HA Architecture for 12 nodes


Rancher recommend to split workers from controleplane and etcs roles (https://rancher.com/docs/rancher/v2.x/en/cluster-provisioning/production/)

I have a 12 nodes cluster with 40 CPU and 128 GO memory each.

For the moment, i have 3 nodes with controleplane, etcd and worker roles and 9 nodes with worker role.

What should I do to have a better architecture ?

  • can i use external virtual machines to host controleplane and etcd roles ?
  • 3 controleplane/etcd are enough for a 12 nodes cluster ?
  • is it really bad to mix worker role with etcd and controleplane ?



There’s a best practices guide that covers this to some extent.

To answer your questions:

  • Yes, it’s good to put controlplane and etcd on VMs.
  • Yes, three etcd nodes is a good number. Chris Kim, an engineer at Rancher, says “the number of etcd nodes you have for a cluster actually really only increases your fault tolerance, rather than increasing performance. Conventionally speaking, 3 is the magic number because you need a majority of nodes live and talking in order to maintain quorum, and I wouldn’t necessarily recommend scaling past 5 or 7”
  • It’s better to put worker roles on a separate host so that you can get more scalability

Hello Catherine,

Thanks for the answers.