Azure LoadBalancer

Azure LoadBalancer basics.

The way Azure defines a LoadBalancer is different from GCE or AWS. Azure’s LB can have multiple frontend IP refs. GCE and AWS only allow one, if you want more, you would need multiple LBs. Since Public IP’s are not part of the LB in Azure, an NSG is not part of the LB in Azure either. However, you cannot delete them in parallel, a Public IP can only be deleted after the LB’s frontend IP ref is removed.

The different Azure Resources such as LB, Public IP, and NSG are the same tier of Azure resources and circular dependencies need to be avoided. In other words, they should only depend on service state.

By default the basic SKU is selected for a load balancer. Services can be annotated to allow auto selection of available load balancers. Service annotations can also be used to provide specific availability sets that host the load balancers. Note that in case of auto selection or specific availability set selection, services are currently not auto-reassigned to an available loadbalancer when the availability set is lost in case of downtime or cluster scale down.

LoadBalancer annotations

Below is a list of annotations supported for Kubernetes services with type LoadBalancer:

AnnotationValueDescriptionKubernetes Version or falseSpecify whether the load balancer should be internal. It’s defaulting to public if not set.v1.10.0 and later of the subnetSpecify which subnet the internal load balancer should be bound to. It’s defaulting to the subnet configured in cloud config file if not set.v1.10.0 and later, {vmset-name}Specify the Azure load balancer selection algorithm based on vm sets (VMSS or VMAS). There are currently three possible load balancer selection modes : default, auto or “{vmset-name}”. This is only working for basic LB or multiple standard LB (see below for how it works)v1.10.0 and later of the PIP DNS labelSpecify the DNS label name for the service’s public IP address (PIP). If it is set to empty string, DNS in PIP would be deleted. Because of a bug, before v1.15.10/v1.16.7/v1.17.3, the DNS label on PIP would also be deleted if the annotation is not specified.v1.15.0 and later or falseSpecify that the service should be exposed using an Azure security rule that may be shared with another service, trading specificity of rules for an increase in the number of services that can be exposed. This relies on the Azure “augmented security rules” feature.v1.10.0 and later of the PIP resource groupSpecify the resource group of the service’s PIP that are not in the same resource group as the cluster.v1.10.0 and later of allowed service tagsSpecify a list of allowed service tags separated by comma.v1.11.0 and later idle timeouts in minutesSpecify the time, in minutes, for TCP connection idle timeouts to occur on the load balancer. Default and minimum value is 4. Maximum value is 30. Must be an integer.v1.11.4, v1.12.0 and later of PIPSpecify the PIP that will be applied to load balancer.v1.16 and later of Public IP PrefixSpecify the Public IP Prefix that will be applied to load balancer.v1.21 and later with out-of-tree cloud provider of the PIPSpecify the tags of the PIP that will be associated to the load balancer typed service. Docv1.20 and later probe intervalRefer to the detailed docs herev1.21 and later with out-of-tree cloud provider minimum number of unhealthy responses of health probeRefer to the detailed docs herev1.21 and later with out-of-tree cloud provider path of the health probeRefer to the detailed docs herev1.20 and later with out-of-tree cloud provider{port}_health-probe_intervalHealth probe interval{port} is port number of service. Refer to the detailed docs herev1.21 and later with out-of-tree cloud provider{port}_health-probe_num-of-probeThe minimum number of unhealthy responses of health probe{port} is port number of service. Refer to the detailed docs herev1.21 and later with out-of-tree cloud provider{port}_health-probe_request-pathRequest path of the health probe{port} is port number of service. Refer to the detailed docs herev1.20 and later with out-of-tree cloud provider high availability ports on internal SLBHA ports is required when applications require IP fragmentsv1.20 and later or falseDeny all traffic to the service. This is helpful when the service.Spec.LoadBalancerSourceRanges is set to an internal load balancer typed service. When set the loadBalancerSourceRanges field on the service in order to whitelist ip src addresses, although the generated NSG has added the rules for loadBalancerSourceRanges, the default rule (65000) will allow any vnet traffic, basically meaning the whitelist is of no use. This annotation solves this issue.v1.21 and later public IPs besides the service’s own public IPIt is mainly used for global VIP on Azure cross-region LoadBalancerv1.20 and later with out-of-tree cloud provider or falseDisable Floating IP configuration for load balancerv1.21 and later with out-of-tree cloud provider

