Using pyarrow from C++ and Cython Code

pyarrow features both a Cython and C++ API.

C++ API

The Arrow C++ header files are bundled with a pyarrow installation. To get the absolute path to this directory (like numpy.get_include()), use:

import pyarrow as pa
pa.get_include()

Assuming the path above is on your compiler’s include path, the pyarrow API can be included using the following directive:

#include <arrow/python/pyarrow.h>

This will not include other parts of the Arrow API, which you will need to include yourself (for example arrow/api.h).

When building C extensions that use the Arrow C++ libraries, you must add appropriate linker flags. We have provided functions pyarrow.get_libraries and pyarrow.get_library_dirs which return a list of library names and likely library install locations (if you installed pyarrow with pip or conda). These must be included when declaring your C extensions with distutils (see below).

Initializing the API

int import_pyarrow()

Initialize inner pointers of the pyarrow API. On success, 0 is returned. Otherwise, -1 is returned and a Python exception is set.

It is mandatory to call this function before calling any other function in the pyarrow C++ API. Failing to do so will likely lead to crashes.

Wrapping and Unwrapping

pyarrow provides the following functions to go back and forth between Python wrappers (as exposed by the pyarrow Python API) and the underlying C++ objects.

bool is_array(PyObject *obj)

Return whether obj wraps an Arrow C++ Array pointer; in other words, whether obj is a pyarrow.Array instance.

bool is_buffer(PyObject *obj)

Return whether obj wraps an Arrow C++ Buffer pointer; in other words, whether obj is a pyarrow.Buffer instance.

bool is_column(PyObject *obj)

Return whether obj wraps an Arrow C++ Column pointer; in other words, whether obj is a pyarrow.Column instance.

bool is_data_type(PyObject *obj)

Return whether obj wraps an Arrow C++ DataType pointer; in other words, whether obj is a pyarrow.DataType instance.

bool is_field(PyObject *obj)

Return whether obj wraps an Arrow C++ Field pointer; in other words, whether obj is a pyarrow.Field instance.

bool is_record_batch(PyObject *obj)

Return whether obj wraps an Arrow C++ RecordBatch pointer; in other words, whether obj is a pyarrow.RecordBatch instance.

bool is_schema(PyObject *obj)

Return whether obj wraps an Arrow C++ Schema pointer; in other words, whether obj is a pyarrow.Schema instance.

bool is_table(PyObject *obj)

Return whether obj wraps an Arrow C++ Table pointer; in other words, whether obj is a pyarrow.Table instance.

bool is_tensor(PyObject *obj)

Return whether obj wraps an Arrow C++ Tensor pointer; in other words, whether obj is a pyarrow.Tensor instance.

The following functions expect a pyarrow object, unwrap the underlying Arrow C++ API pointer, and put it in the out parameter. The returned Status object must be inspected first to know whether any error occurred. If successful, out is guaranteed to be non-NULL.

Status unwrap_array(PyObject *obj, std::shared_ptr<Array> *out)

Unwrap the Arrow C++ Array pointer from obj and put it in out.

Status unwrap_buffer(PyObject *obj, std::shared_ptr<Buffer> *out)

Unwrap the Arrow C++ Buffer pointer from obj and put it in out.

Status unwrap_column(PyObject *obj, std::shared_ptr<Column> *out)

Unwrap the Arrow C++ Column pointer from obj and put it in out.

Status unwrap_data_type(PyObject *obj, std::shared_ptr<DataType> *out)

Unwrap the Arrow C++ DataType pointer from obj and put it in out.

Status unwrap_field(PyObject *obj, std::shared_ptr<Field> *out)

Unwrap the Arrow C++ Field pointer from obj and put it in out.

Status unwrap_record_batch(PyObject *obj, std::shared_ptr<RecordBatch> *out)

Unwrap the Arrow C++ RecordBatch pointer from obj and put it in out.

Status unwrap_schema(PyObject *obj, std::shared_ptr<Schema> *out)

Unwrap the Arrow C++ Schema pointer from obj and put it in out.

Status unwrap_table(PyObject *obj, std::shared_ptr<Table> *out)

Unwrap the Arrow C++ Table pointer from obj and put it in out.

Status unwrap_tensor(PyObject *obj, std::shared_ptr<Tensor> *out)

Unwrap the Arrow C++ Tensor pointer from obj and put it in out.

The following functions take an Arrow C++ API pointer and wrap it in a pyarray object of the corresponding type. A new reference is returned. On error, NULL is returned and a Python exception is set.

PyObject *wrap_array(const std::shared_ptr<Array> &array)

Wrap the Arrow C++ array in a pyarrow.Array instance.

PyObject *wrap_buffer(const std::shared_ptr<Buffer> &buffer)

Wrap the Arrow C++ buffer in a pyarrow.Buffer instance.

PyObject *wrap_column(const std::shared_ptr<Column> &column)

Wrap the Arrow C++ column in a pyarrow.Column instance.

PyObject *wrap_data_type(const std::shared_ptr<DataType> &data_type)

Wrap the Arrow C++ data_type in a pyarrow.DataType instance.

PyObject *wrap_field(const std::shared_ptr<Field> &field)

Wrap the Arrow C++ field in a pyarrow.Field instance.

PyObject *wrap_record_batch(const std::shared_ptr<RecordBatch> &batch)

Wrap the Arrow C++ record batch in a pyarrow.RecordBatch instance.

PyObject *wrap_schema(const std::shared_ptr<Schema> &schema)

Wrap the Arrow C++ schema in a pyarrow.Schema instance.

