Flyte enables production-grade orchestration for machine learning workflows and data processing created to accelerate local workflows to production.

Flyte

<https://github.com/flyteorg/flyte/issues/7005|#7005 Support Ray AutoscalerOptions in RayCluster spec>
Issue created by <https://github.com/pingsutw|pingsutw>
*Motivation*

Users who need to configure Ray autoscaler behavior (e.g., environment variables, upscaling mode, resource limits) currently cannot do so because the Flyte Ray integration doesn't pass `AutoscalerOptions` through to the KubeRay `RayCluster` CR.
Currently, only `enable_autoscaling` (bool) and per-worker `min_replicas`/`max_replicas` are supported. The KubeRay `RayCluster` CRD has a rich <https://docs.ray.io/en/latest/cluster/kubernetes/user-guides/configuring-autoscaling.html|`autoscalerOptions`> field that allows configuring:
• *`upscalingMode`* — `Default`, `Aggressive`, or `Conservative`
• *`env`* — environment variables for the autoscaler sidecar container
• *`resources`* — CPU/memory requests and limits for the autoscaler container
• *`image`* — custom autoscaler container image
• *`idleTimeoutSeconds`* — how long idle workers are kept before scale-down

*Proposal*

Expose these options end-to-end across the Flyte stack:

*1. flyteidl (protobuf definitions)*

Add an `AutoscalerOptions` message to <https://github.com/flyteorg/flyte/blob/master/flyteidl/protos/flyteidl/plugins/ray.proto|`flyteidl/plugins/ray.proto`>:

message AutoscalerOptions {
  string upscaling_mode = 1;           // "Default", "Aggressive", "Conservative"
  int32 idle_timeout_seconds = 2;
  repeated EnvVar env = 3;             // autoscaler sidecar env vars
  string image = 4;                    // custom autoscaler image
  Resources resources = 5;             // autoscaler container resources
}

Add a field to the existing `RayCluster` message:

message RayCluster {
  ...
  AutoscalerOptions autoscaler_options = N;
}

*2. flytepropeller (backend Ray plugin)*

Update the <https://github.com/flyteorg/flyte/tree/master/flyteplugins/go/tasks/plugins/k8s/ray|Ray plugin handler> to:
• Read `AutoscalerOptions` from the task's protobuf `RayCluster` message
• Map it onto the KubeRay `RayCluster` CR's `.spec.autoscalerOptions`

*3. flytekit / flyte-sdk (Python SDK)*

In the Ray plugin (`plugins/ray/src/flyteplugins/ray/task.py` in <https://github.com/flyteorg/flyte-sdk|flyte-sdk>):
• Add an `AutoscalerOptions` dataclass:

@dataclass
class AutoscalerOptions:
    upscaling_mode: Optional[str] = None        # "Default", "Aggressive", "Conservative"
    idle_timeout_seconds: Optional[int] = None
    env: Optional[Dict[str, str]] = None
    image: Optional[str] = None
    resources: Optional[Resources] = None

• Add `autoscaler_options: Optional[AutoscalerOptions] = None` to `RayJobConfig`
• Update `RayFunctionTask.custom_config()` to serialize autoscaler options into the `RayCluster` protobuf
Currently `RayJobConfig` only has:

enable_autoscaling: bool = False

The new field would sit alongside it:

enable_autoscaling: bool = False
autoscaler_options: Optional[AutoscalerOptions] = None

*Goal*

Allow users to write:

RayJobConfig(
    enable_autoscaling=True,
    autoscaler_options=AutoscalerOptions(
        upscaling_mode="Conservative",
        idle_timeout_seconds=120,
        image="rayproject/ray:2.9.0",
    ),
    worker_node_config=[...],
)

And have KubeRay receive the full `autoscalerOptions` on the resulting `RayCluster` CR.

*Are you sure this issue hasn't been raised already?*

Yes

*Have you read the Code of Conduct?*

Yes
<https://github.com/flyteorg/flyte|flyteorg/flyte>