diff --git a/python/src/space/core/ops/append.py b/python/src/space/core/ops/append.py
index 6c0f84c..761cce6 100644
--- a/python/src/space/core/ops/append.py
+++ b/python/src/space/core/ops/append.py
@@ -84,12 +84,12 @@ def __init__(self, location: str, metadata: meta.StorageMetadata):
     StoragePaths.__init__(self, location)
 
     self._metadata = metadata
-    self._schema = arrow.arrow_schema(self._metadata.schema.fields)
-
+    record_fields = set(self._metadata.schema.record_fields)
+    self._schema = arrow.arrow_schema(self._metadata.schema.fields,
+                                      record_fields,
+                                      physical=True)
     self._index_fields, self._record_fields = arrow.classify_fields(
-        self._schema,
-        set(self._metadata.schema.record_fields),
-        selected_fields=None)
+        self._schema, record_fields, selected_fields=None)
 
     # Data file writers.
     self._index_writer_info: Optional[_IndexWriterInfo] = None
diff --git a/python/src/space/core/schema/arrow.py b/python/src/space/core/schema/arrow.py
index 898c240..8f7961b 100644
--- a/python/src/space/core/schema/arrow.py
+++ b/python/src/space/core/schema/arrow.py
@@ -20,9 +20,9 @@
 import pyarrow as pa
 from substrait.type_pb2 import NamedStruct, Type
 
-from space.core.utils.constants import UTF_8
 from space.core.schema import constants
 from space.core.schema.types import TfFeatures
+from space.core.utils.constants import UTF_8
 
 _PARQUET_FIELD_ID_KEY = b"PARQUET:field_id"
 
@@ -49,11 +49,13 @@ def next(self) -> str:
     return name
 
 
-def arrow_schema(fields: NamedStruct) -> pa.Schema:
+def arrow_schema(fields: NamedStruct, record_fields: Set[str],
+                 physical: bool) -> pa.Schema:
   """Return Arrow schema from Substrait fields.
   
   Args:
     fields: schema fields in the Substrait format.
+    record_fields: a set of record field names.
     physical: if true, return the physical schema. Physical schema matches with
       the underlying index (Parquet) file schema. Record fields are stored by
       their references, e.g., row position in ArrayRecord file.
@@ -61,19 +63,28 @@ def arrow_schema(fields: NamedStruct) -> pa.Schema:
   return pa.schema(
       _arrow_fields(
           _NamesVisitor(fields.names),  # type: ignore[arg-type]
-          fields.struct.types))  # type: ignore[arg-type]
+          fields.struct.types,  # type: ignore[arg-type]
+          record_fields,
+          physical))
 
 
-def _arrow_fields(names_visitor: _NamesVisitor,
-                  types: List[Type]) -> List[pa.Field]:
+def _arrow_fields(names_visitor: _NamesVisitor, types: List[Type],
+                  record_fields: Set[str], physical: bool) -> List[pa.Field]:
   fields: List[pa.Field] = []
 
   for type_ in types:
     name = names_visitor.next()
-    arrow_field = pa.field(name,
-                           _arrow_type(type_, names_visitor),
-                           metadata=field_metadata(_substrait_field_id(type_)))
-    fields.append(arrow_field)
+
+    if physical and name in record_fields:
+      arrow_type: pa.DataType = pa.struct(
+          record_address_types())  # type: ignore[arg-type]
+    else:
+      arrow_type = _arrow_type(type_, names_visitor)
+
+    fields.append(
+        pa.field(name,
+                 arrow_type,
+                 metadata=field_metadata(_substrait_field_id(type_))))
 
   return fields
 
diff --git a/python/src/space/core/schema/types/tf_features.py b/python/src/space/core/schema/types/tf_features.py
index 1e3dcbd..c183000 100644
--- a/python/src/space/core/schema/types/tf_features.py
+++ b/python/src/space/core/schema/types/tf_features.py
@@ -15,7 +15,7 @@
 """Define a custom Arrow type for Tensorflow Dataset Features."""
 
 from __future__ import annotations
-from typing import Any
+from typing import Any, Union
 
 import json
 import pyarrow as pa
@@ -47,9 +47,12 @@ def __arrow_ext_serialize__(self) -> bytes:
   def __arrow_ext_deserialize__(
       cls,
       storage_type: pa.DataType,  # pylint: disable=unused-argument
-      serialized: bytes) -> TfFeatures:
-    return TfFeatures(
-        f.FeaturesDict.from_json(json.loads(serialized.decode(UTF_8))))
+      serialized: Union[bytes, str]
+  ) -> TfFeatures:
+    if isinstance(serialized, bytes):
+      serialized = serialized.decode(UTF_8)
+
+    return TfFeatures(f.FeaturesDict.from_json(json.loads(serialized)))
 
   def serialize(self, value: Any) -> bytes:
     """Serialize value using the provided features_dict."""
diff --git a/python/tests/core/manifests/test_record.py b/python/tests/core/manifests/test_record.py
index b254915..a77be0f 100644
--- a/python/tests/core/manifests/test_record.py
+++ b/python/tests/core/manifests/test_record.py
@@ -21,7 +21,6 @@
 class TestRecordManifestWriter:
 
