Installing PyArrow#

System Compatibility#

PyArrow is regularly built and tested on Windows, macOS and various Linux distributions. We strongly recommend using a 64-bit system.

Python Compatibility#

PyArrow is currently compatible with Python 3.10, 3.11, 3.12, 3.13 and 3.14.

Using Conda#

Install the latest version of PyArrow from conda-forge using Conda:

conda install -c conda-forge pyarrow

Using Pip#

Install the latest version from PyPI (Windows, Linux, and macOS):

pip install pyarrow

If you encounter any importing issues of the pip wheels on Windows, you may need to install the latest Visual C++ Redistributable for Visual Studio.

Installing nightly packages or from source#

See Python Development.

Dependencies#

Optional dependencies

• NumPy 1.21.2 or higher.

• pandas 1.3.4 or higher,

• cffi.

Additional packages PyArrow is compatible with are fsspec and pytz, dateutil or tzdata package for timezones.

tzdata on Windows#

On Linux and macOS, Arrow uses the OS-provided timezone database. On Windows, Arrow uses the Windows timezone database when built with MSVC or recent MinGW GCC (version 13+), which covers most pre-built packages. No additional setup is needed for these builds.

However, when PyArrow is built with Clang/libc++ on Windows, a user-provided IANA timezone database is required. To download and extract the text version of the IANA timezone database follow the instructions in the C++ Runtime Dependencies or use the (deprecated) pyarrow utility function pyarrow.util.download_tzdata_on_windows().

By default, the timezone database will be detected at %USERPROFILE%\Downloads\tzdata. If the database has been downloaded in a different location, you will need to set a custom path to the database from Python using the (deprecated) pa.set_timezone_db_path("custom_path") function.

import tzdata
print(tzdata.__file__)  # path\to\.venv\Lib\site-packages\tzdata\__init__.py

Differences between conda-forge packages#

On conda-forge, PyArrow is published as three separate packages, each providing varying levels of functionality. This is in contrast to PyPi, where only a single PyArrow package is provided.

The purpose of this split is to minimize the size of the installed package for most users (pyarrow), provide a smaller, minimal package for specialized use cases (pyarrow-core), while still providing a complete package for users who require it (pyarrow-all). What was historically pyarrow on conda-forge is now pyarrow-all, though most users can continue using pyarrow.

The pyarrow-core package includes the following functionality:

• Data Types and In-Memory Data Model

• Compute Functions (i.e., pyarrow.compute)

• Memory and IO Interfaces

• Streaming, Serialization, and IPC (i.e., pyarrow.ipc)

• Filesystem Interface (i.e., pyarrow.fs. Note: Its planned to move cloud fileystems (i.e., S3, GCS, etc) into pyarrow in a future release though Local FS will remain in pyarrow-core.)

• File formats: Arrow/Feather, JSON, CSV, ORC (but not Parquet)

The pyarrow package adds the following:

• Acero (i.e., pyarrow.acero)

• Tabular Datasets (i.e., pyarrow.dataset)

• Parquet (i.e., pyarrow.parquet)

• Substrait (i.e., pyarrow.substrait)

Finally, pyarrow-all adds:

• Arrow Flight RPC and Flight SQL (i.e., pyarrow.flight)

• Gandiva (i.e., pyarrow.gandiva)

The following table lists the functionality provided by each package and may be useful when deciding to use one package over another or when Creating A Custom Selection.

Creating A Custom Selection#

If you know which components you need and want to control whats installed, you can create a custom selection of packages to include only the extra features you need. For example, to install pyarrow-core and add support for reading and writing Parquet, install libparquet alongside pyarrow-core:

conda install -c conda-forge pyarrow-core libparquet

Or if you wish to use pyarrow but need support for Flight RPC:

conda install -c conda-forge pyarrow libarrow-flight