Pin autoscaling requester (ray-project#42254)

Pin the Ray Data autoscaling requester to the head node so that the actor cannot prevent down-scaling of an idle worker node. Signed-off-by: Stephanie Wang <swang@cs.berkeley.edu>
ROCm · Jan 12, 2024 · 66d2367 · 66d2367
1 parent 8058cf1
commit 66d2367
Showing 1 changed file with 1 addition and 2 deletions.
diff --git a/python/ray/data/_internal/execution/autoscaling_requester.py b/python/ray/data/_internal/execution/autoscaling_requester.py
@@ -119,8 +119,7 @@ def get_or_create_autoscaling_requester_actor():
     # point to the head node.
     scheduling_strategy = NodeAffinitySchedulingStrategy(
         ray.get_runtime_context().get_node_id(),
-        soft=True,
-        _spill_on_unavailable=True,
+        soft=False,
     )
     with _autoscaling_requester_lock:
         return AutoscalingRequester.options(