azure_vmss.go

/*
Copyright 2017 The Kubernetes Authors.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/

package azure

import (
	"errors"
	"fmt"
	"regexp"
	"sort"
	"strconv"
	"strings"

	"github.com/Azure/azure-sdk-for-go/arm/network"
	computepreview "github.com/Azure/azure-sdk-for-go/services/compute/mgmt/2017-12-01/compute"
	"github.com/Azure/go-autorest/autorest/to"
	"github.com/golang/glog"

	"k8s.io/api/core/v1"
	"k8s.io/apimachinery/pkg/types"
	"k8s.io/apimachinery/pkg/util/wait"
	"k8s.io/kubernetes/pkg/cloudprovider"
)

var (
	// ErrorNotVmssInstance indicates an instance is not belongint to any vmss.
	ErrorNotVmssInstance = errors.New("not a vmss instance")

	scaleSetNameRE        = regexp.MustCompile(`.*/subscriptions/(?:.*)/Microsoft.Compute/virtualMachineScaleSets/(.+)/virtualMachines(?:.*)`)
	vmssMachineIDTemplate = "/subscriptions/%s/resourceGroups/%s/providers/Microsoft.Compute/virtualMachineScaleSets/%s/virtualMachines/%s"
)

// scaleSet implements VMSet interface for Azure scale set.
type scaleSet struct {
	*Cloud

	// availabilitySet is also required for scaleSet because some instances
	// (e.g. master nodes) may not belong to any scale sets.
	availabilitySet VMSet

	vmssCache                      *timedCache
	vmssVMCache                    *timedCache
	nodeNameToScaleSetMappingCache *timedCache
	availabilitySetNodesCache      *timedCache
}

// newScaleSet creates a new scaleSet.
func newScaleSet(az *Cloud) (VMSet, error) {
	var err error
	ss := &scaleSet{
		Cloud:           az,
		availabilitySet: newAvailabilitySet(az),
	}

	ss.nodeNameToScaleSetMappingCache, err = ss.newNodeNameToScaleSetMappingCache()
	if err != nil {
		return nil, err
	}

	ss.availabilitySetNodesCache, err = ss.newAvailabilitySetNodesCache()
	if err != nil {
		return nil, err
	}

	ss.vmssCache, err = ss.newVmssCache()
	if err != nil {
		return nil, err
	}

	ss.vmssVMCache, err = ss.newVmssVMCache()
	if err != nil {
		return nil, err
	}

	return ss, nil
}

// getVmssVM gets virtualMachineScaleSetVM by nodeName from cache.
// It returns cloudprovider.InstanceNotFound if node does not belong to any scale sets.
func (ss *scaleSet) getVmssVM(nodeName string) (ssName, instanceID string, vm computepreview.VirtualMachineScaleSetVM, err error) {
	instanceID, err = getScaleSetVMInstanceID(nodeName)
	if err != nil {
		return ssName, instanceID, vm, err
	}

	ssName, err = ss.getScaleSetNameByNodeName(nodeName)
	if err != nil {
		return ssName, instanceID, vm, err
	}

	if ssName == "" {
		return "", "", vm, cloudprovider.InstanceNotFound
	}

	glog.V(4).Infof("getVmssVM gets scaleSetName (%q) and instanceID (%q) for node %q", ssName, instanceID, nodeName)
	cachedVM, err := ss.vmssVMCache.Get(ss.makeVmssVMName(ssName, instanceID))
	if err != nil {
		return ssName, instanceID, vm, err
	}

	if cachedVM == nil {
		glog.Errorf("Can't find node (%q) in any scale sets", nodeName)
		return ssName, instanceID, vm, cloudprovider.InstanceNotFound
	}

	return ssName, instanceID, *(cachedVM.(*computepreview.VirtualMachineScaleSetVM)), nil
}

