Merge pull request kubernetes#5695 from kawych/tpu

Add atomic scale down option for node groups

cluster-autoscaler/cloudprovider/test/test_cloud_provider.go

@@ -443,6 +443,9 @@ func (tng *TestNodeGroup) DeleteNodes(nodes []*apiv1.Node) error {
 	id := tng.id
 	tng.targetSize -= len(nodes)
 	tng.Unlock()
+	if tng.opts != nil && tng.opts.AtomicScaling && tng.targetSize != 0 {
+		return fmt.Errorf("TestNodeGroup: attempted to partially scale down a node group that should be scaled down atomically")
+	}
 	for _, node := range nodes {
 		err := tng.cloudProvider.onScaleDown(id, node.Name)
 		if err != nil {

cluster-autoscaler/config/autoscaling_options.go

@@ -46,6 +46,8 @@ type NodeGroupAutoscalingOptions struct {
 	ScaleDownUnreadyTime time.Duration
 	// Maximum time CA waits for node to be provisioned
 	MaxNodeProvisionTime time.Duration
+	// AtomicScaling means that all nodes should be provisioned and brought down all at once instead of one-by-one
+	AtomicScaling bool
 }
 
 // GCEOptions contain autoscaling options specific to GCE cloud provider.