Please note that

  • When loadBalancerSourceRanges have been set on service spec, won’t work because of DROP iptables rules from kube-proxy. The CIDRs from service tags should be merged into loadBalancerSourceRanges to make it work.

Load balancer selection modes

There are currently three possible load balancer selection modes :

  1. Default mode - service has no annotation (“”). In this case the Loadbalancer of the primary Availability set is selected
  2. auto” mode - service is annotated with __auto__ value. In this case, services would be associated with the Loadbalancer with the minimum number of rules.
  3. “{vmset-name}” mode - service is annotated with the name of a VMSS/VMAS. In this case, only load balancers of the specified VMSS/VMAS would be selected, and services would be associated with the one with the minimum number of rules.

Note that the “auto” mode is valid only if the service is newly created. It is not allowed to change the annotation value to __auto__ of an existed service.

The selection mode for a load balancer only works for basic load balancers or multiple standard load balancers. Following is the detailed information of allowed number of VMSS/VMAS in a load balancer.

  • Standard SKU supports any virtual machine in a single virtual network, including a mix of virtual machines, availability sets, and virtual machine scale sets. So all the nodes would be added to the same standard LB backend pool with a max size of 1000.
  • Basic SKU only supports virtual machines in a single availability set, or a virtual machine scale set. Only nodes with the same availability set or virtual machine scale set would be added to the basic LB backend pool.

LoadBalancer SKUs

Azure cloud provider supports both basic and standard SKU load balancers, which can be set via loadBalancerSku option in cloud config file. A list of differences between these two SKUs can be found here.

Note that the public IPs used in load balancer frontend configurations should be the same SKU. That is a standard SKU public IP for standard load balancer and a basic SKU public IP for a basic load balancer.

Azure doesn’t support a network interface joining load balancers with different SKUs, hence migration dynamically between them is not supported.

If you do require migration, please delete all services with type LoadBalancer (or change to other type)

Outbound connectivity

Outbound connectivity is also different between the two load balancer SKUs:

  • For the basic SKU, the outbound connectivity is opened by default. If multiple frontends are set, then the outbound IP is selected randomly (and configurable) from them.

  • For the standard SKU, the outbound connectivity is disabled by default. There are two ways to open the outbound connectivity: use a standard public IP with the standard load balancer or define outbound rules.

Standard LoadBalancer

Because the load balancer in a Kubernetes cluster is managed by the Azure cloud provider, and it may change dynamically (e.g. the public load balancer would be deleted if no services defined with type LoadBalancer), outbound rules are the recommended path if you want to ensure the outbound connectivity for all nodes.

Especially note:

  • In the context of outbound connectivity, a single standalone VM, all the VM’s in an Availability Set, all the instances in a VMSS behave as a group. This means, if a single VM in an Availability Set is associated with a Standard SKU, all VM instances within this Availability Set now behave by the same rules as if they are associated with Standard SKU, even if an individual instance is not directly associated with it.

  • Public IP’s used as instance-level public IP are mutually exclusive with outbound rules.