// getCachedVirtualMachineByInstanceID gets scaleSetVMInfo from cache.
// The node must belong to one of scale sets.
func (ss *scaleSet) getVmssVMByInstanceID(scaleSetName, instanceID string) (vm computepreview.VirtualMachineScaleSetVM, err error) {
	vmName := ss.makeVmssVMName(scaleSetName, instanceID)
	cachedVM, err := ss.vmssVMCache.Get(vmName)
	if err != nil {
		return vm, err
	}

	if cachedVM == nil {
		glog.Errorf("cound't find vmss virtual machine by scaleSetName (%q) and instanceID (%q)", scaleSetName, instanceID)
		return vm, cloudprovider.InstanceNotFound
	}

	return *(cachedVM.(*computepreview.VirtualMachineScaleSetVM)), nil
}

// GetInstanceIDByNodeName gets the cloud provider ID by node name.
// It must return ("", cloudprovider.InstanceNotFound) if the instance does
// not exist or is no longer running.
func (ss *scaleSet) GetInstanceIDByNodeName(name string) (string, error) {
	managedByAS, err := ss.isNodeManagedByAvailabilitySet(name)
	if err != nil {
		glog.Errorf("Failed to check isNodeManagedByAvailabilitySet: %v", err)
		return "", err
	}
	if managedByAS {
		// vm is managed by availability set.
		return ss.availabilitySet.GetInstanceIDByNodeName(name)
	}

	_, _, vm, err := ss.getVmssVM(name)
	if err != nil {
		return "", err
	}

	return *vm.ID, nil
}

// GetNodeNameByProviderID gets the node name by provider ID.
func (ss *scaleSet) GetNodeNameByProviderID(providerID string) (types.NodeName, error) {
	// NodeName is not part of providerID for vmss instances.
	scaleSetName, err := extractScaleSetNameByExternalID(providerID)
	if err != nil {
		glog.V(4).Infof("Can not extract scale set name from providerID (%s), assuming it is mananaged by availability set: %v", providerID, err)
		return ss.availabilitySet.GetNodeNameByProviderID(providerID)
	}

	instanceID, err := getLastSegment(providerID)
	if err != nil {
		glog.V(4).Infof("Can not extract instanceID from providerID (%s), assuming it is mananaged by availability set: %v", providerID, err)
		return ss.availabilitySet.GetNodeNameByProviderID(providerID)
	}

	vm, err := ss.getVmssVMByInstanceID(scaleSetName, instanceID)
	if err != nil {
		return "", err
	}

	if vm.OsProfile != nil && vm.OsProfile.ComputerName != nil {
		nodeName := strings.ToLower(*vm.OsProfile.ComputerName)
		return types.NodeName(nodeName), nil
	}

	return "", nil
}

// GetInstanceTypeByNodeName gets the instance type by node name.
func (ss *scaleSet) GetInstanceTypeByNodeName(name string) (string, error) {
	managedByAS, err := ss.isNodeManagedByAvailabilitySet(name)
	if err != nil {
		glog.Errorf("Failed to check isNodeManagedByAvailabilitySet: %v", err)
		return "", err
	}
	if managedByAS {
		// vm is managed by availability set.
		return ss.availabilitySet.GetInstanceTypeByNodeName(name)
	}

	_, _, vm, err := ss.getVmssVM(name)
	if err != nil {
		return "", err
	}

	if vm.Sku != nil && vm.Sku.Name != nil {
		return *vm.Sku.Name, nil
	}

	return "", nil
}

// GetZoneByNodeName gets cloudprovider.Zone by node name.
func (ss *scaleSet) GetZoneByNodeName(name string) (cloudprovider.Zone, error) {
	managedByAS, err := ss.isNodeManagedByAvailabilitySet(name)
	if err != nil {
		glog.Errorf("Failed to check isNodeManagedByAvailabilitySet: %v", err)
		return cloudprovider.Zone{}, err
	}
	if managedByAS {
		// vm is managed by availability set.
		return ss.availabilitySet.GetZoneByNodeName(name)
	}

