PanicError in read_ndjson and polars.str.json_decode with empty struct

[deep8324]

PanicError in read_ndjson and polars.str.json_decode with empty struct

[deep8324]

The issue exists even with polars v 0.20.3

[jcmuel]

Hi @deep8324,

the data schema of the DataFrame in your example is not well-defined, so it's no surprise that you get an error. It seems like the actual schema that the example seems to represent is the following:

OrderedDict([('offer', Struct({'my_value': Int64, 'condition': List(Struct({'applicationId': Int64, 'conditionRequestReason': List(Struct({'a': Int64}))}))}))])

The data for the list of struct with the field a contains a list with an empty struct [{}]. It is not clear what this is supposed to represent. When you explicitly specify the schema as mentioned above, then this would represent a list with a single struct where the value for a is null. This is also the behavior that you can observe when you run the following example. Note that in the example below, the well_defined_data contains an empty list [] instead of [{}]. The incomplete_schema_data is exactly your example, but with additional schema information, such that you can read the data anyway.

# pylint: disable=missing-class-docstring, missing-function-docstring
""" Test for reading NDJSON files with nested fields. """
import io
import json
import unittest
from typing import Sequence, Any, cast

import polars as pl
import polars.testing


class TestNdJson(unittest.TestCase):
    def test_json_format(self):
        well_defined_data = io.StringIO(
            """{"offer": {"my_value": 0, "condition": [{"applicationId": 0, "conditionRequestReason": []}]}}
            {"offer": {"my_value": 0, "condition": [{"applicationId": 0, "conditionRequestReason": [{"a":0}]}]}}""")

        # Initialize the frame from a sequence of dictionaries.
        dicts = cast(Sequence[dict[str, Any]],
                     (json.loads(line) for line in well_defined_data))
        frame_from_dicts = pl.from_dicts(dicts)
        print(frame_from_dicts.schema)

        # Initialize the frame from the NDJSON file.
        frame_from_ndjson = pl.read_ndjson(well_defined_data)
        print(frame_from_ndjson.schema)

        # Compare the content of the frames.
        polars.testing.assert_frame_equal(frame_from_dicts, frame_from_ndjson)

        # Initialize the frame from a malformed NDJSON
        incomplete_schema_data = io.StringIO(
            """{"offer": {"my_value": 0, "condition": [{"applicationId": 0, "conditionRequestReason": [{}]}]}}""")
        frame_with_schema = pl.read_ndjson(incomplete_schema_data, schema=frame_from_ndjson.schema)
        frame_with_schema_export = frame_with_schema.write_ndjson()

        self.assertEqual(
            '{"offer":{"my_value":0,"condition":[{"applicationId":0,"conditionRequestReason":[{"a":null}]}]}}\n',
            frame_with_schema_export)


if __name__ == '__main__':
    unittest.main()

However, I can observe that without schema information, the behavior of from_dicts and read_ndjson differs when the data contains the awkward list with the empty struct [{}].

    def test_read_ndjson_list_of_awkward_struct(self):
        input_data = io.StringIO(
            """{"offer": {"my_value": 0, "condition": [{"applicationId": 0, "conditionRequestReason": [{}]}]}}
            {"offer": {"my_value": 0, "condition": [{"applicationId": 0, "conditionRequestReason": [{"a":0}]}]}}""")

        # Initialize the frame from a sequence of dictionaries.
        dicts = cast(Sequence[dict[str, Any]],
                     (json.loads(line) for line in input_data))
        frame_from_dicts = pl.from_dicts(dicts)
        print(frame_from_dicts.schema)

        # Initialize the frame from the NDJSON file.
        frame_from_ndjson = pl.read_ndjson(input_data)
        print(frame_from_ndjson.schema)

        # Compare the content of the frames.
        polars.testing.assert_frame_equal(frame_from_dicts, frame_from_ndjson)

The behavior of from_dicts seems to be a bit unexpected:

>>> frame_from_dicts
shape: (2, 1)
┌─────────────────────────┐
│ offer                   │
│ ---                     │
│ struct[2]               │
╞═════════════════════════╡
│ {0,[{0,[{null,null}]}]} │
│ {0,[{0,[{null,0}]}]}    │
└─────────────────────────┘
>>> frame_from_dicts.schema
OrderedDict([('offer', Struct({'my_value': Int64, 'condition': List(Struct({'applicationId': Int64, 'conditionRequestReason': List(Struct({'': Null, 'a': Int64}))}))}))])

[jcmuel]

Enclosed is a small simple example, where JSON decode doesn't determine the correct schema for a list of integer. It works for list of string, but not for list of integer.

