Fix un-merged frames

distributed/protocol/core.py

@@ -51,11 +51,14 @@ def _encode_default(obj):
                 if typ is Serialize:
                     obj = obj.data
                 offset = len(frames)
-                sub_header, sub_frames = serialize_and_split(
-                    obj, serializers=serializers, on_error=on_error, context=context
-                )
+                if typ is Serialized:
+                    sub_header, sub_frames = obj.header, obj.frames
+                else:
+                    sub_header, sub_frames = serialize_and_split(
+                        obj, serializers=serializers, on_error=on_error, context=context
+                    )
+                    _inplace_compress_frames(sub_header, sub_frames)
                 sub_header["num-sub-frames"] = len(sub_frames)
-                _inplace_compress_frames(sub_header, sub_frames)
                 frames.append(
                     msgpack.dumps(
                         sub_header, default=msgpack_encode_default, use_bin_type=True

distributed/protocol/tests/test_serialize.py

@@ -13,7 +13,7 @@
 
 from dask.utils_test import inc
 
-from distributed import wait
+from distributed import Nanny, wait
 from distributed.comm.utils import from_frames, to_frames
 from distributed.protocol import (
     Serialize,
@@ -499,3 +499,17 @@ def test_ser_memoryview_object():
     data_in = memoryview(np.array(["hello"], dtype=object))
     with pytest.raises(TypeError):
         serialize(data_in, on_error="raise")
+
+
+@gen_cluster(client=True, Worker=Nanny)
+async def test_large_pickled_object(c, s, a, b):
+    np = pytest.importorskip("numpy")
+
+    class Data:
+        def __init__(self, n):
+            self.data = np.empty(n, dtype="u1")
+
+    x = Data(100_000_000)
+    y = await c.scatter(x, workers=[a.worker_address])
+    z = c.submit(lambda x: x, y, workers=[b.worker_address])
+    await z