	_, _, vm, err := ss.getVmssVM(name)
	if err != nil {
		return cloudprovider.Zone{}, err
	}

	if vm.InstanceView != nil && vm.InstanceView.PlatformFaultDomain != nil {
		return cloudprovider.Zone{
			FailureDomain: strconv.Itoa(int(*vm.InstanceView.PlatformFaultDomain)),
			Region:        *vm.Location,
		}, nil
	}

	return cloudprovider.Zone{}, nil
}

// GetPrimaryVMSetName returns the VM set name depending on the configured vmType.
// It returns config.PrimaryScaleSetName for vmss and config.PrimaryAvailabilitySetName for standard vmType.
func (ss *scaleSet) GetPrimaryVMSetName() string {
	return ss.Config.PrimaryScaleSetName
}

// GetIPByNodeName gets machine IP by node name.
func (ss *scaleSet) GetIPByNodeName(nodeName, vmSetName string) (string, error) {
	nic, err := ss.GetPrimaryInterface(nodeName, vmSetName)
	if err != nil {
		glog.Errorf("error: ss.GetIPByNodeName(%s), GetPrimaryInterface(%q, %q), err=%v", nodeName, nodeName, vmSetName, err)
		return "", err
	}

	ipConfig, err := getPrimaryIPConfig(nic)
	if err != nil {
		glog.Errorf("error: ss.GetIPByNodeName(%s), getPrimaryIPConfig(%v), err=%v", nodeName, nic, err)
		return "", err
	}

	targetIP := *ipConfig.PrivateIPAddress
	return targetIP, nil
}

// This returns the full identifier of the primary NIC for the given VM.
func (ss *scaleSet) getPrimaryInterfaceID(machine computepreview.VirtualMachineScaleSetVM) (string, error) {
	if len(*machine.NetworkProfile.NetworkInterfaces) == 1 {
		return *(*machine.NetworkProfile.NetworkInterfaces)[0].ID, nil
	}

	for _, ref := range *machine.NetworkProfile.NetworkInterfaces {
		if *ref.Primary {
			return *ref.ID, nil
		}
	}

	return "", fmt.Errorf("failed to find a primary nic for the vm. vmname=%q", *machine.Name)
}

// machineName is composed of computerNamePrefix and 36-based instanceID.
// And instanceID part if in fixed length of 6 characters.
// Refer https://msftstack.wordpress.com/2017/05/10/figuring-out-azure-vm-scale-set-machine-names/.
func getScaleSetVMInstanceID(machineName string) (string, error) {
	nameLength := len(machineName)
	if nameLength < 6 {
		return "", ErrorNotVmssInstance
	}

	instanceID, err := strconv.ParseUint(machineName[nameLength-6:], 36, 64)
	if err != nil {
		return "", ErrorNotVmssInstance
	}

	return fmt.Sprintf("%d", instanceID), nil
}

// extractScaleSetNameByExternalID extracts the scaleset name by node's externalID.
func extractScaleSetNameByExternalID(externalID string) (string, error) {
	matches := scaleSetNameRE.FindStringSubmatch(externalID)
	if len(matches) != 2 {
		return "", ErrorNotVmssInstance
	}

	return matches[1], nil
}

// listScaleSets lists all scale sets.
func (ss *scaleSet) listScaleSets() ([]string, error) {
	var err error
	ctx, cancel := getContextWithCancel()
	defer cancel()

	allScaleSets, err := ss.VirtualMachineScaleSetsClient.List(ctx, ss.ResourceGroup)
	if err != nil {
		glog.Errorf("VirtualMachineScaleSetsClient.List failed: %v", err)
		return nil, err
	}

	ssNames := make([]string, len(allScaleSets))
	for i := range allScaleSets {
		ssNames[i] = *(allScaleSets[i].Name)
	}

	return ssNames, nil
}

// listScaleSetVMs lists VMs belonging to the specified scale set.
func (ss *scaleSet) listScaleSetVMs(scaleSetName string) ([]computepreview.VirtualMachineScaleSetVM, error) {
	var err error
	ctx, cancel := getContextWithCancel()
	defer cancel()

