Python Development

This page provides general Python development guidelines and source build instructions for all platforms.

Coding Style

We follow a similar PEP8-like coding style to the pandas project.

The code must pass flake8 (available from pip or conda) or it will fail the build. Check for style errors before submitting your pull request with:

flake8 .
flake8 --config=.flake8.cython .

The package autopep8 (also available from pip or conda) can automatically fix many of the errors reported by flake8:

autopep8 --in-place ../integration/
autopep8 --in-place --global-config=.flake8.cython pyarrow/table.pxi

Unit Testing

We are using pytest to develop our unit test suite. After building the project (see below) you can run its unit tests like so:

pytest pyarrow

Package requirements to run the unit tests are found in requirements-test.txt and can be installed if needed with pip install -r requirements-test.txt.

The project has a number of custom command line options for its test suite. Some tests are disabled by default, for example. To see all the options, run

pytest pyarrow --help

and look for the “custom options” section.

Test Groups

We have many tests that are grouped together using pytest marks. Some of these are disabled by default. To enable a test group, pass --$GROUP_NAME, e.g. --parquet. To disable a test group, prepend disable, so --disable-parquet for example. To run only the unit tests for a particular group, prepend only- instead, for example --only-parquet.

The test groups currently include:

  • gandiva: tests for Gandiva expression compiler (uses LLVM)

  • hdfs: tests that use libhdfs or libhdfs3 to access the Hadoop filesystem

  • hypothesis: tests that use the hypothesis module for generating random test cases. Note that --hypothesis doesn’t work due to a quirk with pytest, so you have to pass --enable-hypothesis

  • large_memory: Test requiring a large amount of system RAM

  • orc: Apache ORC tests

  • parquet: Apache Parquet tests

  • plasma: Plasma Object Store tests

  • s3: Tests for Amazon S3

  • tensorflow: Tests that involve TensorFlow


For running the benchmarks, see Benchmarks.

Building on Linux and MacOS

System Requirements

On macOS, any modern XCode (6.4 or higher; the current version is 10) is sufficient.

On Linux, for this guide, we require a minimum of gcc 4.8, or clang 3.7 or higher. You can check your version by running

$ gcc --version

If the system compiler is older than gcc 4.8, it can be set to a newer version using the $CC and $CXX environment variables:

export CC=gcc-4.8
export CXX=g++-4.8

Environment Setup and Build

First, let’s clone the Arrow git repository:

mkdir repos
cd repos
git clone

You should now see

$ ls -l
total 8
drwxrwxr-x 12 wesm wesm 4096 Apr 15 19:19 arrow/

Using Conda


Using conda to build Arrow on macOS is complicated by the fact that the conda-forge compilers require an older macOS SDK. Conda offers some installation instructions; the alternative would be to use Homebrew and pip instead.

Let’s create a conda environment with all the C++ build and Python dependencies from conda-forge, targeting development for Python 3.7:

On Linux and macOS:

conda create -y -n pyarrow-dev -c conda-forge \
    --file arrow/ci/conda_env_unix.yml \
    --file arrow/ci/conda_env_cpp.yml \
    --file arrow/ci/conda_env_python.yml \
    compilers \

As of January 2019, the compilers package is needed on many Linux distributions to use packages from conda-forge.

With this out of the way, you can now activate the conda environment

conda activate pyarrow-dev

For Windows, see the Building on Windows section below.

We need to set some environment variables to let Arrow’s build system know about our build toolchain:


Using pip


If you installed Python using the Anaconda distribution or Miniconda, you cannot currently use virtualenv to manage your development. Please follow the conda-based development instructions instead.

On macOS, use Homebrew to install all dependencies required for building Arrow C++:

brew update && brew bundle --file=arrow/cpp/Brewfile

See here for a list of dependencies you may need.

On Debian/Ubuntu, you need the following minimal set of dependencies. All other dependencies will be automatically built by Arrow’s third-party toolchain.

$ sudo apt-get install libjemalloc-dev libboost-dev \
                       libboost-filesystem-dev \
                       libboost-system-dev \
                       libboost-regex-dev \
                       python-dev \
                       autoconf \
                       flex \

If you are building Arrow for Python 3, install python3-dev instead of python-dev.

On Arch Linux, you can get these dependencies via pacman.

