[Bug] Allow zero replica for workers for Helm

[ducviet00]

Why are these changes needed?

We are currently using Ray for computing heavily tasks on GKE. When initializing, it spawns a worker each worker group. Then, it triggers GKE scale up node. It's money cost.

This happens because ternary function in template file. {{ 0 | 1 }} = 1

kuberay/helm-chart/ray-cluster/templates/raycluster-cluster.yaml

Line 91 in 87dde22

minReplicas: {{ $values.minReplicas | default (default 1 $values.miniReplicas) }}

kuberay/helm-chart/ray-cluster/templates/raycluster-cluster.yaml

Line 157 in 87dde22

minReplicas: {{ .Values.worker.minReplicas | default (default 1 .Values.worker.miniReplicas) }}

workaround by setting default replica to zero.

Related issue number

Open #965

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

[kevin85421]

Thank you for the contribution! I am wondering is there any difference for your use case between disabled: true and replicas: 0?

[ducviet00]

Thank you for the contribution! I am wondering is there any difference for your use case between disabled: true and replicas: 0?

as I understand, minReplicas: 0 allow scaling down pod to zero and disabled: true doesn't allow scale up pods.

[ducviet00]

I think setting replica: 0 as the default is better than setting replica: 1 because we shouldn't create a replica initially. A worker requests much memory, so it's a waste of resources. The autoscaler will handle it based on the job's resource demand. Setting minReplicas: 0 to allows it and allows scale down to zero when no job running.

[ducviet00]

@kevin85421 Could you make a review?

[kevin85421]

Test this PR manually using this gist.

# Step 0: Replace values.yaml with the gist
# (path: helm-chart/ray-cluster)
helm install ray-cluster .

# Step 1: Try to scale up the cluster
export HEAD_POD=$(kubectl get pods -o custom-columns=POD:metadata.name | grep raycluster-autoscaler-head)
kubectl exec $HEAD_POD -it -c ray-head -- python -c "import ray;ray.init();ray.autoscaler.sdk.request_resources(num_cpus=4)"

# Step 2: The RayCluster will scale from 0 worker to 3 workers.

[yc2984]

@kevin85421 is this available on 0.5.2?

[yc2984]

@kevin85421 is this available on 0.5.2?

I see it's only on 0.6.0. Is it stable or still WIP?