// ScaleUp tries to scale the cluster up. Return true if it found a way to increase the size,
// false if it didn't and error if an error occured. Assumes that all nodes in the cluster are
// ready and in sync with instance groups.
func ScaleUp(unschedulablePods []*kube_api.Pod, nodes []*kube_api.Node, cloudProvider cloudprovider.CloudProvider, kubeClient *kube_client.Client,
predicateChecker *simulator.PredicateChecker, recorder kube_record.EventRecorder, maxNodesTotal int,
estimatorName string) (bool, error) {
// From now on we only care about unschedulable pods that were marked after the newest
// node became available for the scheduler.
if len(unschedulablePods) == 0 {
glog.V(1).Info("No unschedulable pods")
return false, nil
}
for _, pod := range unschedulablePods {
glog.V(1).Infof("Pod %s/%s is unschedulable", pod.Namespace, pod.Name)
}
expansionOptions := make([]ExpansionOption, 0)
nodeInfos, err := GetNodeInfosForGroups(nodes, cloudProvider, kubeClient)
if err != nil {
return false, fmt.Errorf("failed to build node infos for node groups: %v", err)
}
podsRemainUnshedulable := make(map[*kube_api.Pod]struct{})
for _, nodeGroup := range cloudProvider.NodeGroups() {
currentSize, err := nodeGroup.TargetSize()
if err != nil {
glog.Errorf("Failed to get node group size: %v", err)
continue
}
if currentSize >= nodeGroup.MaxSize() {
// skip this node group.
glog.V(4).Infof("Skipping node group %s - max size reached", nodeGroup.Id())
continue
}
option := ExpansionOption{
nodeGroup: nodeGroup,
pods: make([]*kube_api.Pod, 0),
}
nodeInfo, found := nodeInfos[nodeGroup.Id()]
if !found {
glog.Errorf("No node info for: %s", nodeGroup.Id())
continue
}
for _, pod := range unschedulablePods {
err = predicateChecker.CheckPredicates(pod, nodeInfo)
if err == nil {
option.pods = append(option.pods, pod)
} else {
glog.V(2).Infof("Scale-up predicate failed: %v", err)
podsRemainUnshedulable[pod] = struct{}{}
}
}
if len(option.pods) > 0 {
if estimatorName == BinpackingEstimatorName {
binpackingEstimator := estimator.NewBinpackingNodeEstimator(predicateChecker)
option.nodeCount = binpackingEstimator.Estimate(option.pods, nodeInfo)
} else if estimatorName == BasicEstimatorName {
basicEstimator := estimator.NewBasicNodeEstimator()
for _, pod := range option.pods {
basicEstimator.Add(pod)
}
option.nodeCount, option.debug = basicEstimator.Estimate(nodeInfo.Node())
} else {
glog.Fatalf("Unrecognized estimator: %s", estimatorName)
}
expansionOptions = append(expansionOptions, option)
}
}
// Pick some expansion option.
