Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
7.0.0
-
None
-
pyarrow 7.0.0 installed via pip.
Description
I've been happily making ExtensionArrays, but recently noticed that they aren't preserved by round-trips through Parquet files when use_compliant_nested_type=True.
Consider this writer.py:
import json import numpy as np import pyarrow as pa import pyarrow.parquet as pq class AnnotatedType(pa.ExtensionType): def __init__(self, storage_type, annotation): self.annotation = annotation super().__init__(storage_type, "my:app") def __arrow_ext_serialize__(self): return json.dumps(self.annotation).encode() @classmethod def __arrow_ext_deserialize__(cls, storage_type, serialized): annotation = json.loads(serialized.decode()) return cls(storage_type, annotation) @property def num_buffers(self): return self.storage_type.num_buffers @property def num_fields(self): return self.storage_type.num_fields pa.register_extension_type(AnnotatedType(pa.null(), None)) array = pa.Array.from_buffers( AnnotatedType(pa.list_(pa.float64()), {"cool": "beans"}), 3, [None, pa.py_buffer(np.array([0, 3, 3, 5], np.int32))], children=[pa.array([1.1, 2.2, 3.3, 4.4, 5.5])], ) table = pa.table({"": array}) print(table) pq.write_table(table, "tmp.parquet", use_compliant_nested_type=True)
And this reader.py:
import json import numpy as np import pyarrow as pa import pyarrow.parquet as pq class AnnotatedType(pa.ExtensionType): def __init__(self, storage_type, annotation): self.annotation = annotation super().__init__(storage_type, "my:app") def __arrow_ext_serialize__(self): return json.dumps(self.annotation).encode() @classmethod def __arrow_ext_deserialize__(cls, storage_type, serialized): annotation = json.loads(serialized.decode()) return cls(storage_type, annotation) @property def num_buffers(self): return self.storage_type.num_buffers @property def num_fields(self): return self.storage_type.num_fields pa.register_extension_type(AnnotatedType(pa.null(), None)) table = pq.read_table("tmp.parquet") print(table)
(The AnnotatedType is the same; I wrote it twice for explicitness.)
When the writer.py has use_compliant_nested_type=False, the output is
% python writer.py pyarrow.Table : extension<my:app<AnnotatedType>> ---- : [[[1.1,2.2,3.3],[],[4.4,5.5]]] % python reader.py pyarrow.Table : extension<my:app<AnnotatedType>> ---- : [[[1.1,2.2,3.3],[],[4.4,5.5]]]
In other words, the AnnotatedType is preserved. When use_compliant_nested_type=True, however,
% rm tmp.parquet rm: remove regular file 'tmp.parquet'? y % python writer.py pyarrow.Table : extension<my:app<AnnotatedType>> ---- : [[[1.1,2.2,3.3],[],[4.4,5.5]]] % python reader.py pyarrow.Table : list<element: double> child 0, element: double ---- : [[[1.1,2.2,3.3],[],[4.4,5.5]]]
The issue doesn't seem to be in the writing, but in the reading: regardless of whether use_compliant_nested_type is True or False, I can see the extension metadata in the Parquet → Arrow converted schema.
>>> import pyarrow.parquet as pq >>> pq.ParquetFile("tmp.parquet").schema.to_arrow_schema() : list<item: double> child 0, item: double -- field metadata -- ARROW:extension:metadata: '{"cool": "beans"}' ARROW:extension:name: 'my:app'
versus
>>> import pyarrow.parquet as pq >>> pq.ParquetFile("tmp.parquet").schema.to_arrow_schema() : list<element: double> child 0, element: double -- field metadata -- ARROW:extension:metadata: '{"cool": "beans"}' ARROW:extension:name: 'my:app'
Note that the first has "item: double" and the second has "element: double".
(I'm also rather surprised that use_compliant_nested_type=False is an option. Wouldn't you want the Parquet files to always be written with compliant lists? I noticed this when I was having trouble getting the data into BigQuery.)