	allVMs, err := ss.VirtualMachineScaleSetVMsClient.List(ctx, ss.ResourceGroup, scaleSetName, "", "", string(computepreview.InstanceView))
	if err != nil {
		glog.Errorf("VirtualMachineScaleSetVMsClient.List failed: %v", err)
		return nil, err
	}

	return allVMs, nil
}

// getAgentPoolScaleSets lists the virtual machines for the resource group and then builds
// a list of scale sets that match the nodes available to k8s.
func (ss *scaleSet) getAgentPoolScaleSets(nodes []*v1.Node) (*[]string, error) {
	agentPoolScaleSets := &[]string{}
	for nx := range nodes {
		if isMasterNode(nodes[nx]) {
			continue
		}

		nodeName := nodes[nx].Name
		ssName, err := ss.getScaleSetNameByNodeName(nodeName)
		if err != nil {
			return nil, err
		}

		if ssName == "" {
			glog.V(3).Infof("Node %q is not belonging to any known scale sets", nodeName)
			continue
		}

		*agentPoolScaleSets = append(*agentPoolScaleSets, ssName)
	}

	return agentPoolScaleSets, nil
}

// GetVMSetNames selects all possible availability sets or scale sets
// (depending vmType configured) for service load balancer. If the service has
// no loadbalancer mode annotation returns the primary VMSet. If service annotation
// for loadbalancer exists then return the eligible VMSet.
func (ss *scaleSet) GetVMSetNames(service *v1.Service, nodes []*v1.Node) (vmSetNames *[]string, err error) {
	hasMode, isAuto, serviceVMSetNames := getServiceLoadBalancerMode(service)
	if !hasMode {
		// no mode specified in service annotation default to PrimaryScaleSetName.
		scaleSetNames := &[]string{ss.Config.PrimaryScaleSetName}
		return scaleSetNames, nil
	}

	scaleSetNames, err := ss.getAgentPoolScaleSets(nodes)
	if err != nil {
		glog.Errorf("ss.GetVMSetNames - getAgentPoolScaleSets failed err=(%v)", err)
		return nil, err
	}
	if len(*scaleSetNames) == 0 {
		glog.Errorf("ss.GetVMSetNames - No scale sets found for nodes in the cluster, node count(%d)", len(nodes))
		return nil, fmt.Errorf("No scale sets found for nodes, node count(%d)", len(nodes))
	}

	// sort the list to have deterministic selection
	sort.Strings(*scaleSetNames)

	if !isAuto {
		if serviceVMSetNames == nil || len(serviceVMSetNames) == 0 {
			return nil, fmt.Errorf("service annotation for LoadBalancerMode is empty, it should have __auto__ or availability sets value")
		}
		// validate scale set exists
		var found bool
		for sasx := range serviceVMSetNames {
			for asx := range *scaleSetNames {
				if strings.EqualFold((*scaleSetNames)[asx], serviceVMSetNames[sasx]) {
					found = true
					serviceVMSetNames[sasx] = (*scaleSetNames)[asx]
					break
				}
			}
			if !found {
				glog.Errorf("ss.GetVMSetNames - scale set (%s) in service annotation not found", serviceVMSetNames[sasx])
				return nil, fmt.Errorf("scale set (%s) - not found", serviceVMSetNames[sasx])
			}
		}
		vmSetNames = &serviceVMSetNames
	}

	return vmSetNames, nil
}

// GetPrimaryInterface gets machine primary network interface by node name and vmSet.
func (ss *scaleSet) GetPrimaryInterface(nodeName, vmSetName string) (network.Interface, error) {
	managedByAS, err := ss.isNodeManagedByAvailabilitySet(nodeName)
	if err != nil {
		glog.Errorf("Failed to check isNodeManagedByAvailabilitySet: %v", err)
		return network.Interface{}, err
	}
	if managedByAS {
		// vm is managed by availability set.
		return ss.availabilitySet.GetPrimaryInterface(nodeName, "")
	}

	ssName, instanceID, vm, err := ss.getVmssVM(nodeName)
	if err != nil {
		glog.Errorf("error: ss.GetPrimaryInterface(%s), ss.getVmssVM(%s), err=%v", nodeName, nodeName, err)
		return network.Interface{}, err
	}

