Installation

PySpark is included in the official releases of Spark available in the Apache Spark website. For Python users, PySpark also provides pip installation from PyPI. This is usually for local usage or as a client to connect to a cluster instead of setting up a cluster itself.

This page includes instructions for installing PySpark by using pip, Conda, downloading manually, and building from the source.

Python Version Supported

Python 3.6 and above.

Using PyPI

PySpark installation using PyPI is as follows:

pip install pyspark

If you want to install extra dependencies for a specific component, you can install it as below:

# Spark SQL
pip install pyspark[sql]
# pandas API on Spark
pip install pyspark[pandas_on_spark] plotly  # to plot your data, you can install plotly together.

For PySpark with/without a specific Hadoop version, you can install it by using PYSPARK_HADOOP_VERSION environment variables as below:

PYSPARK_HADOOP_VERSION=2.7 pip install pyspark

The default distribution uses Hadoop 3.2 and Hive 2.3. If users specify different versions of Hadoop, the pip installation automatically downloads a different version and use it in PySpark. Downloading it can take a while depending on the network and the mirror chosen. PYSPARK_RELEASE_MIRROR can be set to manually choose the mirror for faster downloading.

PYSPARK_RELEASE_MIRROR=http://mirror.apache-kr.org PYSPARK_HADOOP_VERSION=2.7 pip install

It is recommended to use -v option in pip to track the installation and download status.

PYSPARK_HADOOP_VERSION=2.7 pip install pyspark -v

Supported values in PYSPARK_HADOOP_VERSION are:

  • without: Spark pre-built with user-provided Apache Hadoop

  • 2.7: Spark pre-built for Apache Hadoop 2.7

  • 3.2: Spark pre-built for Apache Hadoop 3.2 and later (default)

Note that this installation way of PySpark with/without a specific Hadoop version is experimental. It can change or be removed between minor releases.

Using Conda

Conda is an open-source package management and environment management system (developed by Anaconda), which is best installed through Miniconda or Miniforge. The tool is both cross-platform and language agnostic, and in practice, conda can replace both pip and virtualenv.

Conda uses so-called channels to distribute packages, and together with the default channels by Anaconda itself, the most important channel is conda-forge, which is the community-driven packaging effort that is the most extensive & the most current (and also serves as the upstream for the Anaconda channels in most cases).

To create a new conda environment from your terminal and activate it, proceed as shown below:

conda create -n pyspark_env
conda activate pyspark_env

After activating the environment, use the following command to install pyspark, a python version of your choice, as well as other packages you want to use in the same session as pyspark (you can install in several steps too).

conda install -c conda-forge pyspark  # can also add "python=3.8 some_package [etc.]" here

Note that PySpark for conda is maintained separately by the community; while new versions generally get packaged quickly, the availability through conda(-forge) is not directly in sync with the PySpark release cycle.

While using pip in a conda environment is technically feasible (with the same command as above), this approach is discouraged, because pip does not interoperate with conda.

For a short summary about useful conda commands, see their cheat sheet.

Manually Downloading

PySpark is included in the distributions available at the Apache Spark website. You can download a distribution you want from the site. After that, uncompress the tar file into the directory where you want to install Spark, for example, as below:

tar xzvf spark-3.0.0-bin-hadoop2.7.tgz

Ensure the SPARK_HOME environment variable points to the directory where the tar file has been extracted. Update PYTHONPATH environment variable such that it can find the PySpark and Py4J under SPARK_HOME/python/lib. One example of doing this is shown below:

cd spark-3.0.0-bin-hadoop2.7
export SPARK_HOME=`pwd`
export PYTHONPATH=$(ZIPS=("$SPARK_HOME"/python/lib/*.zip); IFS=:; echo "${ZIPS[*]}"):$PYTHONPATH

Installing from Source

To install PySpark from source, refer to Building Spark.

Dependencies

Package

Minimum supported version

Note

pandas

0.23.2

Optional for Spark SQL

NumPy

1.7

Required for MLlib DataFrame-based API

pyarrow

1.0.0

Optional for Spark SQL

Py4J

0.10.9.5

Required

pandas

0.23.2

Required for pandas API on Spark

pyarrow

1.0.0

Required for pandas API on Spark

Numpy

1.14

Required for pandas API on Spark

Note that PySpark requires Java 8 or later with JAVA_HOME properly set. If using JDK 11, set -Dio.netty.tryReflectionSetAccessible=true for Arrow related features and refer to Downloading.

Note for AArch64 (ARM64) users: PyArrow is required by PySpark SQL, but PyArrow support for AArch64 is introduced in PyArrow 4.0.0. If PySpark installation fails on AArch64 due to PyArrow installation errors, you can install PyArrow >= 4.0.0 as below:

pip install "pyarrow>=4.0.0" --prefer-binary