Why is pandas.read_json, modifying the value of long integers?

• Using python 3.8.5 and pandas 1.1.1

Current Implementation

• First, the code reads the file in and converts it from a str type to a dict, with json.loads

with open (file, "r")  as f:
    all_data = json.loads(f.read()) 

• Then 'value' is converted back to a str

json.dumps(all_data['objects']['value'])

• Using orient='index' sets the keys as columns headers and the values are in the rows.
  • The data is also converted to an int at this point, and the value changes.
  • I'm guessing that there's some floating point conversion issue in this step
    • pandas issues: read_json reads large integers as strings incorrectly if dtype not explicitly mentioned #20608

pd.read_json(json.dumps(all_data['objects']['value']), orient='index')

Updated code

Option 1

• Use pandas.DataFrame.from_dict and then convert to numeric.

file = "test_data.json"
with open (file, "r")  as f:
    all_data = json.loads(f.read()) 

# use .from_dict
data = pd.DataFrame.from_dict(all_data['objects']['value'], orient='index')

# convert columns to numeric
data[['id_1', 'id_2']] = data[['id_1', 'id_2']].apply(pd.to_numeric, errors='coerce')

data = data.reset_index(drop=True)

# display(data)
                        timestamp                 id_1                 id_2
0  Wed Aug 26 08:52:57 +0000 2020  1298543947669573634  1298519559306190850

print(data.info())
[out]:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1 entries, 0 to 0
Data columns (total 3 columns):
 #   Column     Non-Null Count  Dtype 
---  ------     --------------  ----- 
 0   timestamp  1 non-null      object
 1   id_1       1 non-null      int64 
 2   id_2       1 non-null      int64 
dtypes: int64(2), object(1)
memory usage: 152.0+ bytes

Option 2

• Use pandas.json_normalize and then convert columns to numeric.

file = "test_data.json"
with open (file, "r")  as f:
    all_data = json.loads(f.read()) 

# read all_data into a dataframe
df = pd.json_normalize(all_data['objects']['value'])

# rename the columns
df.columns = [x.split('.')[1] for x in df.columns]

# convert to numeric
df[['id_1', 'id_2']] = df[['id_1', 'id_2']].apply(pd.to_numeric, errors='coerce')

# display(df)
                        timestamp                 id_1                 id_2
0  Wed Aug 26 08:52:57 +0000 2020  1298543947669573634  1298519559306190850

print(df.info()
[out]:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1 entries, 0 to 0
Data columns (total 3 columns):
 #   Column     Non-Null Count  Dtype 
---  ------     --------------  ----- 
 0   timestamp  1 non-null      object
 1   id_1       1 non-null      int64 
 2   id_2       1 non-null      int64 
dtypes: int64(2), object(1)
memory usage: 152.0+ bytes

This is caused by issue 20608 and still happens in the current 1.2.4 version of Pandas.

Here's my workaround, which is even slightly faster on my data than read_json:

def broken_load_json(path):
    """There's an open issue: https://github.com/pandas-dev/pandas/issues/20608
    about read_csv loading large integers incorrectly because it's converting
    from string to float to int, losing precision."""
    df = pd.read_json(pathlib.Path(path), orient='index')
    return df

def orjson_load_json(path):
    import orjson  # The builting json module would also work
    with open(path) as f:
        d = orjson.loads(f.read())
    df = pd.DataFrame.from_dict(d, orient='index')  # Builds the index from the dict's keys as strings, sadly
    # Fix the dtype of the index
    df = df.reset_index()
    df['index'] = df['index'].astype('int64')
    df = df.set_index('index')
    return df

Note that my answer preserves the values of the IDs, which are meaningful in my use case.