Sample PySpark-on-YARN Application

This sample PySpark application demonstrates how to dynamically package your Python dependencies and isolate your application from any other jobs running on a YARN cluster. It builds on the discussion @nchammas had with several other PySpark users on SPARK-13587. The key pattern is captured in setup-and-submit.sh.

This sample application also demonstrates how to structure and run your Spark tests, both locally and on Travis CI.

Approach

The gist of the approach is to use spark-submit's --archives option to ship a complete Python virtual environment (Conda works too!) to all your YARN executors, isolated to a specific application. This lets you use any version of Python and any combination of Python libraries with your application, without requiring them to be pre-installed on your cluster. More importantly, this lets PySpark applications with conflicting dependencies run side-by-side on the same YARN cluster without issue.

Caveats

The main downside of this approach is that the setup and distribution of the virtual environment adds many seconds -- or even minutes, depending on the size of the environment -- to the job's initialization time. So depending on your job latency requirements and the complexity of your environment, this approach may not be worth it to you.

This approach also does not cover any non-Python dependencies your PySpark application may have unless a) they are Java or Scala dependencies, which spark-submit already supports via --packages; or b) they are Conda or manylinux packages, which your Python package manager puts in your virtual environment for you.

A typical example of a dependency this approach won't handle is a standalone C library that you need to install using your Linux distribution's package manager. If you want to isolate non-Python dependencies like this to your application, you need to look into deployment options that somehow leverage Linux containers.

Running Tests

Run this project's tests as follows:

spark-submit run_tests.py

Normally, with pytest, you'd do this to run your tests:

PYTHONPATH=./ pytest

However, this won't work here since we need pytest to run within PySpark. Spark 2.2.0 added support for pip installation, which may simplify this pattern. However, we won't use it here.

Running Locally

Although the point of this project is to demonstrate how to run a self-contained PySpark application on a YARN cluster, you can run it locally as follows:

python3 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -r requirements.pip

spark-submit hello.py

Running on YARN

To run this project on YARN, simply set some environment variables and you're good to go.

export HADOOP_CONF_DIR="/path/to/hadoop/conf"
export PYTHON="/path/to/python/executable"
export SPARK_HOME="/path/to/spark/home"
export PATH="$SPARK_HOME/bin:$PATH"

./setup-and-submit.sh

The key pattern for packaging your PySpark application is captured in setup-and-submit.sh.