Autoscaling in non-default namespaces is broken

[josephburnett]

The ela-autoscaler service account, which the autoscaler uses to update deployment sizes, is created in the default namespace. This works fine when revisions are also created in the default namespace, but it breaks when otherwise.

An error shows when trying to schedule the autoscaler pods. This doesn't break the normal serving path, but it prevents the revision from scaling up. And it slows down scheduling enough to cause the conformance tests to timeout in various places.

[josephburnett]

To fix this, we should either 1) create a service account per namespace or 2) make the autoscaler a single, multi-tenant component in the ela-system namespace.

[josephburnett]

or 3) move all single-tenant autoscalers to the ela-system namespace.

I think we should go with the third option. To do so, we must move the service account into ela-system: https://github.com/elafros/elafros/blob/8ef2aeebd9dbc2730996a3d03546e970887bd3a6/serviceaccount.yaml#L25 and create the autoscaler deployement: https://github.com/elafros/elafros/blob/8ef2aeebd9dbc2730996a3d03546e970887bd3a6/pkg/controller/revision/ela_autoscaler.go#L39 and service: https://github.com/elafros/elafros/blob/8ef2aeebd9dbc2730996a3d03546e970887bd3a6/pkg/controller/revision/ela_autoscaler.go#L103 in the ela-namespace too. We will also have to point the queue at the autoscaler in the new namespace: https://github.com/elafros/elafros/blob/8ef2aeebd9dbc2730996a3d03546e970887bd3a6/cmd/ela-queue/main.go#L98

[markusthoemmes]

Is this still a valid issue given the referenced PR is merged?