In a Kubernetes cluster, depending on the flow of traffic direction, there are two kinds of load balancing:
- Load balancing for the traffic within a Kubernetes cluster
- Load balancing for the traffic coming from outside the cluster
Load balancing for internal traffic
Load balancing within a Kubernetes cluster is accessed through a ClusterIP
service type.
ClusterIP
presents a single IP address to the client and load balances the traffic going to this IP to the backend servers.
The actual load balancing happens using iptables
or IPVS configuration. The kube-proxy
Kubernetes component programs these.
The iptables
mode of operation uses DNAT rules to distribute direct trafik to real servers, whereas IPVS leverages in-kernel transport-layer load-balancing.
A comparision between these two methods can be found here.
By default, kube-proxy
runs in iptables
mode.
The kube-proxy configuration can be altered by updating the kube-proxy
configmap in the kube-system
namespace.
Load balancing for external traffic
External traffic destined for the Kubernetes service requires a service of type LoadBalancer
, through which external clients connect to your internal service.
Under the hood, it uses a load balancer provided by the underlying infrastructre to direct the traffic.
In the Cloud
In cloud deployments, the load balancer is provided by the cloud provider. For example, in AWS, the service controller communicates with the AWS API to provision an ELB service which targets the cluster nodes.
On-premises
For an on-premises deployment, Kommander ships with MetalLB, which provides load-balancing services.
kubectl -n kommander delete pod -l app=metallb,component=controller
To use MetalLB:
-
Identify and reserve a virtual IP (VIP) address range in your networking infrastructure.
-
Configure your networking infrastructure so that the reserved IP addresses is reachable:
- from all hosts specified in the inventory file.
- from the computer used to deploy Kubernetes.
Your configuration is complete if the reserved virtual IP addresses are in the same subnet as the rest of the cluster nodes. If it is in a different subnet you may need to configure appropriate routes to ensure connectivity with the virtual IP address. If the virtual IP addresses share an interface with the primary IP address of the interface, you must disable any IP or MAC spoofing from the infrastructure firewall.
MetalLB can be configured in two modes: Layer2 and BGP.
The following example illustrates how to enable MetalLB and configure it with the Layer2 mode:
---
apiVersion: v1
kind: ConfigMap
metadata:
name: metallb-overrides
data:
values.yaml: |
configInline:
address-pools:
- name: default
protocol: layer2
addresses:
- 10.0.50.25-10.0.50.50
---
apiVersion: apps.kommander.d2iq.io/v1alpha1
kind: AppDeployment
metadata:
name: metallb
spec:
appRef:
name: metallb-0.12.2
configOverrides:
name: metallb-overrides
The number of virtual IP addresses in the reserved range determines the maximum number of LoadBalancer
service types you can create in the cluster.
MetalLB in bgp
mode implements only a subset of the BGP protocol. In particular, it only advertises the virtual IP to peer BGP agent.
The following example illustrates the BGP configuration in the overrides ConfigMap
:
apiVersion: v1
kind: ConfigMap
metadata:
name: metallb-overrides
data:
values.yaml: |
configInline:
peers:
- my-asn: 64500
peer-asn: 64500
peer-address: 172.17.0.4
address-pools:
- name: my-ip-space
protocol: bgp
addresses:
- 172.40.100.0/24
In the above configuration, peers
defines the configuration of the BGP peer, such as peer ip address and autonomous system number
(asn
).
The address-pools
section is similar to layer2
, except for the protocol.
MetalLB also supports advanced BGP configuration, which can be found here.