	// Check scale set name.
	if vmSetName != "" && !strings.EqualFold(ssName, vmSetName) {
		return network.Interface{}, errNotInVMSet
	}

	primaryInterfaceID, err := ss.getPrimaryInterfaceID(vm)
	if err != nil {
		glog.Errorf("error: ss.GetPrimaryInterface(%s), ss.getPrimaryInterfaceID(), err=%v", nodeName, err)
		return network.Interface{}, err
	}

	nicName, err := getLastSegment(primaryInterfaceID)
	if err != nil {
		glog.Errorf("error: ss.GetPrimaryInterface(%s), getLastSegment(%s), err=%v", nodeName, primaryInterfaceID, err)
		return network.Interface{}, err
	}

	nic, err := ss.InterfacesClient.GetVirtualMachineScaleSetNetworkInterface(ss.ResourceGroup, ssName, instanceID, nicName, "")
	if err != nil {
		glog.Errorf("error: ss.GetPrimaryInterface(%s), ss.GetVirtualMachineScaleSetNetworkInterface.Get(%s, %s, %s), err=%v", nodeName, ss.ResourceGroup, ssName, nicName, err)
		return network.Interface{}, err
	}

	// Fix interface's location, which is required when updating the interface.
	// TODO: is this a bug of azure SDK?
	if nic.Location == nil || *nic.Location == "" {
		nic.Location = vm.Location
	}

	return nic, nil
}

// getScaleSetWithRetry gets scale set with exponential backoff retry
func (ss *scaleSet) getScaleSetWithRetry(name string) (computepreview.VirtualMachineScaleSet, bool, error) {
	var result computepreview.VirtualMachineScaleSet
	var exists bool

	err := wait.ExponentialBackoff(ss.requestBackoff(), func() (bool, error) {
		cached, retryErr := ss.vmssCache.Get(name)
		if retryErr != nil {
			glog.Errorf("backoff: failure for scale set %q, will retry,err=%v", name, retryErr)
			return false, nil
		}
		glog.V(4).Info("backoff: success for scale set %q", name)

		if cached != nil {
			exists = true
			result = *(cached.(*computepreview.VirtualMachineScaleSet))
		}

		return true, nil
	})

	return result, exists, err
}

// getPrimaryNetworkConfiguration gets primary network interface configuration for scale sets.
func (ss *scaleSet) getPrimaryNetworkConfiguration(networkConfigurationList *[]computepreview.VirtualMachineScaleSetNetworkConfiguration, scaleSetName string) (*computepreview.VirtualMachineScaleSetNetworkConfiguration, error) {
	networkConfigurations := *networkConfigurationList
	if len(networkConfigurations) == 1 {
		return &networkConfigurations[0], nil
	}

	for idx := range networkConfigurations {
		networkConfig := &networkConfigurations[idx]
		if networkConfig.Primary != nil && *networkConfig.Primary == true {
			return networkConfig, nil
		}
	}

	return nil, fmt.Errorf("failed to find a primary network configuration for the scale set %q", scaleSetName)
}

func (ss *scaleSet) getPrimaryIPConfigForScaleSet(config *computepreview.VirtualMachineScaleSetNetworkConfiguration, scaleSetName string) (*computepreview.VirtualMachineScaleSetIPConfiguration, error) {
	ipConfigurations := *config.IPConfigurations
	if len(ipConfigurations) == 1 {
		return &ipConfigurations[0], nil
	}

	for idx := range ipConfigurations {
		ipConfig := &ipConfigurations[idx]
		if ipConfig.Primary != nil && *ipConfig.Primary == true {
			return ipConfig, nil
		}
	}

	return nil, fmt.Errorf("failed to find a primary IP configuration for the scale set %q", scaleSetName)
}

