fix: add missing .attrs support in parquet serialisation

src/awkward/operations/ak_from_parquet.py

@@ -268,7 +268,10 @@ def _load(
     assert len(arrays) != 0
     if len(arrays) == 1:
         return wrap_layout(
-            arrays[0], highlevel=highlevel, attrs=attrs, behavior=behavior
+            arrays[0],
+            highlevel=highlevel,
+            attrs=(attrs if attrs else {}) | arrays[0].attrs,
+            behavior=behavior,
         )
     else:
         # TODO: if each array is a record?
@@ -334,7 +337,19 @@ def _read_parquet_file(
             arrow_table = parquetfile.read_row_groups(row_groups, parquet_columns)
 
     arrow_table = ak._connect.pyarrow.convert_native_arrow_table_to_awkward(arrow_table)
-    return ak.operations.ak_from_arrow._impl(
+
+    array_attrs = {}
+    if arrow_table.schema.metadata:
+        if b"PANDAS_ATTRS" in arrow_table.schema.metadata:
+            array_metadata = arrow_table.schema.metadata[b"PANDAS_ATTRS"]
+            array_attrs = json.loads(array_metadata)
+        elif b"AWKWARD_ATTRS" in arrow_table.schema.metadata:
+            array_metadata = arrow_table.schema.metadata[b"AWKWARD_ATTRS"]
+            array_attrs = json.loads(array_metadata)
+    else:
+        array_attrs = {}
+
+    result = ak.operations.ak_from_arrow._impl(
         arrow_table,
         generate_bitmasks,
         # why is high-level False here?
@@ -343,6 +358,10 @@ def _read_parquet_file(
         None,
     )
 
+    result.attrs = array_attrs
+
+    return result
+
 
 class _DictOfEmptyBuffers:
     def __getitem__(self, where):

src/awkward/operations/ak_to_parquet.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 from collections.abc import Mapping, Sequence
 from os import fsdecode
 
@@ -405,6 +406,12 @@ def parquet_columns(specifier, only=None):
     if extensionarray:
         table = convert_awkward_arrow_table_to_native(table)
 
+    if hasattr(array, "attrs") and array.attrs:
+        df_metadata = {"AWKWARD_ATTRS": json.dumps(array.attrs.to_dict())}
+        existing_metadata = table.schema.metadata
+        merged_metadata = {**existing_metadata, **df_metadata}
+        table = table.replace_schema_metadata(merged_metadata)
+
     if parquet_extra_options is None:
         parquet_extra_options = {}
 

tests/test_1294_to_and_from_parquet.py

@@ -852,7 +852,7 @@ def test_unionarray(tmp_path, through, extensionarray):
 @pytest.fixture()
 def generate_datafiles(tmp_path):
     fs = fsspec.filesystem("file")
-    data1 = ak.from_iter([[1, 2, 3], [4, 5]])
+    data1 = ak.from_iter([[1, 2, 3], [4, 5]], attrs={"property": "value"})
     data2 = data1 + 1
     md1 = ak.to_parquet(data1, os.path.join(tmp_path, "data1.parq"))
     md2 = ak.to_parquet(data2, os.path.join(tmp_path, "data2.parq"))
@@ -920,3 +920,21 @@ def test_select(with_global_metadata):
 
     with pytest.raises(ValueError):
         ak.metadata_from_parquet(with_global_metadata, row_groups=[4])
+
+
+def test_awkward_attr_serialisation(generate_datafiles):
+    path, _mdlist, _fs = generate_datafiles
+    assert ak.from_parquet(f"{path}/data1.parq").attrs == {"property": "value"}
+
+
+def test_pandas_attr_serialisation(generate_datafiles):
+
+    pd = pytest.importorskip("pandas", "2.1.0")
+
+    path, _mdlist, _fs = generate_datafiles
+
+    df = pd.DataFrame({"x": [1, 2, 3], "y": [4, 5, 6]})
+    df.attrs = {"property": "value", 1: 2}
+    df.to_parquet(f"{path}/data3.parq")
+
+    assert ak.from_parquet(f"{path}/data3.parq").attrs == {"property": "value", "1": 2}