Here is the recommended way to define the outbound rules when using separate provisioning tools:

  • Create a separate IP (or multiple IPs for scale) in a standard SKU for outbound rules. Make use of the allocatedOutboundPorts parameter to allocate sufficient ports for your desired scenario scale.
  • Create a separate pool definition for outbound, and ensure all virtual machines or VMSS virtual machines are in this pool. Azure cloud provider will manage the load balancer rules with another pool, so that provisioning tools and the Azure cloud provider won’t affect each other.
  • Define inbound with load balancing rules and inbound NAT rules as needed, and set disableOutboundSNAT to true on the load balancing rule(s). Don’t rely on the side effect from these rules for outbound connectivity. It makes it messier than it needs to be and limits your options. Use inbound NAT rules to create port forwarding mappings for SSH access to the VM’s rather than burning public IPs per instance.

Exclude nodes from the load balancer

Excluding nodes from Azure LoadBalancer is supported since v1.20.0.

The kubernetes controller manager supports excluding nodes from the load balancer backend pools by enabling the feature gate ServiceNodeExclusion. To exclude nodes from Azure LoadBalancer, label should be added to the nodes.

  1. To use the feature, the feature gate ServiceNodeExclusion should be on (enabled by default since its beta on v1.19).

  2. The labeled nodes would be excluded from the LB in the next LB reconcile loop, which needs one or more LB typed services to trigger. Basically, users could trigger the update by creating a service. If there are one or more LB typed services existing, no extra operations are needed.

  3. To re-include the nodes, just remove the label and the update would be operated in the next LB reconcile loop.

Using SCTP

SCTP protocol services are only supported on internal standard LoadBalancer, hence annotation "true" should be added to SCTP protocol services. See below for an example:

apiVersion: v1
kind: Service
  name: sctpservice
  annotations: "true"
  type: LoadBalancer
    app: sctpserver
    - name: sctpserver
      protocol: SCTP
      port: 30102
      targetPort: 30102

Multiple Standard LoadBalancer per cluster

This feature is supported since v1.20.0.

There is only one external and one internal Standard Load Balancer (SLB) at most per cluster. Set enableMultipleStandardLoadBalancers=true in the cloud config if you want to turn on the multiple SLB mode. Similar to the basic LB, there will be a 1:1 mapping between each SLB and VMSS/VMAS. The SLB of the primary VMSS/VMAS will be named after clusterName in the cloud config (in AKS, it would be kubernetes) while the name of those belonging to non-primary VMSS/VMAS will be the name of the corresponding vmSet.

If the cluster provisioning tools like ASK-Engine and CAPZ don’t proactively create a dedicated SLB for each VMSS/VMAS when enabling multiple SLB, only the primary SLB would be created. You could manually trigger the creation by setting the service annotation to bind the service to that VMSS/VMAS. The dedicated SLB would be created once the service reconcile loop is done. Unlike the primary SLB, there is no default outbound rules/IPs for the non-primary SLBs. That means the SLB would be deleted once all the services referencing it are deleted.

Choose which SLB to use

The service annotation will be respected as long as enableMultipleStandardLoadBalancers=true when using standard LB, and the usage is the same as it is in the basic LB clusters. Specifically, there are three selection mode: default to select the primary SLB; __auto__ to select the SLB with minimum rules and vmSetName to select the dedicated SLB of that VMSS/VMAS.

Outbound Connections of non-primary VMSS/VMAS

The outbound rules of the non-primary SLB are not managed by cloud provider azure. Instead, it should be managed by cluster provisioning tools. For now, there is no outbound configuration for the non-primary VMSS/VMAS, but we plan to support customized outbound configurations in AKS and CAPZ in the future.

Sharing the primary SLB with multiple VMSS/VMAS

This feature is supported since v1.21.0

For each non-primary VMSS/VMAS, one can determine to use dedicated SLB or share the primary SLB. If the VMSS/VMAS names are in the cloud config nodepoolsWithoutDedicatedSLB, those would join the backend pool of the primary SLB while the others would remain to have dedicated SLBs. If the VMSS/VMAS supposed to share the primary SLB owns a dedicated SLB, the dedicated one would be deleted, and the VMSS/VMAS would be joint the primary SLB’s backend pool.

