Control plane availability

[somehowchris]

In general I really like karmada but there's 1 bugging me.

Is the karmada control plane itself HA?
-> generally having a small outage over a few minutes/hours isn't really bad [interms of control of the cluster ] but as more and more things are managed via karmada, how does one keep up with leader election in pull mode? is there even something like leader election? also in the case of taking down a node i.e. migrating between zones, how does one demote an old host cluster to have another one (preferably automatically) be prompoted?

[whitewindmills]

it's a good question, but currently Karmada control plane does not support HA. we are also considering the HA solution for the Karmada control plane.

[RainbowMango]

From the Karmada users practice, Karmada control plane HA could be achieved in the following two ways:

1. All Karmada components could be run with leader election mode, which means people can run multiple replicas for each component, just like the practice of how we run Kubernetes in HA mode. This is a common way.
2. Run multiple Karmada components, a member cluster might be registered to more than one Karmada control plane, so even in case of one Karmada downs, people can still manage the cluster as well as the resources by another Karmada instance. This approach is still less practiced, but it does exist.

[somehowchris]

Thanks for the swift reply.

Interesting approaches. I have read about the leader election flags but didn't have any further context, is there a page or something like a blog explaining a bit more? This for sure would be lovely to read and tinker more about.

About the second approach, so to my understanding, karmada only propagates ressources which it is told to share like deployments or pods. So with that in mind how would other karamda clusters have the same context? Also there is the FedereatedHPA, how would one keep track which karmada host cluster schedules which scaling?

[RainbowMango]

Unfortunately, Karmada doesn't have specific documents for HA yet in addition to general installation. Thanks for spotting this, I just opened an issue to track this.

Before the official document, i can briefly share some ideas about how to achieve HA for Karmada.
If you have experience deploying HA Kubernetes cluster, then deploying HA Karmada will be very simple and similar:

• karmada-apiserver: it's essentially a kube-apiserver.
• karmada-controller-manager: similar to kube-controller-manager.
• karmada-scheduler: similar to kube-scheduler.
• so on
  The working mechanism of Karmada components is similar to Kubernetes, I don't think there is anything would block Karmada work in HA mode.

The second approach may be a higher level of HA and may exceed the scope of this conversation. On the other hand, I will also gather feedback from users and hope to explain this practice through an official doc.