bestOption := BestExpansionOption(expansionOptions)
if bestOption != nil && bestOption.nodeCount > 0 {
glog.V(1).Infof("Best option to resize: %s", bestOption.nodeGroup.Id())
if len(bestOption.debug) > 0 {
glog.V(1).Info(bestOption.debug)
}
glog.V(1).Infof("Estimated %d nodes needed in %s", bestOption.nodeCount, bestOption.nodeGroup.Id())
currentSize, err := bestOption.nodeGroup.TargetSize()
if err != nil {
return false, fmt.Errorf("failed to get node group size: %v", err)
}
newSize := currentSize + bestOption.nodeCount
if newSize >= bestOption.nodeGroup.MaxSize() {
glog.V(1).Infof("Capping size to MAX (%d)", bestOption.nodeGroup.MaxSize())
newSize = bestOption.nodeGroup.MaxSize()
}
if maxNodesTotal > 0 && len(nodes)+(newSize-currentSize) > maxNodesTotal {
glog.V(1).Infof("Capping size to max cluster total size (%d)", maxNodesTotal)
newSize = maxNodesTotal - len(nodes) + currentSize
if newSize < currentSize {
return false, fmt.Errorf("max node total count already reached")
}
}
glog.V(0).Infof("Scale-up: setting group %s size to %d", bestOption.nodeGroup.Id(), newSize)
if err := bestOption.nodeGroup.IncreaseSize(newSize - currentSize); err != nil {
return false, fmt.Errorf("failed to increase node group size: %v", err)
}
for _, pod := range bestOption.pods {
recorder.Eventf(pod, kube_api.EventTypeNormal, "TriggeredScaleUp",
"pod triggered scale-up, group: %s, sizes (current/new): %d/%d", bestOption.nodeGroup.Id(), currentSize, newSize)
}
return true, nil
}
for pod := range podsRemainUnshedulable {
recorder.Event(pod, kube_api.EventTypeNormal, "NotTriggerScaleUp",
"pod didn't trigger scale-up (it wouldn't fit if a new node is added)")
}
return false, nil
}
func main() {
flag.Var(&migConfigFlag, "nodes", "sets min,max size and url of a MIG to be controlled by Cluster Autoscaler. "+
"Can be used multiple times. Format: <min>:<max>:<migurl>")
flag.Parse()
url, err := url.Parse(*kubernetes)
if err != nil {
glog.Fatalf("Failed to parse Kuberentes url: %v", err)
}
kubeConfig, err := config.GetKubeClientConfig(url)
if err != nil {
glog.Fatalf("Failed to build Kuberentes client configuration: %v", err)
}
kubeClient := kube_client.NewOrDie(kubeConfig)
unschedulablePodLister := NewUnschedulablePodLister(kubeClient)
nodeLister := NewNodeLister(kubeClient)
migConfigs := make([]*config.MigConfig, 0, len(migConfigFlag))
for i := range migConfigFlag {
migConfigs = append(migConfigs, &migConfigFlag[i])
}
gceManager, err := gce.CreateGceManager(migConfigs)
if err != nil {
glog.Fatalf("Failed to create GCE Manager %v", err)
}
predicateChecker := simulator.NewPredicateChecker()
for {
select {
case <-time.After(time.Minute):
{
nodes, err := nodeLister.List()
if err != nil {
glog.Errorf("Failed to list nodes: %v", err)
continue
}
if len(nodes) == 0 {
glog.Errorf("No nodes in the cluster")
continue
}
if err := CheckMigsAndNodes(nodes, gceManager); err != nil {
glog.Warningf("Cluster is not ready for autoscaling: %v", err)
continue
}
pods, err := unschedulablePodLister.List()
if err != nil {
glog.Errorf("Failed to list unscheduled pods: %v", err)
continue
}
// We need to reset all pods that have been marked as unschedulable not after
// the newest node became available for the scheduler.
allNodesAvailableTime := GetAllNodesAvailableTime(nodes)
resetOldPods(kubeClient, pods, allNodesAvailableTime)
// From now on we only care about unschedulable pods that were marked after the newest
// node became available for the scheduler.
pods = filterOldPods(pods, allNodesAvailableTime)
if len(pods) == 0 {
glog.V(1).Info("No unschedulable pods")
continue
}
for _, pod := range pods {
glog.V(1).Infof("Pod %s/%s is unschedulable", pod.Namespace, pod.Name)
}
expansionOptions := make([]ExpansionOption, 0)
nodeInfos, err := GetNodeInfosForMigs(nodes, gceManager, kubeClient)
if err != nil {
glog.Errorf("Failed to build node infors for migs: %v", err)
continue
}
for _, migConfig := range migConfigs {
currentSize, err := gceManager.GetMigSize(migConfig)
if err != nil {
glog.Errorf("Failed to get MIG size: %v", err)
continue
}
if currentSize >= int64(migConfig.MaxSize) {
// skip this mig.
