Support Infinity, -Infinity and NaN in read_json

[maxnoe]

While these special values are not strictly standard conform, most implementations do allow or use them.

For example Google's GSON Library or python's json module in the standard library.

[jreback]

nan's are already supported.

In [1]: df = DataFrame({'A' : [np.nan,1,np.inf,-np.inf]})

In [2]: df.to_json(None)
Out[2]: '{"A":{"0":null,"1":1.0,"2":null,"3":null}}'

[maxnoe]

But not NaN for reading: pd.read_json('{"a": [NaN, Infinity, -Infinity]}')

[jreback]

that's not valid json

[maxnoe]

I know, that's what I wrote in the first post. But it is a commonly used extension.

See the description in Google's GSON package:

Section 2.4 of JSON specification disallows special double values (NaN, Infinity, -Infinity). However, Javascript specification (see section 4.3.20, 4.3.22, 4.3.23) allows these values as valid Javascript values. Moreover, most JavaScript engines will accept these special values in JSON without problem. So, at a practical level, it makes sense to accept these values as valid JSON even though JSON specification disallows them.

Or python's json module:

It also understands NaN, Infinity, and -Infinity as their corresponding float values, which is outside the JSON spec.

[maxnoe]

Also interesting: the json specifications explicitly allow to accept extensions of the standard:
https://tools.ietf.org/html/rfc7159.html#section-9

A JSON parser transforms a JSON text into another representation. A
JSON parser MUST accept all texts that conform to the JSON grammar.
A JSON parser MAY accept non-JSON forms or extensions.

[kbruegge]

👍 This would indeed be very helpful for interoperability between a pandas backend and a JS frontend.

[zack-sampson]

I took a crack at this and discovered that the json library underlying read_json, ujson, doesn't handle NaN nor infinity. In previous discussions (more) the authors of ujson resisted adding support because it's out of spec, and there's no way to specify a custom decodor nor encoder with ujson.

[maciejkula]

The newest master throws a ValueError: Expected object or value exception when trying to parse JSON containing inf of nan values. While I understand that these are not valid JSON, it might be helpful to throw more informative exceptions to help the user debug the issue.

Script to reproduce (using Hypothesis):

import json

from hypothesis import assume, example, given
import hypothesis.strategies as st

import pandas as pd

FAILING_EXAMPLES = (
    [{u'': float('inf')}],
    [{u'': float('nan')}],
)

@given(st.lists(st.dictionaries(st.text(),
                                st.floats() | st.booleans() | st.text() | st.none())))
@example([{u'': float('inf')}])
@example([{u'': float('nan')}])
def test_load_json(test_input):

    df = pd.read_json(json.dumps(test_input))

[yifeim]

Hi there,

Thanks for the thorough discussions. We would suppose that pandas to be friendly to data scientists and a little surprised to find compatibility issues with IEEE float standards. Infinity/-Infinity is particularly useful because it is flexible enough to be clipped to the max/min value in any valid ranges. This is not possible with NaN/null.

As of today, it is possible to encode Infinity using default_handler, but not to decode it, correct?

Thanks.