Custom Load Balancer health probe

As documented here, Tcp, Http and Https are three protocols supported by load balancer service.

Currently, the default protocol of the health probe varies among services with different transport protocols, app protocols, annotations and external traffic policies.

  1. for local services, HTTP and /healthz would be used. The health probe will query NodeHealthPort rather than actual backend service
  2. for cluster TCP services, TCP would be used.
  3. for cluster UDP services, no health probes.

Since v1.20, service annotation is introduced to determine the health probe behavior.

  • For clusters <=1.23, spec.ports.appProtocol would only be used as probe protocol when is also set.
  • For clusters >1.24, spec.ports.appProtocol would be used as probe protocol and / would be used as default probe request path ( could be used to change to a different request path).

Note that the request path would be ignored when using TCP or the spec.ports.appProtocol is empty. More specifically:

loadbalancer Probe ProtocolLB Probe Request Path
standardclustertcphttp/httpsTCP(<=1.23) or http/https(>=1.24)null(<=1.23) or /(>=1.24)
standardclustertcpunsupported protocol/custom-pathtcpnull
basicclustertcphttpTCP(<=1.23) or http/https(>=1.24)null(<=1.23) or /(>=1.24)
basicclustertcpunsupported protocol/custom-pathtcpnull

Since v1.21, two service annotations and load-balancer-health-probe-num-of-probe are introduced, which customize the configuration of health probe. If is not set, Default value of 5 is applied. If load-balancer-health-probe-num-of-probe is not set, Default value of 2 is applied. And total probe should be less than 120 seconds.

Custom Load Balancer health probe for port

Because MixedProtocolLBService feature is in alpha stage, Ports in one service may have different probe configurations. Following annotations are introduced to customize probe configuration for one port.

port specific annotationglobal probe annotation{port}{port}{port}

For following manifest, probe rule for port httpsserver is different from the one for httpserver because annoations for port httpsserver are specified.

apiVersion: v1
kind: Service
  name: appservice
  annotations: "5" "4"
  type: LoadBalancer
    app: server
    - name: httpserver
      protocol: TCP
      port: 80
      targetPort: 30102
    - name: httpsserver
      protocol: TCP
      appProtocol: HTTPS
      port: 443
      targetPort: 30104

Configure Load Balancer backend

This feature is supported since v1.23.0

The backend pool type can be configured by specifying loadBalancerBackendPoolConfigurationType in the cloud configuration file. There are three possible values:

  1. nodeIPConfiguration (default). In this case we attach nodes to the LB by calling the VMSS/NIC API to associate the corresponding node IP configuration with the LB backend pool.
  2. nodeIP. In this case we attach nodes to the LB by calling the LB API to add the node private IP addresses to the LB backend pool.
  3. podIP (not supported yet). In this case we do not attach nodes to the LB. Instead we directly adding pod IPs to the LB backend pool.

Load balancer limits

The limits of the load balancer related resources are listed below:

Standard Load Balancer

Load balancers1,000
Rules per resource1,500
Rules per NIC (across all IPs on a NIC)300
Frontend IP configurations600
Backend pool size1,000 IP configurations, single virtual network
Backend resources per Load Balancer150
High-availability ports1 per internal frontend
Outbound rules per Load Balancer600
Load Balancers per VM2 (1 Public and 1 internal)

The limit is up to 150 resources, in any combination of standalone virtual machine resources, availability set resources, and virtual machine scale-set placement groups.

Basic Load Balancer

Load balancers1,000
Rules per resource250
Rules per NIC (across all IPs on a NIC)300
Frontend IP configurations200
Backend pool size300 IP configurations, single availability set
Availability sets per Load Balancer1
Load Balancers per VM2 (1 Public and 1 internal)

There is a restriction of 300 rules per NIC, hence for single SLB mode 300 services are allowed at most. If more services are required, try to enable multiple SLBs.

Last modified July 1, 2022 : Update load balancer doc (627a1a264)