Correct type inference for UInt64Index during access

doc/source/whatsnew/v1.0.0.rst

@@ -345,6 +345,7 @@ Numeric
 - Improved error message when using `frac` > 1 and `replace` = False (:issue:`27451`)
 - Bug in numeric indexes resulted in it being possible to instantiate an :class:`Int64Index`, :class:`UInt64Index`, or :class:`Float64Index` with an invalid dtype (e.g. datetime-like) (:issue:`29539`)
 - Bug in :class:`UInt64Index` precision loss while constructing from a list with values in the ``np.uint64`` range (:issue:`29526`)
+- Bug in :meth:`NumericIndex._convert_arr_indexer` that caused indexing to fail when integers in the ``np.uint64`` range were used (:issue:`28023`)
 -
 
 Conversion

pandas/core/indexes/numeric.py

@@ -2,7 +2,7 @@
 
 import numpy as np
 
-from pandas._libs import index as libindex
+from pandas._libs import index as libindex, lib
 from pandas.util._decorators import Appender, cache_readonly
 
 from pandas.core.dtypes.cast import astype_nansafe
@@ -331,13 +331,15 @@ def _convert_scalar_indexer(self, key, kind=None):
 
     @Appender(_index_shared_docs["_convert_arr_indexer"])
     def _convert_arr_indexer(self, keyarr):
-        # Cast the indexer to uint64 if possible so
-        # that the values returned from indexing are
-        # also uint64.
-        keyarr = com.asarray_tuplesafe(keyarr)
-        if is_integer_dtype(keyarr):
-            return com.asarray_tuplesafe(keyarr, dtype=np.uint64)
-        return keyarr
+        # Cast the indexer to uint64 if possible so that the values returned
+        # from indexing are also uint64.
+        dtype = None
+        if is_integer_dtype(keyarr) or (
+            lib.infer_dtype(keyarr, skipna=False) == "integer"
+        ):
+            dtype = np.uint64
+
+        return com.asarray_tuplesafe(keyarr, dtype=dtype)
 
     @Appender(_index_shared_docs["_convert_index_indexer"])
     def _convert_index_indexer(self, keyarr):

pandas/tests/indexes/test_numeric.py

@@ -1209,3 +1209,22 @@ def test_range_float_union_dtype():
 
     result = other.union(index)
     tm.assert_index_equal(result, expected)
+
+
+def test_uint_index_does_not_convert_to_float64():
+    # https://github.com/pandas-dev/pandas/issues/28279
+    series = pd.Series(
+        [0, 1, 2, 3, 4],
+        index=[
+            7606741985629028552,
+            17876870360202815256,
+            13106359306506049338,
+            8991270399732411471,
+            8991270399732411472,
+        ],
+    )
+
+    result = series.loc[[7606741985629028552, 17876870360202815256]].index
+    expected = UInt64Index([7606741985629028552, 17876870360202815256], dtype="uint64")
+
+    tm.assert_index_equal(result, expected)