Issue evidence · Run · ray-project/ray

Actor-pool scale-down leaves primary object-store pins, so worker nodes never go idle

GitHub issue: https://github.com/ray-project/ray/issues/66903

Reproduced

After Ray Data's actor pool scales down, objects left in the object store keep the node from ever being reported idle, so an autoscaler never removes it.

What Badgr ran

Ran the issue's script unchanged on Ray 2.59.0 (the issue reports 2.56.1) on a local 3-node cluster, locally on CPU.

What came back

  • Actors died with "all references to the actor were removed" while the dataset was still running.
  • A worker node with CPU use 0.0 showed about 15.7 MB of object store in use, idle_duration_ms 0, and its only activity "Resource: object_store_memory currently in use."

Reproduced on a newer Ray than the issue names. The KubeRay behavior in the issue was not tested.

Checked 2026-10-11 in Badgr’s local development environment.

← All issue evidence