NumPy Interoperability

Snowpark pandas provides limited interoperability with NumPy functions through the NumPy NEP18 and NEP13 specifications defined by __array_ufunc__ and __array_function__. A discrete number of NumPy APIs are translated to distributed snowpark pandas functions.

NumPy method

Notes for current implementation

np.where

Mapped to np.where(cond, x, y) to x.where(cond, y) cond, x, and y should have the same shapes or be scalars. The result is always a Snowpark pandas DataFrame.

Since this function maps to df.where the column and index labels are considered, as opposed strict positional indexing in NumPy.

cond, x, and y can either be all non-scalars or a mix of scalars and non-scalars, such that non-scalars have the same shape. (If cond, x, and y are all scalars, NumPy will not call the dispatcher at all, and the normal NumPy behavior will occur.)

np.add

Mapped to df.__add__(df2)

np.logical_and

Mapped to df.__and__(df2)

np.logical_or

Mapped to df.__or__(df2)

np.logical_xor

Mapped to df.__xor__(df2)

np.logical_not

Mapped to ~df.astype(bool)

NEP18 Implementation Details

NumPy differs from pandas and Snowflake pandas in several key respects. It is important to understand that the interoperability provided is to support common pandas use-cases, rather than matrix or linear algebra operations. NumPy functions are mapped, with some transformation, to their pandas analogues.

Return Value

NEP18 does not specify the return value when implementing a function like np.where, but they suggest that the return value should match the input types. We follow that suggestion here and return a Snowpark pandas DataFrame.

Broadcasting

NumPy will “broadcast” all arguments into the same array shape so operations can be vectorized on the CPU. Snowpark pandas should not do this because all execution runs within Snowflake. All input DataFrames or Series should be of the same shape and will not be broadcast. Scalar values can also be used as an input.

Positional Operations

NumPy always performs positional operations on input datatypes, assuming they are similarly shaped and meaningful arrays. Pandas can have DataFrames which represent the same data but with different column ordering. Even when a numpy method is called on a Snow pandas DataFrame we continue to consider the labels while performing the operation.