diff --git a/mars/dataframe/utils.py b/mars/dataframe/utils.py
index 8958a0b485..f0ebfb3fba 100644
--- a/mars/dataframe/utils.py
+++ b/mars/dataframe/utils.py
@@ -33,10 +33,6 @@
 from ..core import Entity, ExecutableTuple
 from ..core.context import Context, get_context
 from ..lib.mmh3 import hash as mmh_hash
-from ..services.task.execution.ray.context import (
-    RayExecutionContext,
-    RayExecutionWorkerContext,
-)
 from ..tensor.utils import dictify_chunk_size, normalize_chunk_sizes
 from ..typing import ChunkType, TileableType
 from ..utils import (
@@ -46,7 +42,7 @@
     ModulePlaceholder,
     is_full_slice,
     parse_readable_size,
-    is_ray_address,
+    is_on_ray,
 )
 
 try:
@@ -1437,6 +1433,9 @@ def _concat_chunks(merge_chunks: List[ChunkType], output_index: int):
     return new_op.new_tileable(df_or_series.op.inputs, kws=[params])
 
 
+# TODO: clean_up_func, is_on_ray and restore_func functions may be 
+# removed or refactored in the future to calculate func size 
+# with more accuracy as well as address some serialization issues.
 def clean_up_func(op):
     closure_clean_up_bytes_threshold = int(
         os.getenv("MARS_CLOSURE_CLEAN_UP_BYTES_THRESHOLD", 10**4)
@@ -1461,7 +1460,7 @@ def clean_up_func(op):
             op.logic_key is not None
         ), f"Logic key of {op} wasn't calculated before cleaning up func."
         logger.debug(f"{op} need cleaning up func.")
-        if _is_on_ray(ctx):
+        if is_on_ray(ctx):
             import ray
 
             op.func_key = ray.put(op.func)
@@ -1470,27 +1469,10 @@ def clean_up_func(op):
             op.func = cloudpickle.dumps(op.func)
 
 
-def _is_on_ray(ctx):
-    # There are three conditions
-    #   a. mars backend
-    #   b. ray backend(oscar), c. ray backend(dag)
-    # When a. or b. is selected, ctx is an instance of ThreadedServiceContext.
-    #   The main difference between them is whether worker_address matches ray scheme.
-    #   To avoid duplicated checks, here we choose the first worker address.
-    # When c. is selected, ctx is an instance of RayExecutionContext or RayExecutionWorkerContext,
-    #   while get_worker_addresses method isn't currently implemented in RayExecutionWorkerContext.
-    try:
-        worker_addresses = ctx.get_worker_addresses()
-    except AttributeError:  # pragma: no cover
-        assert isinstance(ctx, RayExecutionWorkerContext)
-        return True
-    return isinstance(ctx, RayExecutionContext) or is_ray_address(worker_addresses[0])
-
-
 def restore_func(ctx: Context, op):
     if op.need_clean_up_func and ctx is not None:
         logger.debug(f"{op} need restoring func.")
-        if _is_on_ray(ctx):
+        if is_on_ray(ctx):
             import ray
 
             op.func = ray.get(op.func_key)
diff --git a/mars/deploy/oscar/tests/test_ray_dag_oscar.py b/mars/deploy/oscar/tests/test_ray_dag_oscar.py
index 41d28c8c6c..dc658527b4 100644
--- a/mars/deploy/oscar/tests/test_ray_dag_oscar.py
+++ b/mars/deploy/oscar/tests/test_ray_dag_oscar.py
@@ -52,3 +52,23 @@ async def test_iterative_tiling(ray_start_regular_shared2, create_cluster):
 @require_ray
 async def test_execute_describe(ray_start_regular_shared2, create_cluster):
     await test_local.test_execute_describe(create_cluster)
+
+
+@require_ray
+@pytest.mark.parametrize(
+    "create_cluster",
+    [
+        {
+            "config": {
+                "task.task_preprocessor_cls": "mars.deploy.oscar.tests.test_clean_up_and_restore_func.RayBackendFuncTaskPreprocessor",
+                "subtask.subtask_processor_cls": "mars.deploy.oscar.tests.test_clean_up_and_restore_func.RayBackendFuncSubtaskProcessor",
+            }
+        }
+    ],
+    indirect=True,
+)
+@pytest.mark.asyncio
+async def test_ray_dag_oscar_clean_up_and_restore_func(
+    ray_start_regular_shared2, create_cluster
+):
+    await test_local.test_execute_apply_closure(create_cluster)
\ No newline at end of file
diff --git a/mars/utils.py b/mars/utils.py
index 3ce87e8523..a392f2575e 100644
--- a/mars/utils.py
+++ b/mars/utils.py
@@ -1701,6 +1701,30 @@ def is_ray_address(address: str) -> bool:
         return False
 
 
+# TODO: clean_up_func, is_on_ray and restore_func functions may be 
+# removed or refactored in the future to calculate func size 
+# with more accuracy as well as address some serialization issues.
+def is_on_ray(ctx):
+    from .services.task.execution.ray.context import (
+        RayExecutionContext,
+        RayExecutionWorkerContext,
+    )
+    # There are three conditions
+    #   a. mars backend
+    #   b. ray backend(oscar), c. ray backend(dag)
+    # When a. or b. is selected, ctx is an instance of ThreadedServiceContext.
+    #   The main difference between them is whether worker address matches ray scheme.
+    #   To avoid duplicated checks, here we choose the first worker address.
+    # When c. is selected, ctx is an instance of RayExecutionContext or RayExecutionWorkerContext,
+    #   while get_worker_addresses method isn't currently implemented in RayExecutionWorkerContext.
+    try:
+        worker_addresses = ctx.get_worker_addresses()
+    except AttributeError:  # pragma: no cover
+        assert isinstance(ctx, RayExecutionWorkerContext)
+        return True
+    return isinstance(ctx, RayExecutionContext) or is_ray_address(worker_addresses[0])
+
+
 def cache_tileables(*tileables):
     from .core import ENTITY_TYPE