BUG: Integer column index breaks json roundtrip with orient=table

[rikardn]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

col1 = [1.0, 2.0, 3.5, 6.75]
col2 = [2.1, 3.1, 4.1, 5.1]
df = pd.DataFrame({1: col1, 2:col2}, index=[110, 112, 113, 121])
df.index.name = 'ID'
s = df.to_json(orient='table')
new = pd.read_json(s, orient='table')

Issue Description

The new dataframe will become

      1   2
ID         
110 NaN NaN
112 NaN NaN
113 NaN NaN
121 NaN NaN

Expected Behavior

The expected dataframe would look like this:

        1    2
ID            
110  1.00  2.1
112  2.00  3.1
113  3.50  4.1
121  6.75  5.1

Changing to strings instead of integers in the column index will give the expected result:

col1 = [1.0, 2.0, 3.5, 6.75]
col2 = [2.1, 3.1, 4.1, 5.1]
df = pd.DataFrame({'1': col1, '2':col2}, index=[110, 112, 113, 121])
df.index.name = 'ID'
s = df.to_json(orient='table')
new = pd.read_json(s, orient='table')

Installed Versions

This crashed in my environment with the error assert '_distutils' in core.__file__, core.__file__ raised from lib/python3.9/site-packages/_distutils_hack/__init__.py", line 59, in ensure_local_distutils

[SeanRBurton]

I'm willing to tackle this, but I'm not sure of the best approach:

1. Disallow non-string keys unless orient="values".
2. If orient="table", reconstruct the keys with the correct type using the schema. Raise an exception if the table has two keys which serialize to the same value, e.g. 1 and "1".
3. Something else?

[rikardn]

I can add my user perspective:

The use case I have is to serialize a DataFrame to json and be able to get back the exact same DataFrame regardless of type of Index etc. Given this use case it is really not interesting to select orient.

[gpapadok]

Playing around I found another issue.

col1 = [1.0, 2.0, 3.5, 6.75]
col2 = [2.1, 3.1, 4.1, 5.1]
col3 = [2.2, 3.3, 4.4, 5.5]
df = pd.DataFrame({'1': col1, '2':col2, 1:col3}, index=[110, 112, 113, 121])
df.index.name = 'ID'
s = df.to_json()
new = pd.read_json(s)

The df data frame has two columns with same name but different type:

        1    2    1
ID                 
110  1.00  2.1  2.2
112  2.00  3.1  3.3
113  3.50  4.1  4.4
121  6.75  5.1  5.5

By serializing and deserializing we get back new as:

       1    2
110  2.2  2.1
112  3.3  3.1
113  4.4  4.1
121  5.5  5.1

Column types also get converted from strings to ints:

In [118]: df.columns, new.columns
Out[118]: (Index(['1', '2', 1], dtype='object'), Int64Index([1, 2], dtype='int64'))

[jmg-duarte]

The issue is in:

pandas/pandas/core/internals/construction.py

Line 955 in 1b2646a

content = lib.dicts_to_array(data, list(columns))

This is the result:

Called by:

pandas/pandas/io/json/_table_schema.py

Line 353 in 1b2646a

df = DataFrame(table["data"], columns=col_order)[col_order]

I'm interested in helping more, do you where can I look further @mroeschke?

[mroeschke]

Not too familiar with this code path, so I'm not exactly sure off the top of my head how nans are getting introduced, but further investigation welcome!

[jmg-duarte]

I found the issue and have a fix, but the code I am touching is ~10 years old according to Gitlens git blame.

Either way, the problem is that when reading the JSON, pandas will have an array of dicts like so:

{"ID":110,"1":1.0,"2":2.1}

Since JSON only supports strings as keys, we have an issue here, because we don't have a 1:1 mapping (https://stackoverflow.com/questions/1450957/pythons-json-module-converts-int-dictionary-keys-to-strings).

In the following code, the row will be a dictionary with strings as keys, and columns (and k) will be the proper column array.

pandas/pandas/_libs/lib.pyx

Lines 432 to 455 in 1b2646a

	@cython.wraparound(False)
	@cython.boundscheck(False)
	def dicts_to_array(dicts: list, columns: list):
	cdef:
	Py_ssize_t i, j, k, n
	ndarray[object, ndim=2] result
	dict row
	object col, onan = np.nan
	k = len(columns)
	n = len(dicts)
	result = np.empty((n, k), dtype='O')
	for i in range(n):
	row = dicts[i]
	for j in range(k):
	col = columns[j]
	if col in row:
	result[i, j] = row[col]
	else:
	result[i, j] = onan
	return result

Meaning there is a mismatch as 1 is not in {"ID":110,"1":1.0,"2":2.1}, though "1" is.

I've tested the following fix:

- col = columns[j]
+ col = str(columns[j])

And it solves this issue, I'm currently running the test suite to check if nothing else got broken.

[jmg-duarte]

I've finished running the test suite, before and after, this change seems to 4 introduce failures. Is there a way I can check the summary of both runs and diff between them, even if manually?

[coatless]

@jmg-duarte did you end up solving the issue?

If not, @mroeschke could you suggest a way for @jmg-duarte to diff between summary runs? It's not ideal that the JSON being generated is invalid.