They say the real value of Kubernetes is its uniform declarative style, extensibility with CRDs, consistent generated clients with the ability to watch, and not necessarily its scheduler or container runtimes [citation needed].
Let's see if that's true, by running a nodeless Kube API server on Cloud Run, that's capable of handling CRUD API requests, accepting new CRD definitions, the works, but will never have any Nodes, so any Pods will just remaining pending forever. Then let's see if that's actually useful for anything. If not, maybe try to find a way to run Pods.
This would require implementing the etcd API, possibly also on Cloud Run, backed by Datastore, which is itself backed by Spanner so it's consistent. This might be all that's really required, I don't know.
You could also validate this theory by running a GKE cluster without any node pools, but that still incurs the $.10/hour (~$7/month) cluster management fee. Not serverless enough!
They say the real value of Kubernetes is its uniform declarative style, extensibility with CRDs, consistent generated clients with the ability to watch, and not necessarily its scheduler or container runtimes [citation needed].
Let's see if that's true, by running a nodeless Kube API server on Cloud Run, that's capable of handling CRUD API requests, accepting new CRD definitions, the works, but will never have any Nodes, so any Pods will just remaining pending forever. Then let's see if that's actually useful for anything. If not, maybe try to find a way to run Pods.
This would require implementing the etcd API, possibly also on Cloud Run, backed by Datastore, which is itself backed by Spanner so it's consistent. This might be all that's really required, I don't know.
You could also validate this theory by running a GKE cluster without any node pools, but that still incurs the $.10/hour (~$7/month) cluster management fee. Not serverless enough!