Implement AtomicScaleUp ProvisioningClass

[yaroslava-serdiuk]

The Atomic-scale-up Provisioning Class aims to provision the resources required for the specified pods in an atomic way.

Tasks

Beta Give feedback

1. Update API to support AtomicScaleUp (the implementation of the method is a cloud provider responsibility, if the method is not implemented, fallback to IncreaseSize)
2. Implement ScaleUp for Provisioning Class
3. Update ProvisioningRequest Injector
4. Update ProvisioningRequest Processor
5. Fix ProvisioningRequest booked capacity is scaled down (and shouldn't be) #6517

[yaroslava-serdiuk]

@aleksandra-malinowska is already working on ScaleUp implementation.
/assign @aleksandra-malinowska

[aleksandra-malinowska]

/cc @kisieland

[aleksandra-malinowska]

@yaroslava-serdiuk regarding the 'update ProvisioningRequest injector', I assume you meant adding the new class to supported classes list. But it doesn't seem like this list isn't used anywhere (we don't validate if the request is of the supported class in injector). Is this WAI?

I can see why we wouldn't validate class here - we should take into account the requests that were provisioned and will cause a workload to be admitted - but I'm not sure if this was a conscious decision. LMK if I should add validation or just clean up this var.

Relevant code: https://github.com/kubernetes/autoscaler/blob/master/cluster-autoscaler/processors/provreq/provisioning_request_injector.go#L41C5-L41C33

[yaroslava-serdiuk]

I think supported classes list is actually a leftover. We validate provisioning class separately in ScaleUp loop, so it's ok to remove the const to me. I think this additional validation is not needed (and it's easy forget to update the list if we add another prov class).

we should take into account the requests that were provisioned

If ProvReq is failed or Provisioned we do not inject pods: https://github.com/kubernetes/autoscaler/blob/master/cluster-autoscaler/processors/provreq/provisioning_request_injector.go#L60

[Shubham82]

/area cluster-autoscaler
/area core-autoscaler

[yaroslava-serdiuk]

/assign