glog.V(4).Infof("Skipping MIG %s - max size reached", migConfig.Url())
continue
}
option := ExpansionOption{
migConfig: migConfig,
estimator: estimator.NewBasicNodeEstimator(),
}
migHelpsSomePods := false
nodeInfo, found := nodeInfos[migConfig.Url()]
if !found {
glog.Errorf("No node info for: %s", migConfig.Url())
continue
}
for _, pod := range pods {
err = predicateChecker.CheckPredicates(pod, nodeInfo)
if err == nil {
migHelpsSomePods = true
option.estimator.Add(pod)
} else {
glog.V(2).Infof("Scale-up predicate failed: %v", err)
}
}
if migHelpsSomePods {
expansionOptions = append(expansionOptions, option)
}
}
// Pick some expansion option.
bestOption := BestExpansionOption(expansionOptions)
if bestOption != nil {
glog.V(1).Infof("Best option to resize: %s", bestOption.migConfig.Url())
nodeInfo, found := nodeInfos[bestOption.migConfig.Url()]
if !found {
glog.Errorf("No sample node for: %s", bestOption.migConfig.Url())
continue
}
node := nodeInfo.Node()
estimate, report := bestOption.estimator.Estimate(node)
glog.V(1).Info(bestOption.estimator.GetDebug())
glog.V(1).Info(report)
glog.V(1).Infof("Estimated %d nodes needed in %s", estimate, bestOption.migConfig.Url())
currentSize, err := gceManager.GetMigSize(bestOption.migConfig)
if err != nil {
glog.Errorf("Failed to get MIG size: %v", err)
continue
}
newSize := currentSize + int64(estimate)
if newSize >= int64(bestOption.migConfig.MaxSize) {
newSize = int64(bestOption.migConfig.MaxSize)
}
glog.V(1).Infof("Setting %s size to %d", bestOption.migConfig.Url(), newSize)
if err := gceManager.SetMigSize(bestOption.migConfig, newSize); err != nil {
glog.Errorf("Failed to set MIG size: %v", err)
}
}
}
}
}
}
// ScaleUp tries to scale the cluster up. Return true if it found a way to increase the size,
// false if it didn't and error if an error occured.
func ScaleUp(unschedulablePods []*kube_api.Pod, nodes []*kube_api.Node, migConfigs []*config.MigConfig,
gceManager *gce.GceManager, kubeClient *kube_client.Client,
predicateChecker *simulator.PredicateChecker, recorder kube_record.EventRecorder) (bool, error) {
// From now on we only care about unschedulable pods that were marked after the newest
// node became available for the scheduler.
if len(unschedulablePods) == 0 {
glog.V(1).Info("No unschedulable pods")
return false, nil
}
for _, pod := range unschedulablePods {
glog.V(1).Infof("Pod %s/%s is unschedulable", pod.Namespace, pod.Name)
}
expansionOptions := make([]ExpansionOption, 0)
nodeInfos, err := GetNodeInfosForMigs(nodes, gceManager, kubeClient)
if err != nil {
return false, fmt.Errorf("failed to build node infors for migs: %v", err)
}
podsRemainUnshedulable := make(map[*kube_api.Pod]struct{})
for _, migConfig := range migConfigs {
currentSize, err := gceManager.GetMigSize(migConfig)
if err != nil {
glog.Errorf("Failed to get MIG size: %v", err)
continue
}
if currentSize >= int64(migConfig.MaxSize) {
// skip this mig.
