diff --git a/src/snowflake/snowpark/modin/plugin/extensions/base_overrides.py b/src/snowflake/snowpark/modin/plugin/extensions/base_overrides.py index 5aa937b809b..abbcb9bc762 100644 --- a/src/snowflake/snowpark/modin/plugin/extensions/base_overrides.py +++ b/src/snowflake/snowpark/modin/plugin/extensions/base_overrides.py @@ -2072,7 +2072,7 @@ def abs(self): # noqa: RT01, D200 Return a `BasePandasDataset` with absolute numeric value of each element. """ # TODO: SNOW-1119855: Modin upgrade - modin.pandas.base.BasePandasDataset - return self.__constructor__(query_compiler=self._query_compiler.unary_op("abs")) + return self.__constructor__(query_compiler=self._query_compiler.abs()) # Modin does dtype validation on unary ops that Snowpark pandas does not. @@ -2101,7 +2101,7 @@ def __neg__(self): BasePandasDataset """ # TODO: SNOW-1119855: Modin upgrade - modin.pandas.base.BasePandasDataset - return self.__constructor__(query_compiler=self._query_compiler.unary_op("__neg__")) + return self.__constructor__(query_compiler=self._query_compiler.negative()) # Modin needs to add a check for mapper is not None, which changes query counts in test_concat.py @@ -2199,7 +2199,7 @@ def __array_ufunc__(self, ufunc: np.ufunc, method: str, *inputs, **kwargs): # Use pandas version of ufunc if it exists if method != "__call__": # Return sentinel value NotImplemented - return NotImplemented + return NotImplemented # pragma: no cover from snowflake.snowpark.modin.plugin.utils.numpy_to_pandas import ( numpy_to_pandas_universal_func_map, ) @@ -2208,7 +2208,7 @@ def __array_ufunc__(self, ufunc: np.ufunc, method: str, *inputs, **kwargs): ufunc = numpy_to_pandas_universal_func_map[ufunc.__name__] return ufunc(self, inputs[1:], kwargs) # return the sentinel NotImplemented if we do not support this function - return NotImplemented + return NotImplemented # pragma: no cover # Snowpark pandas does extra argument validation.