$ sudo pacman -S jemalloc boost

Now, let’s create a Python virtualenv with all Python dependencies in the same folder as the repositories and a target installation folder:

virtualenv pyarrow
source ./pyarrow/bin/activate
pip install six numpy pandas cython pytest hypothesis

# This is the folder where we will install the Arrow libraries during
# development
mkdir dist

If your cmake version is too old on Linux, you could get a newer one via pip install cmake.

We need to set some environment variables to let Arrow’s build system know about our build toolchain:

export ARROW_HOME=$(pwd)/dist
export LD_LIBRARY_PATH=$(pwd)/dist/lib:$LD_LIBRARY_PATH

Build and test

Now build and install the Arrow C++ libraries:

mkdir arrow/cpp/build
pushd arrow/cpp/build

      -DARROW_ORC=ON \
make -j4
make install

Many of these components are optional, and can be switched off by setting them to OFF:

  • ARROW_FLIGHT: RPC framework

  • ARROW_GANDIVA: LLVM-based expression compiler

  • ARROW_ORC: Support for Apache ORC file format

  • ARROW_PARQUET: Support for Apache Parquet file format

  • ARROW_PLASMA: Shared memory object store

If multiple versions of Python are installed in your environment, you may have to pass additional parameters to cmake so that it can find the right executable, headers and libraries. For example, specifying -DPYTHON_EXECUTABLE=$VIRTUAL_ENV/bin/python (assuming that you’re in virtualenv) enables cmake to choose the python executable which you are using.


On Linux systems with support for building on multiple architectures, make may install libraries in the lib64 directory by default. For this reason we recommend passing -DCMAKE_INSTALL_LIBDIR=lib because the Python build scripts assume the library directory is lib


If you have conda installed but are not using it to manage dependencies, and you have trouble building the C++ library, you may need to set -DARROW_DEPENDENCY_SOURCE=AUTO or some other value (described here) to explicitly tell CMake not to use conda.

For any other C++ build challenges, see C++ Development.

Now, build pyarrow:

pushd arrow/python
python build_ext --inplace

If you did not build one of the optional components, set the corresponding PYARROW_WITH_$COMPONENT environment variable to 0.

Now you are ready to install test dependencies and run Unit Testing, as described above.

To build a self-contained wheel (including the Arrow and Parquet C++ libraries), one can set --bundle-arrow-cpp:

pip install wheel  # if not installed
python build_ext --build-type=$ARROW_BUILD_TYPE \
       --bundle-arrow-cpp bdist_wheel

Building with CUDA support

The pyarrow.cuda module offers support for using Arrow platform components with Nvidia’s CUDA-enabled GPU devices. To build with this support, pass -DARROW_CUDA=ON when building the C++ libraries, and set the following environment variable when building pyarrow:



Since pyarrow depends on the Arrow C++ libraries, debugging can frequently involve crossing between Python and C++ shared libraries.

Using gdb on Linux

To debug the C++ libraries with gdb while running the Python unit

test, first start pytest with gdb:

gdb --args python -m pytest pyarrow/tests/ -k $TEST_TO_MATCH

To set a breakpoint, use the same gdb syntax that you would when debugging a C++ unitttest, for example:

(gdb) b src/arrow/python/
No source file named src/arrow/python/
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (src/arrow/python/ pending.

Building on Windows

First, we bootstrap a conda environment similar to above, but skipping some of the Linux/macOS-only packages:

First, starting from fresh clones of Apache Arrow:

git clone
 conda create -y -n pyarrow-dev -c conda-forge ^
     --file arrow\ci\conda_env_cpp.yml ^
     --file arrow\ci\conda_env_python.yml ^
conda activate pyarrow-dev

Now, we build and install Arrow C++ libraries

mkdir cpp\build
cd cpp\build
set ARROW_HOME=C:\thirdparty
cmake -G "Visual Studio 14 2015 Win64" ^
      -DARROW_GANDIVA=on ^
      -DARROW_PARQUET=on ^
      -DARROW_PYTHON=on ..
cmake --build . --target INSTALL --config Release
cd ..\..

After that, we must put the install directory’s bin path in our %PATH%:


Now, we can build pyarrow:

cd python
python build_ext --inplace

Then run the unit tests with:

py.test pyarrow -v

Running C++ unit tests for Python integration

Getting python-test.exe to run is a bit tricky because your %PYTHONHOME% must be configured to point to the active conda environment:


Now python-test.exe or simply ctest (to run all tests) should work.

Windows Caveats

Some components are not supported yet on Windows:

  • Flight RPC

  • Plasma