glog.V(4).Infof("Skipping MIG %s - max size reached", migConfig.Url())
continue
}
option := ExpansionOption{
migConfig: migConfig,
estimator: estimator.NewBasicNodeEstimator(),
}
migHelpsSomePods := false
nodeInfo, found := nodeInfos[migConfig.Url()]
if !found {
glog.Errorf("No node info for: %s", migConfig.Url())
continue
}
for _, pod := range unschedulablePods {
err = predicateChecker.CheckPredicates(pod, nodeInfo)
if err == nil {
migHelpsSomePods = true
option.estimator.Add(pod)
} else {
glog.V(2).Infof("Scale-up predicate failed: %v", err)
podsRemainUnshedulable[pod] = struct{}{}
}
}
if migHelpsSomePods {
expansionOptions = append(expansionOptions, option)
}
}
// Pick some expansion option.
bestOption := BestExpansionOption(expansionOptions)
if bestOption != nil && bestOption.estimator.GetCount() > 0 {
glog.V(1).Infof("Best option to resize: %s", bestOption.migConfig.Url())
nodeInfo, found := nodeInfos[bestOption.migConfig.Url()]
if !found {
return false, fmt.Errorf("no sample node for: %s", bestOption.migConfig.Url())
}
node := nodeInfo.Node()
estimate, report := bestOption.estimator.Estimate(node)
glog.V(1).Info(bestOption.estimator.GetDebug())
glog.V(1).Info(report)
glog.V(1).Infof("Estimated %d nodes needed in %s", estimate, bestOption.migConfig.Url())
currentSize, err := gceManager.GetMigSize(bestOption.migConfig)
if err != nil {
return false, fmt.Errorf("failed to get MIG size: %v", err)
}
newSize := currentSize + int64(estimate)
if newSize >= int64(bestOption.migConfig.MaxSize) {
glog.V(1).Infof("Capping size to MAX (%d)", bestOption.migConfig.MaxSize)
newSize = int64(bestOption.migConfig.MaxSize)
}
glog.V(1).Infof("Setting %s size to %d", bestOption.migConfig.Url(), newSize)
if err := gceManager.SetMigSize(bestOption.migConfig, newSize); err != nil {
return false, fmt.Errorf("failed to set MIG size: %v", err)
}
for pod := range bestOption.estimator.FittingPods {
recorder.Eventf(pod, kube_api.EventTypeNormal, "TriggeredScaleUp",
"pod triggered scale-up, mig: %s, sizes (current/new): %d/%d", bestOption.migConfig.Name, currentSize, newSize)
}
return true, nil
}
for pod := range podsRemainUnshedulable {
recorder.Event(pod, kube_api.EventTypeNormal, "NotTriggerScaleUp",
"pod didn't trigger scale-up (it wouldn't fit if a new node is added)")
}
return false, nil
}
func main() {
flag.Var(&migConfigFlag, "nodes", "sets min,max size and url of a MIG to be controlled by Cluster Autoscaler. "+
"Can be used multiple times. Format: <min>:<max>:<migurl>")
flag.Parse()
url, err := url.Parse(*kubernetes)
if err != nil {
glog.Fatalf("Failed to parse Kuberentes url: %v", err)
}
kubeConfig, err := config.GetKubeClientConfig(url)
if err != nil {
glog.Fatalf("Failed to build Kuberentes client configuration: %v", err)
}
kubeClient := kube_client.NewOrDie(kubeConfig)
unscheduledPodLister := NewUnscheduledPodLister(kubeClient)
nodeLister := NewNodeLister(kubeClient)
migConfigs := make([]*config.MigConfig, 0, len(migConfigFlag))
for i := range migConfigFlag {
migConfigs = append(migConfigs, &migConfigFlag[i])
}
gceManager, err := gce.CreateGceManager(migConfigs)
if err != nil {
glog.Fatalf("Failed to create GCE Manager %v", err)
}
predicateChecker := simulator.NewPredicateChecker()
for {
select {
case <-time.After(time.Minute):
{
pods, err := unscheduledPodLister.List()
if err != nil {
glog.Errorf("Failed to list pods: %v", err)
continue
}
if len(pods) == 0 {
glog.V(1).Info("No unscheduled pods")
continue
}
for _, pod := range pods {
glog.V(1).Infof("Pod %s/%s is not scheduled", pod.Namespace, pod.Name)
}
nodes, err := nodeLister.List()
if err != nil {
glog.Errorf("Failed to list nodes: %v", err)
continue
}
if len(nodes) == 0 {
glog.Errorf("No nodes in the cluster")
continue
}
if err := CheckMigsAndNodes(nodes, gceManager); err != nil {
glog.Warningf("Cluster is not ready for autoscaling: %v", err)
continue
}
// Checks if scheduler tried to schedule the pods after thew newest node was added.