   def test_write(self, tmp_path):
-    data_dir = tmp_path / "dataset" / "data"
     metadata_dir = tmp_path / "dataset" / "metadata"
     metadata_dir.mkdir(parents=True)
 
diff --git a/python/tests/core/ops/test_append.py b/python/tests/core/ops/test_append.py
index bd4b67b..8e4ab95 100644
--- a/python/tests/core/ops/test_append.py
+++ b/python/tests/core/ops/test_append.py
@@ -12,16 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import List
+import numpy as np
 import pyarrow as pa
 import pyarrow.parquet as pq
+from tensorflow_datasets import features  # type: ignore[import-untyped]
 
 from space.core.ops import LocalAppendOp
 import space.core.proto.metadata_pb2 as meta
+from space.core.schema.types import TfFeatures
 from space.core.storage import Storage
 
 
 class TestLocalAppendOp:
 
+  # TODO: to add tests using Arrow table input.
+
   def test_write_pydict_all_types(self, tmp_path):
     location = tmp_path / "dataset"
     schema = pa.schema([
@@ -75,6 +81,82 @@ def test_write_pydict_all_types(self, tmp_path):
     assert patch.storage_statistics_update == meta.StorageStatistics(
         num_rows=5, index_compressed_bytes=114, index_uncompressed_bytes=126)
 
+  def test_write_pydict_with_record_fields(self, tmp_path):
+    tf_features_images = features.FeaturesDict(
+        {"images": features.Image(shape=(None, None, 3), dtype=np.uint8)})
+    tf_features_objects = features.FeaturesDict({
+        "objects":
+        features.Sequence({
+            "bbox": features.BBoxFeature(),
+            "id": np.int64
+        }),
+    })
+
+    location = tmp_path / "dataset"
+    schema = pa.schema([
+        pa.field("int64", pa.int64()),
+        pa.field("string", pa.string()),
+        pa.field("images", TfFeatures(tf_features_images)),
+        pa.field("objects", TfFeatures(tf_features_objects))
+    ])
+    storage = Storage.create(location=str(location),
+                             schema=schema,
+                             primary_keys=["int64"],
+                             record_fields=["images", "objects"])
+
+    op = LocalAppendOp(str(location), storage.metadata)
+
+    op.write({
+        "int64": [1, 2, 3],
+        "string": ["a", "b", "c"],
+        "images": [b"images0", b"images1", b"images2"],
+        "objects": [b"objects0", b"objects1", b"objects2"]
+    })
+    op.write({
+        "int64": [0, 10],
+        "string": ["A", "z"],
+        "images": [b"images3", b"images4"],
+        "objects": [b"objects3", b"objects4"]
+    })
+
+    patch = op.finish()
+    assert patch is not None
+
+    # Validate index manifest files.
+    index_manifest = self._read_manifests(
+        storage, list(patch.added_index_manifest_files))
+    assert index_manifest == {
+        "_FILE": index_manifest["_FILE"],
+        "_INDEX_COMPRESSED_BYTES": [114],
+        "_INDEX_UNCOMPRESSED_BYTES": [126],
+        "_NUM_ROWS": [5],
+        "_STATS_f0": [{
+            "_MAX": 10,
+            "_MIN": 0
+        }]
+    }
+
+    # Validate record manifest files.
+    record_manifest = self._read_manifests(
+        storage, list(patch.added_record_manifest_files))
+    assert record_manifest == {
+        "_FILE": record_manifest["_FILE"],
+        "_FIELD_ID": [2, 3],
+        "_NUM_ROWS": [5, 5],
+        "_UNCOMPRESSED_BYTES": [55, 60]
+    }
+
+    # Data file exists.
+    self._check_file_exists(location, index_manifest["_FILE"])
+    self._check_file_exists(location, record_manifest["_FILE"])
+
+    # Validate statistics.
+    assert patch.storage_statistics_update == meta.StorageStatistics(
+        num_rows=5,
+        index_compressed_bytes=114,
+        index_uncompressed_bytes=126,
+        record_uncompressed_bytes=115)
+
   def test_empty_op_return_none(self, tmp_path):
     location = tmp_path / "dataset"
     schema = pa.schema([pa.field("int64", pa.int64())])
@@ -85,3 +167,15 @@ def test_empty_op_return_none(self, tmp_path):
 
     op = LocalAppendOp(str(location), storage.metadata)
     assert op.finish() is None
+
+  def _read_manifests(self, storage: Storage,
+                      file_paths: List[str]) -> pa.Table:
+    manifests = []
+    for f in file_paths:
+      manifests.append(pq.read_table(storage.full_path(f)))
+
+    return pa.concat_tables(manifests).to_pydict()
+
+  def _check_file_exists(self, location, file_paths: List[str]):
+    for f in file_paths:
+      assert (location / f).exists()
diff --git a/python/tests/core/schema/conftest.py b/python/tests/core/schema/conftest.py
index 4205696..58d4f5f 100644
--- a/python/tests/core/schema/conftest.py
+++ b/python/tests/core/schema/conftest.py
@@ -12,12 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import pytest
+import json
 
