Spec.Replicas), int(totalScaleDownCount-totalScaledDown)))
newReplicasCount := targetRS.Spec.Replicas - scaleDownCount
if newReplicasCount > targetRS.Spec.Replicas {
return 0, fmt.Errorf("when scaling down old RS, got invalid request to scale down %s/%s %d -> %d", targetRS.Namespace, targetRS.Name, targetRS.Spec.Replicas, newReplicasCount)
}
_, _, err = dc.scaleReplicaSetAndRecordEvent(targetRS, newReplicasCount, deployment)
if err != nil {
return totalScaledDown, err
}
totalScaledDown += scaleDownCount
}
return totalScaledDown, nil
}
// cleanupUnhealthyReplicas will scale down old replica sets with unhealthy replicas, so that all unhealthy replicas will be deleted.
func (dc *DeploymentController) cleanupUnhealthyReplicas(oldRSs []*extensions.ReplicaSet, deployment *extensions.Deployment, minReadySeconds, maxCleanupCount int32) ([]*extensions.ReplicaSet, int32, error) {
sort.Sort(controller.ReplicaSetsByCreationTimestamp(oldRSs))
// Safely scale down all old replica sets with unhealthy replicas. Replica set will sort the pods in the order
// such that not-ready < ready, unscheduled < scheduled, and pending < running. This ensures that unhealthy replicas will
// been deleted first and won't increase unavailability.
totalScaledDown := int32(0)
for i, targetRS := range oldRSs {
if totalScaledDown >= maxCleanupCount {
break
}
if targetRS.Spec.Replicas == 0 {
// cannot scale down this replica set.
continue
}
// TODO: use dc.getAvailablePodsForReplicaSets instead
availablePodCount, err := deploymentutil.GetAvailablePodsForReplicaSets(dc.client, deployment, []*extensions.ReplicaSet{targetRS}, minReadySeconds)
if err != nil {
return nil, totalScaledDown, fmt.Errorf("could not find available pods: %v", err)
}
glog.V(4).Infof("Found %d available pods in old RS %s/%s", availablePodCount, targetRS.Namespace, targetRS.Name)
if targetRS.Spec.Replicas == availablePodCount {
// no unhealthy replicas found, no scaling required.
continue
}
scaledDownCount := int32(integer.IntMin(int(maxCleanupCount-totalScaledDown), int(targetRS.Spec.Replicas-availablePodCount)))
newReplicasCount := targetRS.Spec.Replicas - scaledDownCount
if newReplicasCount > targetRS.Spec.Replicas {
return nil, 0, fmt.Errorf("when cleaning up unhealthy replicas, got invalid request to scale down %s/%s %d -> %d", targetRS.Namespace, targetRS.Name, targetRS.Spec.Replicas, newReplicasCount)
}
_, updatedOldRS, err := dc.scaleReplicaSetAndRecordEvent(targetRS, newReplicasCount, deployment)
if err != nil {
return nil, totalScaledDown, err
}
totalScaledDown += scaledDownCount
oldRSs[i] = updatedOldRS
}
return oldRSs, totalScaledDown, nil
}
func (dc *DeploymentController) reconcileOldReplicaSets(allRSs []*extensions.ReplicaSet, oldRSs []*extensions.ReplicaSet, newRS *extensions.ReplicaSet, deployment *extensions.Deployment) (bool, error) {
oldPodsCount := deploymentutil.GetReplicaCountForReplicaSets(oldRSs)
if oldPodsCount == 0 {
// Can't scale down further
return false, nil
}
minReadySeconds := deployment.Spec.MinReadySeconds
allPodsCount := deploymentutil.GetReplicaCountForReplicaSets(allRSs)
// TODO: use dc.getAvailablePodsForReplicaSets instead
newRSAvailablePodCount, err := deploymentutil.GetAvailablePodsForReplicaSets(dc.client, deployment, []*extensions.ReplicaSet{newRS}, minReadySeconds)
if err != nil {
return false, fmt.Errorf("could not find available pods: %v", err)
}
glog.V(4).Infof("New RS %s/%s has %d available pods.", newRS.Namespace, newRS.Name, newRSAvailablePodCount)
maxUnavailable := deploymentutil.MaxUnavailable(*deployment)
// Check if we can scale down. We can scale down in the following 2 cases:
// * Some old replica sets have unhealthy replicas, we could safely scale down those unhealthy replicas since that won't further
// increase unavailability.
// * New replica set has scaled up and it's replicas becomes ready, then we can scale down old replica sets in a further step.
//
// maxScaledDown := allPodsCount - minAvailable - newReplicaSetPodsUnavailable
// take into account not only maxUnavailable and any surge pods that have been created, but also unavailable pods from
// the newRS, so that the unavailable pods from the newRS would not make us scale down old replica sets in a further
// step(that will increase unavailability).
//
// Concrete example:
//
// * 10 replicas
// * 2 maxUnavailable (absolute number, not percent)
// * 3 maxSurge (absolute number, not percent)
//
// case 1:
// * Deployment is updated, newRS is created with 3 replicas, oldRS is scaled down to 8, and newRS is scaled up to 5.
// * The new replica set pods crashloop and never become available.
// * allPodsCount is 13. minAvailable is 8. newRSPodsUnavailable is 5.
// * A node fails and causes one of the oldRS pods to become unavailable. However, 13 - 8 - 5 = 0, so the oldRS won't be scaled down.
// * The user notices the crashloop and does kubectl rollout undo to rollback.
// * newRSPodsUnavailable is 1, since we rolled back to the good replica set, so maxScaledDown = 13 - 8 - 1 = 4. 4 of the crashlooping pods will be scaled down.
// * The total number of pods will then be 9 and the newRS can be scaled up to 10.
//
// case 2:
// Same example, but pushing a new pod template instead of rolling back (aka "roll over"):
// * The new replica set created must start with 0 replicas because allPodsCount is already at 13.
// * However, newRSPodsUnavailable would also be 0, so the 2 old replica sets could be scaled down by 5 (13 - 8 - 0), which would then
// allow the new replica set to be scaled up by 5.
minAvailable := deployment.Spec.Replicas - maxUnavailable
newRSUnavailablePodCount := newRS.Spec.Replicas - newRSAvailablePodCount
maxScaledDown := allPodsCount - minAvailable - newRSUnavailablePodCount
if maxScaledDown <= 0 {
return false, nil
}
// Clean up unhealthy replicas first, otherwise unhealthy replicas will block deployment
// and cause timeout. See https://github.com/kubernetes/kubernetes/issues/16737
oldRSs, cleanupCount, err := dc.cleanupUnhealthyReplicas(oldRSs, deployment, deployment.Spec.MinReadySeconds, maxScaledDown)
if err != nil {
return false, nil
}
glog.V(4).Infof("Cleaned up unhealthy replicas from old RSes by %d", cleanupCount)
// Scale down old replica sets, need check maxUnavailable to ensure we can scale down
allRSs = append(oldRSs, newRS)
scaledDownCount, err := dc.scaleDownOldReplicaSetsForRollingUpdate(allRSs, oldRSs, deployment)
if err != nil {
return false, nil
}
glog.V(4).Infof("Scaled down old RSes of deployment %s by %d", deployment.Name, scaledDownCount)
totalScaledDown := cleanupCount + scaledDownCount
return totalScaledDown > 0, nil
}