<rekado_>I guess for AWS a reasonable strategy would be: i3.metal instance (autoscaling to add more when needed), firecracker (instead of qemu), shared read-only /gnu, and then launch VMs with Guix VM scripts.
<rekado_>on top of that some DNS thing to re-route networking to individual instances.
<rekado_>more expensive than simpler EC2 instances, but at scale it comes out cheaper and faster to have VMs share a big server and overcommit resources.
<civodul>rekado_: the phrase "container orchestrator" always sounded a bit scary to me
<civodul>but coming from you, it sounds exciting :-)
<civodul>overall i think as a project we're failing to communicate with sysadmins and devops, and that's a shame