# pylint: disable=missing-class-docstring, missing-function-docstring, too-few-public-methods
""" Unit tests for the polars JSON decode functionality. """
import io
import pytest
import polars as pl
import polars.testing


class TestPolarsJsonDecode:
    @pytest.mark.parametrize('test_name, input_ndjson', [
        ("str", """{"list_field": ["a", "b"]}
                   {"list_field": []}
                   {"list_field": ["c", "d", "e"]}"""),
        ("int", """{"list_field": [1, 2]}
                   {"list_field": []}
                   {"list_field": [4, 5, 6]}""")
    ])
    def test_json_decode_list_of_basic_type(self, test_name, input_ndjson):
        print(f"Reading list of {test_name}...")
        input_buf = io.StringIO(input_ndjson)
        frame_from_ndjson = pl.read_ndjson(input_buf)

        # pylint: disable-next=assignment-from-no-return
        series_with_json = pl.Series(values=input_buf).str.strip_chars_start()
        series_decoded_unnested = series_with_json.str.json_decode().struct.unnest()

        assert frame_from_ndjson.schema == series_decoded_unnested.schema
        polars.testing.assert_frame_equal(frame_from_ndjson, series_decoded_unnested)


if __name__ == '__main__':
    pytest.main()

[jcmuel]

I just did some tests with polars 0.20.6 (and pyarrow 13.0.0)

• The issue test_json_decode_list_of_basic_type seems to be resolved and doesn't reproduce anymore (since polars 0.20.4; it was still present in polars 0.20.3.)
• The issue test_read_ndjson_list_of_awkward_struct still reproduces.
• The panic in test_json_format still reproduces.
• The panic in test_json_decode_error still reproduces.

Is polars using arrow for decoding JSON input?

[jcmuel]

The following minimal test reproduces the PanicException that @deep8324 reported above.

import io
import json
import pytest
import polars as pl


@pytest.mark.parametrize('input_ndjson', [
    '{"bar": [{}]}',  # nested_null
    '{"foo":[{"bar":[{}]}]}'  # nested_nested_null
])
def test_ndjson_nested_nested_null(input_ndjson):
    """ Test whether read_ndjson and from_dicts imports nested null structs in the same way. """
    buffer = io.StringIO(input_ndjson)
    json_obj = json.loads(next(buffer))
    df_from_dicts = pl.from_dicts([json_obj])
    df_from_ndjson = pl.read_ndjson(buffer)
    assert df_from_dicts.schema == df_from_ndjson.schema

First we have the nested_null subtest, where the assertion for the expected schema fails:

Expected :OrderedDict([('bar', List(Struct({})))])
Actual   :OrderedDict([('bar', List(Struct({'': Null})))])

And then we have the nested_nested_null subtest, that causes the PanicException, that reports exactly the same unexpected type mismatch:

ListArray's child's DataType must match. However, the expected DataType is 
Struct([Field { name: \"bar\", data_type: LargeList(Field { name: \"item\", data_type: 
>>>>>>>>Struct([]),
is_nullable: true, metadata: {} }), is_nullable: true, metadata: {} }]) 
while it got 
Struct([Field { name: \"bar\", data_type: LargeList(Field { name: \"item\", data_type: 
>>>>>>>>Struct([Field { name: \"\", data_type: Null, is_nullable: true, metadata: {} }]), 
is_nullable: true, metadata: {} }), is_nullable: true, metadata: {} }])."))

The unexpected difference is in the line that I highlighted with the arrows. Instead of an empty struct, the code receives a struct with a field that has the name "" and a null value.

polars unit tested behavior of the import of empty structs from dictionaries and the import of empty structs from NDJSON is inconsistent. Importing {} as Struct([]), like it is done for NDJSON, seems to be the right approach. Importing {} as Struct({'': Null}) seems to be a bug in from_dicts, even if this behavior is unit tested. (I had a look at the polars unit tests.)

Feel free to add this test to the polars unit tests.

[jcmuel]

It still reproduces in polars 0.20.10. I think it really comes down to the fact that polars is treating the empty object "{}" in a different way, depending on where it shows up in the schema.

def test_ndjson_empty_object() -> None:
    """
    The actual type of `empty_object_column` (tested in `test_ndjson_null_buffer()`) is inconsistent with
    the actual type of `nested_empty_object_column` (tested in `test_ndjson_nested_null()`).
    """
    data = io.BytesIO(
        b"""\
    {"id": 1, "empty_object_column": {}, "nested_empty_object_column": [{}]}
    {"id": 2, "empty_object_column": {}, "nested_empty_object_column": [{}]}
    {"id": 3, "empty_object_column": {}, "nested_empty_object_column": [{}]}
    """
    )

    assert pl.read_ndjson(data).schema == {
        "id": pl.Int64,
        "empty_object_column": pl.Struct([]),  # or pl.Struct([pl.Field("", pl.Null)]) ?
        "nested_empty_object_column": pl.List(pl.Struct([]))
    }

The test above fails with the output:

E         -     'empty_object_column': Struct({}),
E         +     'empty_object_column': Struct({'': Null}),
E         ?                                    ++++++++

@ritchie46 : It seems like you wrote the test_ndjson_null_buffer test.
@alexander-beedie : It seems like you wrote the test_ndjson_nested_null test.

I'm mentioning the two of you, since I would be interested to know which of the two behaviors is the expected one.