newestNode := GetNewestNode(nodes)
if newestNode == nil {
glog.Errorf("No newest node")
continue
}
oldestSchedulingTrial := GetOldestFailedSchedulingTrail(pods)
if oldestSchedulingTrial == nil {
glog.Errorf("No oldest unschedueled trial: %v", err)
continue
}
// TODO: Find better way to check if all pods were checked after the newest node
// was added.
if newestNode.CreationTimestamp.After(oldestSchedulingTrial.Add(-1 * time.Minute)) {
// Lets give scheduler another chance.
glog.V(1).Infof("One of the pods have not been tried after adding %s", newestNode.Name)
continue
}
expansionOptions := make([]ExpansionOption, 0)
nodeInfos, sampleNodes, err := GetNodeInfosForMigs(nodes, gceManager, kubeClient)
if err != nil {
glog.Errorf("Failed to build node infors for migs: %v", err)
continue
}
for _, migConfig := range migConfigs {
currentSize, err := gceManager.GetMigSize(migConfig)
if err != nil {
glog.Errorf("Failed to get MIG size: %v", err)
continue
}
if currentSize >= int64(migConfig.MaxSize) {
// skip this mig.
glog.V(4).Infof("Skipping MIG %s - max size reached", migConfig.Url())
continue
}
option := ExpansionOption{
migConfig: migConfig,
estimator: estimator.NewBasicNodeEstimator(),
}
migHelpsSomePods := false
nodeInfo, found := nodeInfos[migConfig.Url()]
if !found {
glog.Errorf("No node info for: %s", migConfig.Url())
continue
}
node, found := sampleNodes[migConfig.Url()]
if !found {
glog.Errorf("No sample node for: %s", migConfig.Url())
continue
}
for _, pod := range pods {
err = predicateChecker.CheckPredicates(pod, node, nodeInfo)
if err == nil {
migHelpsSomePods = true
option.estimator.Add(pod)
} else {
glog.V(2).Infof("Scale-up predicate failed: %v", err)
}
}
if migHelpsSomePods {
expansionOptions = append(expansionOptions, option)
}
}
// Pick some expansion option.
bestOption := BestExpansionOption(expansionOptions)
if bestOption != nil {
glog.V(1).Infof("Best option to resize: %s", bestOption.migConfig.Url())
node, found := sampleNodes[bestOption.migConfig.Url()]
if !found {
glog.Errorf("No sample node for: %s", bestOption.migConfig.Url())
continue
}
estimate := bestOption.estimator.Estimate(node)
glog.V(1).Infof("Estimated %d nodes needed in %s", estimate, bestOption.migConfig.Url())
currentSize, err := gceManager.GetMigSize(bestOption.migConfig)
if err != nil {
glog.Errorf("Failed to get MIG size: %v", err)
continue
}
newSize := currentSize + int64(estimate)
if newSize >= int64(bestOption.migConfig.MaxSize) {
newSize = int64(bestOption.migConfig.MaxSize)
}
glog.V(1).Infof("Setting %s size to %d", bestOption.migConfig.Url(), newSize)
if err := gceManager.SetMigSize(bestOption.migConfig, newSize); err != nil {
glog.Errorf("Failed to set MIG size: %v", err)
}
}
}
}
}
}