BUG: pd.read_json May Not Maintain Numeric String Index (pandas-dev#38727)

Author: theoniko

doc/source/whatsnew/v1.3.0.rst

@@ -276,6 +276,7 @@ I/O
 - Bug in :func:`json_normalize` resulting in the first element of a generator object not being included in the returned ``DataFrame`` (:issue:`35923`)
 - Bug in :func:`read_excel` forward filling :class:`MultiIndex` names with multiple header and index columns specified (:issue:`34673`)
 - :func:`pandas.read_excel` now respects :func:``pandas.set_option`` (:issue:`34252`)
+- Bug in :func:``read_json`` when ``orient="split"`` does not maintan numeric string index (:issue:`28556`)
 
 Period
 ^^^^^^

pandas/io/json/_json.py

@@ -894,14 +894,11 @@ def _try_convert_data(self, name, data, use_dtypes=True, convert_dates=True):
             if result:
                 return new_data, True
 
-        result = False
-
         if data.dtype == "object":
 
             # try float
             try:
                 data = data.astype("float64")
-                result = True
             except (TypeError, ValueError):
                 pass
 
@@ -912,7 +909,6 @@ def _try_convert_data(self, name, data, use_dtypes=True, convert_dates=True):
                 # coerce floats to 64
                 try:
                     data = data.astype("float64")
-                    result = True
                 except (TypeError, ValueError):
                     pass
 
@@ -924,7 +920,6 @@ def _try_convert_data(self, name, data, use_dtypes=True, convert_dates=True):
                 new_data = data.astype("int64")
                 if (new_data == data).all():
                     data = new_data
-                    result = True
             except (TypeError, ValueError, OverflowError):
                 pass
 
@@ -934,11 +929,15 @@ def _try_convert_data(self, name, data, use_dtypes=True, convert_dates=True):
             # coerce floats to 64
             try:
                 data = data.astype("int64")
-                result = True
             except (TypeError, ValueError):
                 pass
 
-        return data, result
+        # if we have an index, we want to preserve dtypes
+        if name == "index" and len(data):
+            if self.orient == "split":
+                return data, False
+
+        return data, True
 
     def _try_convert_to_date(self, data):
         """

pandas/tests/io/json/test_pandas.py

@@ -193,12 +193,13 @@ def test_roundtrip_str_axes(self, request, orient, convert_axes, numpy, dtype):
         # JSON objects. JSON keys are by definition strings, so there's no way
         # to disambiguate whether those keys actually were strings or numeric
         # beforehand and numeric wins out.
-        # TODO: Split should be able to support this
-        if convert_axes and (orient in ("split", "index", "columns")):
+        if convert_axes and (orient in ("index", "columns")):
             expected.columns = expected.columns.astype(np.int64)
             expected.index = expected.index.astype(np.int64)
         elif orient == "records" and convert_axes:
             expected.columns = expected.columns.astype(np.int64)
+        elif convert_axes and orient == "split":
+            expected.columns = expected.columns.astype(np.int64)
 
         assert_json_roundtrip_equal(result, expected, orient)