diff --git a/python/ray/data/_internal/datasource/parquet_datasource.py b/python/ray/data/_internal/datasource/parquet_datasource.py
index b688c2630d68..3b092168707f 100644
--- a/python/ray/data/_internal/datasource/parquet_datasource.py
+++ b/python/ray/data/_internal/datasource/parquet_datasource.py
@@ -23,6 +23,7 @@
 from ray.data._internal.util import (
     _check_pyarrow_version,
     _is_local_scheme,
+    _split_list,
     call_with_retry,
     iterate_with_retry,
 )
@@ -354,9 +355,9 @@ def get_read_tasks(self, parallelism: int) -> List[ReadTask]:
 
         read_tasks = []
         for fragments, paths, metadata in zip(
-            np.array_split(pq_fragments, parallelism),
-            np.array_split(pq_paths, parallelism),
-            np.array_split(pq_metadata, parallelism),
+            _split_list(pq_fragments, parallelism),
+            _split_list(pq_paths, parallelism),
+            _split_list(pq_metadata, parallelism),
         ):
             if len(fragments) <= 0:
                 continue
diff --git a/python/ray/data/tests/test_util.py b/python/ray/data/tests/test_util.py
index b66a9bc5804f..a9bdad3cdc94 100644
--- a/python/ray/data/tests/test_util.py
+++ b/python/ray/data/tests/test_util.py
@@ -1,3 +1,4 @@
+import time
 from typing import Any, Dict, Optional
 
 import numpy as np
@@ -150,6 +151,23 @@ def test_list_splits():
     assert _split_list(["foo", 1, [0], None], 3) == [["foo", 1], [[0]], [None]]
 
 
+def split_performance():
+    int_array = range(1000003)
+    start = time.perf_counter()
+
+    for split in np.array_split(int_array, 100):
+        len(split)
+
+    end1 = time.perf_counter()
+    print(f"============== {end1 - start}")
+
+    for split in _split_list(int_array, 100):
+        len(split)
+    end2 = time.perf_counter()
+    print(f"============== {end2 - end1}")
+    assert (end2 - end1) < (end1 - start)
+
+
 def get_parquet_read_logical_op(
     ray_remote_args: Optional[Dict[str, Any]] = None,
     **read_kwargs,