In layer 2 mode, one node assumes the responsibility of advertising a service to the local network. From the network’s perspective, it simply looks like that machine has multiple IP addresses assigned to its network interface.
The major advantage of the layer 2 mode is its universality: it will work on any Ethernet network, with no special hardware required, not even fancy routers.
In layer 2 mode, all traffic for a service IP goes to one node. From there,
kube-proxy spreads the traffic to all the service’s pods.
In that sense, layer 2 does not implement a load balancer. Rather, it implements a failover mechanism so that a different node can take over should the current leader node fail for some reason.
If the leader node fails for some reason, failover is automatic: the failed node is detected using memberlist, at which point new nodes take over ownership of the IP addresses from the failed node.
Layer 2 mode has two main limitations you should be aware of: single-node bottlenecking, and potentially slow failover.
As explained above, in layer2 mode a single leader-elected node receives all traffic for a service IP. This means that your service’s ingress bandwidth is limited to the bandwidth of a single node. This is a fundamental limitation of using ARP and NDP to steer traffic.
In the current implementation, failover between nodes depends on cooperation from the clients. When a failover occurs, MetalLB sends a number of gratuitous layer 2 packets (a bit of a misnomer - it should really be called “unsolicited layer 2 packets”) to notify clients that the MAC address associated with the service IP has changed.
Most operating systems handle “gratuitous” packets correctly, and update their neighbor caches promptly. In that case, failover happens within a few seconds. However, some systems either don’t implement gratuitous handling at all, or have buggy implementations that delay the cache update.
All modern versions of major OSes (Windows, Mac, Linux) implement layer 2 failover correctly, so the only situation where issues may happen is with older or less common OSes.
To minimize the impact of planned failover on buggy clients, you should keep the old leader node up for a couple of minutes after flipping leadership, so that it can continue forwarding traffic for old clients until their caches refresh.
During an unplanned failover, the service IPs will be unreachable until the buggy clients refresh their cache entries.
If you encounter a situation where layer 2 mode failover is slow (more than about 10s), please file a bug! We can help you investigate and determine if the issue is with the client, or a bug in MetalLB.
The election of the “leader” (the node which is going to advertise the IP) of a given loadbalancer IP is stateless and works in the following way:
This removes the need of having to keep memory of which speaker is in charge of announcing a given IP.
Given the leader election algoritm described above, removing a node does not change the speaker announcing the VIP, while adding a node will change it only if it becomes the new first element of the list.
Given the stateless nature of the mechanism, if a speaker mistakenly detects a set of nodes as non active, it might calculate a different list, resulting in multiple (or no) speakers announcing the same VIP.
MetalLB’s layer2 mode has a lot of similarities to Keepalived, so if you’re familiar with Keepalived, this should all sound fairly familiar. However, there are also a few differences worth mentioning. If you aren’t familiar with Keepalived, you can skip this section.
Keepalived uses the Virtual Router Redundancy Protocol (VRRP). Instances of Keepalived continuously exchange VRRP messages with each other, both to select a leader and to notice when that leader goes away.
MetalLB on the other hand relies on memberlist to know when a node in the cluster is no longer reachable and the service IPs from that node should be moved elsewhere.
Keepalived and MetalLB “look” the same from the client’s perspective: the service IP address seems to migrate from one machine to another when a failover occurs, and the rest of the time it just looks like machines have more than one IP address.
Because it doesn’t use VRRP, MetalLB isn’t subject to some of the limitations of that protocol. For example, the VRRP limit of 255 load balancers per network doesn’t exist in MetalLB. You can have as many load-balanced IPs as you want, as long as there are free IPs in your network. MetalLB also requires less configuration than VRRP–for example, there are no Virtual Router IDs.
On the flip side, because MetalLB relies on memberlist for cluster membership information, it cannot interoperate with third-party VRRP-aware routers and infrastructure. This is working as intended: MetalLB is specifically designed to provide load balancing and failover within a Kubernetes cluster, and in that scenario interoperability with third-party LB software is out of scope.