PyObject *wrap_table(const std::shared_ptr<Table> &table)

Wrap the Arrow C++ table in a pyarrow.Table instance.

PyObject *wrap_tensor(const std::shared_ptr<Tensor> &tensor)

Wrap the Arrow C++ tensor in a pyarrow.Tensor instance.

Cython API

The Cython API more or less mirrors the C++ API, but the calling convention can be different as required by Cython. In Cython, you don’t need to initialize the API as that will be handled automaticalled by the cimport directive.

Note

Classes from the Arrow C++ API are renamed when exposed in Cython, to avoid named clashes with the corresponding Python classes. For example, C++ Arrow arrays have the CArray type and Array is the corresponding Python wrapper class.

Wrapping and Unwrapping

The following functions expect a pyarrow object, unwrap the underlying Arrow C++ API pointer, and return it. NULL is returned (without setting an exception) if the input is not of the right type.

pyarrow.pyarrow_unwrap_array(obj) → shared_ptr[CArray]

Unwrap the Arrow C++ Array pointer from obj.

pyarrow.pyarrow_unwrap_batch(obj) → shared_ptr[CRecordBatch]

Unwrap the Arrow C++ RecordBatch pointer from obj.

pyarrow.pyarrow_unwrap_buffer(obj) → shared_ptr[CBuffer]

Unwrap the Arrow C++ Buffer pointer from obj.

pyarrow.pyarrow_unwrap_column(obj) → shared_ptr[CColumn]

Unwrap the Arrow C++ Column pointer from obj.

pyarrow.pyarrow_unwrap_data_type(obj) → shared_ptr[CDataType]

Unwrap the Arrow C++ CDataType pointer from obj.

pyarrow.pyarrow_unwrap_field(obj) → shared_ptr[CField]

Unwrap the Arrow C++ Field pointer from obj.

pyarrow.pyarrow_unwrap_schema(obj) → shared_ptr[CSchema]

Unwrap the Arrow C++ Schema pointer from obj.

pyarrow.pyarrow_unwrap_table(obj) → shared_ptr[CTable]

Unwrap the Arrow C++ Table pointer from obj.

pyarrow.pyarrow_unwrap_tensor(obj) → shared_ptr[CTensor]

Unwrap the Arrow C++ Tensor pointer from obj.

The following functions take a Arrow C++ API pointer and wrap it in a pyarray object of the corresponding type. An exception is raised on error.

pyarrow.pyarrow_wrap_array(sp_array: const shared_ptr[CArray]& array) → object

Wrap the Arrow C++ array in a Python pyarrow.Array instance.

pyarrow.pyarrow_wrap_batch(sp_array: const shared_ptr[CRecordBatch]& batch) → object

Wrap the Arrow C++ record batch in a Python pyarrow.RecordBatch instance.

pyarrow.pyarrow_wrap_buffer(sp_array: const shared_ptr[CBuffer]& buffer) → object

Wrap the Arrow C++ buffer in a Python pyarrow.Buffer instance.

pyarrow.pyarrow_wrap_column(sp_array: const shared_ptr[CColumn]& column) → object

Wrap the Arrow C++ column in a Python pyarrow.Column instance.

pyarrow.pyarrow_wrap_data_type(sp_array: const shared_ptr[CDataType]& data_type) → object

Wrap the Arrow C++ data_type in a Python pyarrow.DataType instance.

pyarrow.pyarrow_wrap_field(sp_array: const shared_ptr[CField]& field) → object

Wrap the Arrow C++ field in a Python pyarrow.Field instance.

pyarrow.pyarrow_wrap_resizable_buffer(sp_array: const shared_ptr[CResizableBuffer]& buffer) → object

Wrap the Arrow C++ resizable buffer in a Python pyarrow.ResizableBuffer instance.

pyarrow.pyarrow_wrap_schema(sp_array: const shared_ptr[CSchema]& schema) → object

Wrap the Arrow C++ schema in a Python pyarrow.Schema instance.

pyarrow.pyarrow_wrap_table(sp_array: const shared_ptr[CTable]& table) → object

Wrap the Arrow C++ table in a Python pyarrow.Table instance.

pyarrow.pyarrow_wrap_tensor(sp_array: const shared_ptr[CTensor]& tensor) → object

Wrap the Arrow C++ tensor in a Python pyarrow.Tensor instance.

Example

The following Cython module shows how to unwrap a Python object and call the underlying C++ object’s API.

# distutils: language=c++

from pyarrow.lib cimport *


def get_array_length(obj):
    # Just an example function accessing both the pyarrow Cython API
    # and the Arrow C++ API
    cdef shared_ptr[CArray] arr = pyarrow_unwrap_array(obj)
    if arr.get() == NULL:
        raise TypeError("not an array")
    return arr.get().length()

To build this module, you will need a slightly customized setup.py file (this is assuming the file above is named example.pyx):

from distutils.core import setup
from Cython.Build import cythonize

import os
import numpy as np
import pyarrow as pa


ext_modules = cythonize("example.pyx")

for ext in ext_modules:
    # The Numpy C headers are currently required
    ext.include_dirs.append(np.get_include())
    ext.include_dirs.append(pa.get_include())
    ext.libraries.extend(pa.get_libraries())
    ext.library_dirs.extend(pa.get_library_dirs())

    if os.name == 'posix':
        ext.extra_compile_args.append('-std=c++11')

    # Try uncommenting the following line on Linux
    # if you get weird linker errors or runtime crashes
    # ext.define_macros.append(("_GLIBCXX_USE_CXX11_ABI", "0"))


setup(ext_modules=ext_modules)

Compile the extension:

python setup.py build_ext --inplace