Hello Drew-

If you could specify an anti-affinity node selector
<https://kubernetes.io/docs/concepts/configuration/assign-pod-node/#affinity-and-anti-affinity>
would that work?

Brandon

On Mon, Nov 13, 2017 at 7:51 PM Drew Wells <[email protected]> wrote:

> We want to spread etcd deployed by etcd-operator across different AZs. In
> theory, our etcd cluster could then survive a node going down since no more
> than 1 etcd pod would be running on it.
>
>
>  As an example, if we have kubernetes nodes in 3 different AZs: AZ1, AZ2,
> AZ3.
>
> Now I create a deployment with size: 3 for etcd. The desired behavior is
> this pods are created: etcd-0001 on AZ1, etcd-0002 on AZ2, and etcd-003 on
> AZ3. Basically, we desire per pod taints that decrease the change an etcd
> pod will be scheduled with a peer on the same node. I've searched through
> the database for a hack to make this happen. It doesn't appear possible to
> do per pod behavior like this.
>
> Are there recommendations for getting better reliability of an etcd
> cluster when a AZ  becomes unavailable outside of what's been described in
> this example?
>
> Best,
> Drew
>
>
>
>
> --
CTO, CoreOS, Inc
Tectonic is enterprise Kubernetes
https://coreos.com/tectonic

Reply via email to