The module implements an improved (relative to the standard L2 mode in MetalLB) balancing mechanism for services in bare metal clusters when there is no option to use cloud load balancers or MetalLB in BGP mode with Equal-cost multi-path (ECMP) configured.
Principle of operation compared to L2 mode in MetalLB module
MetalLB in L2 mode allows to order Service with LoadBalancer type, the operation of which is based on the fact that balancing nodes simulate ARP-responses from the “public” IP in a peering network. This mode has a significant limitation — only one balancing node handles all the incoming traffic of this service at a time. Therefore:
- The node selected as the leader for the “public” IP becomes a “bottleneck” with no possibility of horizontal scaling.
- If the balancer node fails, all current connections will be dropped for switching to a new balancing node that will be selected as the leader.
This module helps bypass these limitations. It provides a new L2LoadBalancer interface that:
- Allows linking a group of nodes and an IP address pool using
nodeSelector.
After that, we can create a standard Service resource with the type LoadBalancer and specify the following using annotations:
- Which L2LoadBalancer to use, thereby defining the set of nodes and addresses.
- Specify the required number of IP addresses for L2 advertisement.
Thus:
- The application will receive not a single, but several (according to the number of balancer nodes) “public” IPs. These IPs will need to be configured as A-records for the application’s public domain. For further horizontal scaling, additional balancer nodes will need to be added, the corresponding Service will be created automatically, you just need to add them to the list of A-records for the application domain.
- If one of the balancer nodes fails, only a part of the connections will be prone to failover to the healthy node.