airflow-livy-plugins 0.2

Plugins for Airflow to run Spark jobs via Livy: sessions and batches

Airflow Livy Plugins

Plugins for Airflow to run Spark jobs via Livy:

• Sessions,
• Batches. This mode supports additional verification via Spark/YARN REST API.

See this blog post for more information and detailed comparison of ways to run Spark jobs from Airflow.

Directories and files of interest

• airflow_home: example DAGs and plugins for Airflow. Can be used as Airflow home path.
• batches: Spark jobs code, to be used in Livy batches.
• sessions: (Optionally) templated Spark code for Livy sessions.
• airflow.sh: helper shell script. Can be used to run sample DAGs, prep development environment and more. Run it to find out what other commands are available.

How do I...

...run the examples?

Prerequisites:

• Python 3. Make sure it's installed and in $PATH

Now,

1. Do you have a Spark cluster with Livy running somewhere?
   1. No. Either get one, or "mock" it with my Spark cluster on Docker Compose.
   2. Yes. You're golden!
2. Optional - this step can be skipped if you're mocking a cluster on your machine. Open airflow.sh. Inside init_airflow () function you'll see Airflow Connections for Livy, Spark and YARN. Redefine as appropriate.
3. run ./airflow.sh up to bring up the whole infrastructure. Airflow UI will be available at localhost:8080.
4. Ctrl+C to stop Airflow. Then ./airflow.sh down to dispose of remaining Airflow processes (shouldn't be needed there if everything goes well).

... use it in my project?

pip install airflow-livy-plugins

Then link or copy the plugin files into $AIRFLOW_HOME/plugins (see how I do that in ./airflow.sh). They'll get loaded into Airflow via Plugin Manager automatically. This is how you import the plugins:

from airflow.operators import LivySessionOperator
from airflow.operators import LivyBatchOperator

Plugins are loaded at run-time so the imports above will look broken in your IDE, but will work fine in Airflow. Take a look at the sample DAGs to see my walkaround :)

... set up the development environment?

Alright, you want to contribute and need to be able to run the stuff on your machine, as well as the usual niceness that comes with IDEs (debugging, syntax highlighting). How do I

• run ./airflow.sh dev to install all dev dependencies.
• ./airflow.sh updev runs local Airflow with local plugins (as opposed to pulling them from PyPi)
• (Pycharm-specific) point PyCharm to your newly-created virtual environment: go to "Preferences" -> "Project: airflow-livy-plugins" -> "Project interpreter", select "Existing environment" and pick python3 executable from venv folder (venv/bin/python3)
• ./airflow.sh cov - run tests with coverage report (will be saved to htmlcov/).
• ./airflow.sh lint - highlight code style errors.
• ./airflow.sh format to reformat all code (Black + isort)

... debug?

• (Pycharm-specific) Step-by-step debugging with airflow test and running PySpark batch jobs locally (with debugging as well) is supported via run configurations under .idea/runConfigurations. You shouldn't have to do anything to use them - just open the folder in PyCharm as a project.
• An example of how a batch can be ran on local Spark:

python ./batches/join_2_files.py \
"file:////Users/vpanov/data/vpanov/bigdata-docker-compose/data/grades.csv" \
"file:///Users/vpanov/data/vpanov/bigdata-docker-compose/data/ssn-address.tsv" \
-file1_sep=, -file1_header=true \
-file1_schema="\`Last name\` STRING, \`First name\` STRING, SSN STRING, Test1 INT, Test2 INT, Test3 INT, Test4 INT, Final INT, Grade STRING" \
-file1_join_column=SSN -file2_header=false \
-file2_schema="\`Last name\` STRING, \`First name\` STRING, SSN STRING, Address1 STRING, Address2 STRING" \
-file2_join_column=SSN -output_header=true \
-output_columns="file1.\`Last name\` AS LastName, file1.\`First name\` AS FirstName, file1.SSN, file2.Address1, file2.Address2" 

# Optionally append to save result to file
#-output_path="file:///Users/vpanov/livy_batch_example" 

TODO

• airflow.sh - replace with modern tools (e.g. pipenv + Docker image)
• Disable some of flake8 flags for cleaner code