As a customer, I would like to deploy OpenShift On OpenStack, using the IPI workflow where my control plane would have 3 machines and each machine would have use a root volume (a Cinder volume attached to the Nova server) and also an attached ephemeral disk using local storage, that would only be used by etcd.

As this feature will be TechPreview in 4.15, this will only be implemented as a day 2 operation for now. This might or might not change in the future.

We know that etcd requires storage with strong performance capabilities and currently a root volume backed by Ceph has difficulties to provide these capabilities.

By also attaching local storage to the machine and mounting it for etcd would solve the performance issues that we saw when customers were using Ceph as the backend for the control plane disks.

Gophercloud already accepts to create a server with multiple ephemeral disks:

https://github.com/gophercloud/gophercloud/blob/master/openstack/compute/v2/extensions/bootfromvolume/doc.go#L103-L151

We need to figure out how we want to address that in CAPO, probably involving a new API; that later would be used in openshift (MAPO, and probably installer).

We'll also have to update the OpenStack Failure Domain in CPMS.

ARO (Azure) has conducted some benckmarks and is now recommending to put etcd on a separated data disk:

https://docs.google.com/document/d/1O_k6_CUyiGAB_30LuJFI6Hl93oEoKQ07q1Y7N2cBJHE/edit

Also interesting thread: https://groups.google.com/u/0/a/redhat.com/g/aos-devel/c/CztJzGWdsSM/m/jsPKZHSRAwAJ