enable setting failureDomain as part of a cluster's topology

[yastij]

User Story

As a user would like to reference a clusterClass in a cluster, and configure the failure domain of workers in its topology, to be able to re-use ClusterClass more broadly

Detailed Description

Today we don't offer users the ability to configure failureDomain for the various machineDeployment that is part of a ClusterClass which limits a bit the re-usability. We likely also need to define what is the default behaviour is when not setting such field in the cluster's topology.

Anything else you would like to add:

not the same, but #3358 might impact how we do this. We've had asks from CAPV users to spread machineDeployments across failure Domains (e.g. spread an MD that is GPU enabled)

/kind feature

[yastij]

cc @fabriziopandini @CecileRobertMichon

[fabriziopandini]

/area topology
/milestone v1.1
@yastij if I got this right it is your intention to work on this, right?

[fabriziopandini]

Should we consider also failure domains for control plane machines?

[vincepri]

/priority important-soon

[yastij]

@fabriziopandini - yeah, planning to work on this. As for "should we include control plane machines?" I'm not sure we want to.

The options for CP to me are:

1. not specify anything and keep today's behaviour which read from the cluster's status
2. add a failureDomain field, that might not achieve what we want as it'll colocate everything
3. add a list of failureDomains, which is a subset of what's listed in the cluster's status. It gives more control to the user if they don't want to consume an AZ listed in the status, but that would need changes to KCP.

for ControlPlane I'm leaning towards 1, for Workers I'm leaning toward 3 as it has real use cases and value. Also not against aligning options for both (for the sake of consistency)

[yastij]

/assign

[vincepri]

+1 to keep control plane delegated to the control plane provider

[CecileRobertMichon]

+1

[fabriziopandini]

/unassing @yastij
/assign