// createOrUpdateVMSSWithRetry invokes ss.VirtualMachineScaleSetsClient.CreateOrUpdate with exponential backoff retry.
func (ss *scaleSet) createOrUpdateVMSSWithRetry(virtualMachineScaleSet computepreview.VirtualMachineScaleSet) error {
	return wait.ExponentialBackoff(ss.requestBackoff(), func() (bool, error) {
		ctx, cancel := getContextWithCancel()
		defer cancel()
		resp, err := ss.VirtualMachineScaleSetsClient.CreateOrUpdate(ctx, ss.ResourceGroup, *virtualMachineScaleSet.Name, virtualMachineScaleSet)
		glog.V(10).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate(%s): end", *virtualMachineScaleSet.Name)
		return processHTTPRetryResponse(resp, err)
	})
}

// updateVMSSInstancesWithRetry invokes ss.VirtualMachineScaleSetsClient.UpdateInstances with exponential backoff retry.
func (ss *scaleSet) updateVMSSInstancesWithRetry(scaleSetName string, vmInstanceIDs computepreview.VirtualMachineScaleSetVMInstanceRequiredIDs) error {
	return wait.ExponentialBackoff(ss.requestBackoff(), func() (bool, error) {
		ctx, cancel := getContextWithCancel()
		defer cancel()
		resp, err := ss.VirtualMachineScaleSetsClient.UpdateInstances(ctx, ss.ResourceGroup, scaleSetName, vmInstanceIDs)
		glog.V(10).Infof("VirtualMachineScaleSetsClient.UpdateInstances(%s): end", scaleSetName)
		return processHTTPRetryResponse(resp, err)
	})
}

// EnsureHostsInPool ensures the given Node's primary IP configurations are
// participating in the specified LoadBalancer Backend Pool.
func (ss *scaleSet) EnsureHostsInPool(serviceName string, nodes []*v1.Node, backendPoolID string, vmSetName string) error {
	virtualMachineScaleSet, exists, err := ss.getScaleSetWithRetry(vmSetName)
	if err != nil {
		glog.Errorf("ss.getScaleSetWithRetry(%s) for service %q failed: %v", vmSetName, serviceName, err)
		return err
	}
	if !exists {
		errorMessage := fmt.Errorf("Scale set %q not found", vmSetName)
		glog.Errorf("%v", errorMessage)
		return errorMessage
	}

	// Find primary network interface configuration.
	networkConfigureList := virtualMachineScaleSet.VirtualMachineProfile.NetworkProfile.NetworkInterfaceConfigurations
	primaryNetworkConfiguration, err := ss.getPrimaryNetworkConfiguration(networkConfigureList, vmSetName)
	if err != nil {
		return err
	}

	// Find primary IP configuration.
	primaryIPConfiguration, err := ss.getPrimaryIPConfigForScaleSet(primaryNetworkConfiguration, vmSetName)
	if err != nil {
		return err
	}

	// Update primary IP configuration's LoadBalancerBackendAddressPools.
	foundPool := false
	newBackendPools := []computepreview.SubResource{}
	if primaryIPConfiguration.LoadBalancerBackendAddressPools != nil {
		newBackendPools = *primaryIPConfiguration.LoadBalancerBackendAddressPools
	}
	for _, existingPool := range newBackendPools {
		if strings.EqualFold(backendPoolID, *existingPool.ID) {
			foundPool = true
			break
		}
	}
	if !foundPool {
		newBackendPools = append(newBackendPools,
			computepreview.SubResource{
				ID: to.StringPtr(backendPoolID),
			})
		primaryIPConfiguration.LoadBalancerBackendAddressPools = &newBackendPools

		ctx, cancel := getContextWithCancel()
		defer cancel()
		glog.V(3).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate for service (%s): scale set (%s) - updating", serviceName, vmSetName)
		resp, err := ss.VirtualMachineScaleSetsClient.CreateOrUpdate(ctx, ss.ResourceGroup, vmSetName, virtualMachineScaleSet)
		glog.V(10).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate(%q): end", vmSetName)
		if ss.CloudProviderBackoff && shouldRetryHTTPRequest(resp, err) {
			glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate for service (%s): scale set (%s) - updating, err=%v", serviceName, vmSetName, err)
			retryErr := ss.createOrUpdateVMSSWithRetry(virtualMachineScaleSet)
			if retryErr != nil {
				err = retryErr
				glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate for service (%s) abort backoff: scale set (%s) - updating", serviceName, vmSetName)
			}
		}
		if err != nil {
			return err
		}
	}

