[PlacementGroup]Fix bug that kill workers mistakenly when gcs restarts (#12568)

python/ray/tests/test_gcs_fault_tolerance.py

@@ -36,9 +36,6 @@ def test_gcs_server_restart(ray_start_regular):
     ray.worker._global_node.kill_gcs_server()
     ray.worker._global_node.start_gcs_server()
 
-    result = ray.get(actor1.method.remote(7))
-    assert result == 9
-
     actor2 = Increase.remote()
     result = ray.get(actor2.method.remote(2))
     assert result == 4
@@ -46,6 +43,12 @@ def test_gcs_server_restart(ray_start_regular):
     result = ray.get(increase.remote(1))
     assert result == 2
 
+    # Check whether actor1 is alive or not.
+    # NOTE: We can't execute it immediately after gcs restarts
+    # because it takes time for the worker to exit.
+    result = ray.get(actor1.method.remote(7))
+    assert result == 9
+
 
 @pytest.mark.parametrize(
     "ray_start_regular", [

python/ray/tests/test_placement_group.py

@@ -1249,7 +1249,7 @@ def test_create_placement_group_during_gcs_server_restart(
     cluster.head_node.kill_gcs_server()
     cluster.head_node.start_gcs_server()
 
-    for i in range(0, 10):
+    for i in range(0, 100):
         ray.get(placement_groups[i].ready())