We’ve successfully launched Ray Autoscaler on Kubernetes on AWS. However, we’d like to have ray autoscaler support multiple instance types (similar to how the node types work when running the autoscaler natively on AWS). For example, we’ve created node selectors for instance types in k8s via:
# c5.2xlarge
nodeSelector:
beta.kubernetes.io/instance-type: c5.2xlarge
# p2.xlarge
nodeSelector:
beta.kubernetes.io/instance-type: p2.xlarge
However, we are unsure how to link those to ray custom resources, such that if I call an actor with resources p2.xlarge=1
, ray will trigger EKS to autoscale the pods to fill those demands. Is that supported feature, and how do we enable that in the config?