	// Construct instanceIDs from nodes.
	instanceIDs := []string{}
	for _, curNode := range nodes {
		curScaleSetName, err := extractScaleSetNameByExternalID(curNode.Spec.ExternalID)
		if err != nil {
			glog.V(4).Infof("Node %q is not belonging to any scale sets, omitting it", curNode.Name)
			continue
		}
		if curScaleSetName != vmSetName {
			glog.V(4).Infof("Node %q is not belonging to scale set %q, omitting it", curNode.Name, vmSetName)
			continue
		}

		instanceID, err := getLastSegment(curNode.Spec.ExternalID)
		if err != nil {
			glog.Errorf("Failed to get last segment from %q: %v", curNode.Spec.ExternalID, err)
			return err
		}

		instanceIDs = append(instanceIDs, instanceID)
	}

	// Update instances to latest VMSS model.
	vmInstanceIDs := computepreview.VirtualMachineScaleSetVMInstanceRequiredIDs{
		InstanceIds: &instanceIDs,
	}
	ctx, cancel := getContextWithCancel()
	defer cancel()
	instanceResp, err := ss.VirtualMachineScaleSetsClient.UpdateInstances(ctx, ss.ResourceGroup, vmSetName, vmInstanceIDs)
	glog.V(10).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate(%q): end", vmSetName)
	if ss.CloudProviderBackoff && shouldRetryHTTPRequest(instanceResp, err) {
		glog.V(2).Infof("VirtualMachineScaleSetsClient.UpdateInstances for service (%s): scale set (%s) - updating, err=%v", serviceName, vmSetName, err)
		retryErr := ss.updateVMSSInstancesWithRetry(vmSetName, vmInstanceIDs)
		if retryErr != nil {
			err = retryErr
			glog.V(2).Infof("VirtualMachineScaleSetsClient.UpdateInstances for service (%s) abort backoff: scale set (%s) - updating", serviceName, vmSetName)
		}
	}
	if err != nil {
		return err
	}

	return nil
}

// EnsureBackendPoolDeleted ensures the loadBalancer backendAddressPools deleted from the specified vmSet.
func (ss *scaleSet) EnsureBackendPoolDeleted(poolID, vmSetName string) error {
	virtualMachineScaleSet, exists, err := ss.getScaleSetWithRetry(vmSetName)
	if err != nil {
		glog.Errorf("ss.EnsureBackendPoolDeleted(%s, %s) getScaleSetWithRetry(%s) failed: %v", poolID, vmSetName, vmSetName, err)
		return err
	}
	if !exists {
		glog.V(2).Infof("ss.EnsureBackendPoolDeleted(%s, %s), scale set %s has already been non-exist", poolID, vmSetName, vmSetName)
		return nil
	}

	// Find primary network interface configuration.
	networkConfigureList := virtualMachineScaleSet.VirtualMachineProfile.NetworkProfile.NetworkInterfaceConfigurations
	primaryNetworkConfiguration, err := ss.getPrimaryNetworkConfiguration(networkConfigureList, vmSetName)
	if err != nil {
		return err
	}

	// Find primary IP configuration.
	primaryIPConfiguration, err := ss.getPrimaryIPConfigForScaleSet(primaryNetworkConfiguration, vmSetName)
	if err != nil {
		return err
	}