+import numpy as np
 import pyarrow as pa
+import pytest
+from tensorflow_datasets import features  # type: ignore[import-untyped]
 from substrait.type_pb2 import NamedStruct, Type
 
 from space.core.schema.arrow import field_metadata
+from space.core.schema.types import TfFeatures
 
 
 @pytest.fixture
@@ -92,3 +96,31 @@ def sample_arrow_schema():
                ]),
                metadata=field_metadata(260))
   ])
+
+
+@pytest.fixture
+def tf_features():
+  return features.FeaturesDict(
+      {"images": features.Image(shape=(None, None, 3), dtype=np.uint8)})
+
+
+@pytest.fixture
+def tf_features_substrait_fields(tf_features):  # pylint: disable=redefined-outer-name
+  return NamedStruct(
+      names=["int64", "features"],
+      struct=Type.Struct(types=[
+          Type(i64=Type.I64(type_variation_reference=0)),
+          Type(user_defined=Type.UserDefined(type_parameters=[
+              Type.Parameter(string="TF_FEATURES"),
+              Type.Parameter(string=json.dumps(tf_features.to_json()))
+          ],
+                                             type_variation_reference=1))
+      ]))
+
+
+@pytest.fixture
+def tf_features_arrow_schema(tf_features):  # pylint: disable=redefined-outer-name
+  return pa.schema([
+      pa.field("int64", pa.int64(), metadata=field_metadata(0)),
+      pa.field("features", TfFeatures(tf_features), metadata=field_metadata(1))
+  ])
diff --git a/python/tests/core/schema/test_arrow.py b/python/tests/core/schema/test_arrow.py
index 420c220..9d7c409 100644
--- a/python/tests/core/schema/test_arrow.py
+++ b/python/tests/core/schema/test_arrow.py
@@ -15,6 +15,7 @@
 import pyarrow as pa
 
 from space.core.schema import arrow
+from space.core.schema.arrow import field_metadata
 
 
 def test_field_metadata():
@@ -27,8 +28,33 @@ def test_field_id():
                                              b"123"})) == 123
 
 
-def test_arrow_schema(sample_substrait_fields, sample_arrow_schema):
-  assert sample_arrow_schema == arrow.arrow_schema(sample_substrait_fields)
+def test_arrow_schema_logical_without_records(sample_substrait_fields,
+                                              sample_arrow_schema):
+  assert arrow.arrow_schema(sample_substrait_fields, [],
+                            False) == sample_arrow_schema
+
+
+def test_arrow_schema_logical_with_records(tf_features_substrait_fields,
+                                           tf_features_arrow_schema):
+  assert arrow.arrow_schema(tf_features_substrait_fields, [],
+                            False) == tf_features_arrow_schema
+
+
+def test_arrow_schema_physical_without_records(sample_substrait_fields,
+                                               sample_arrow_schema):
+  assert arrow.arrow_schema(sample_substrait_fields, [],
+                            True) == sample_arrow_schema
+
+
+def test_arrow_schema_physical_with_records(tf_features_substrait_fields):
+  arrow_schema = pa.schema([
+      pa.field("int64", pa.int64(), metadata=field_metadata(0)),
+      pa.field("features",
+               pa.struct([("_FILE", pa.string()), ("_ROW_ID", pa.int32())]),
+               metadata=field_metadata(1))
+  ])
+  assert arrow.arrow_schema(tf_features_substrait_fields, ["features"],
+                            True) == arrow_schema
 
 
 def test_field_name_to_id_dict(sample_arrow_schema):
diff --git a/python/tests/core/schema/types/test_tf_features.py b/python/tests/core/schema/types/test_tf_features.py
index 1f988d6..5d0d019 100644
--- a/python/tests/core/schema/types/test_tf_features.py
+++ b/python/tests/core/schema/types/test_tf_features.py
@@ -53,10 +53,16 @@ def test_arrow_ext_serialize_deserialize(self, tf_features, sample_objects):
         "type"] == "tensorflow_datasets.core.features.features_dict.FeaturesDict"  # pylint: disable=line-too-long
     assert "sequence" in features_dict["content"]["features"]["objects"]
 
+    # Bytes input.
     tf_features = TfFeatures.__arrow_ext_deserialize__(storage_type=None,
                                                        serialized=serialized)
     assert len(tf_features.serialize(sample_objects)) > 0
 
+    # String input.
+    tf_features = TfFeatures.__arrow_ext_deserialize__(
+        storage_type=None, serialized=serialized.decode(UTF_8))
+    assert len(tf_features.serialize(sample_objects)) > 0
+
   def test_serialize_deserialize(self, tf_features, sample_objects):
     value_bytes = tf_features.serialize(sample_objects)
     assert len(value_bytes) > 0