	// Construct new loadBalancerBackendAddressPools and remove backendAddressPools from primary IP configuration.
	if primaryIPConfiguration.LoadBalancerBackendAddressPools == nil || len(*primaryIPConfiguration.LoadBalancerBackendAddressPools) == 0 {
		return nil
	}
	existingBackendPools := *primaryIPConfiguration.LoadBalancerBackendAddressPools
	newBackendPools := []computepreview.SubResource{}
	foundPool := false
	for i := len(existingBackendPools) - 1; i >= 0; i-- {
		curPool := existingBackendPools[i]
		if strings.EqualFold(poolID, *curPool.ID) {
			glog.V(10).Infof("EnsureBackendPoolDeleted gets unwanted backend pool %q for scale set %q", poolID, vmSetName)
			foundPool = true
			newBackendPools = append(existingBackendPools[:i], existingBackendPools[i+1:]...)
		}
	}
	if !foundPool {
		// Pool not found, assume it has been already removed.
		return nil
	}

	// Update scale set with backoff.
	primaryIPConfiguration.LoadBalancerBackendAddressPools = &newBackendPools
	glog.V(3).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate: scale set (%s) - updating", vmSetName)
	ctx, cancel := getContextWithCancel()
	defer cancel()
	resp, err := ss.VirtualMachineScaleSetsClient.CreateOrUpdate(ctx, ss.ResourceGroup, vmSetName, virtualMachineScaleSet)
	glog.V(10).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate(%q): end", vmSetName)
	if ss.CloudProviderBackoff && shouldRetryHTTPRequest(resp, err) {
		glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate: scale set (%s) - updating, err=%v", vmSetName, err)
		retryErr := ss.createOrUpdateVMSSWithRetry(virtualMachineScaleSet)
		if retryErr != nil {
			err = retryErr
			glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate abort backoff: scale set (%s) - updating", vmSetName)
		}
	}
	if err != nil {
		return err
	}

	// Update instances to latest VMSS model.
	instanceIDs := []string{"*"}
	vmInstanceIDs := computepreview.VirtualMachineScaleSetVMInstanceRequiredIDs{
		InstanceIds: &instanceIDs,
	}
	instanceResp, err := ss.VirtualMachineScaleSetsClient.UpdateInstances(ctx, ss.ResourceGroup, vmSetName, vmInstanceIDs)
	glog.V(10).Infof("VirtualMachineScaleSetsClient.UpdateInstances(%q): end", vmSetName)
	if ss.CloudProviderBackoff && shouldRetryHTTPRequest(instanceResp, err) {
		glog.V(2).Infof("VirtualMachineScaleSetsClient.UpdateInstances scale set (%s) - updating, err=%v", vmSetName, err)
		retryErr := ss.updateVMSSInstancesWithRetry(vmSetName, vmInstanceIDs)
		if retryErr != nil {
			err = retryErr
			glog.V(2).Infof("VirtualMachineScaleSetsClient.UpdateInstances abort backoff: scale set (%s) - updating", vmSetName)
		}
	}
	if err != nil {
		return err
	}

	// Update virtualMachineScaleSet again. This is a workaround for removing VMSS reference from LB.
	// TODO: remove this workaround when figuring out the root cause.
	if len(newBackendPools) == 0 {
		glog.V(3).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate: scale set (%s) - updating second time", vmSetName)
		resp, err = ss.VirtualMachineScaleSetsClient.CreateOrUpdate(ctx, ss.ResourceGroup, vmSetName, virtualMachineScaleSet)
		glog.V(10).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate(%q): end", vmSetName)
		if ss.CloudProviderBackoff && shouldRetryHTTPRequest(resp, err) {
			glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate: scale set (%s) - updating, err=%v", vmSetName, err)
			retryErr := ss.createOrUpdateVMSSWithRetry(virtualMachineScaleSet)
			if retryErr != nil {
				glog.V(2).Infof("VirtualMachineScaleSetsClient.CreateOrUpdate abort backoff: scale set (%s) - updating", vmSetName)
			}
		}
	}

	return nil
}

// getVmssMachineID returns the full identifier of a vmss virtual machine.
func (az *Cloud) getVmssMachineID(scaleSetName, instanceID string) string {
	return fmt.Sprintf(
		vmssMachineIDTemplate,
		az.SubscriptionID,
		az.ResourceGroup,
		scaleSetName,
		instanceID)
}