So we're all working on data pipelines every day, but wouldn't be nice to just hit a button and have our code automatically deployed to staging or test accounts? I thought so, too, thats why I created the EMR CLI (emr
) that can help you package and deploy your EMR jobs so you don't have to.
The EMR CLI supports a wide variety of configuration options to adapt to your data pipeline, not the other way around.
- Packaging - Ensure a consistent approach to packaging your production Spark jobs.
- Deployment - Easily deploy your Spark jobs across multiple EMR environments or deployment frameworks like EC2, EKS, and Serverless.
- CI/CD - Easily test each iteration of your code without resorting to messy shell scripts. :)
The initial use cases are:
- Consistent packaging for PySpark projects.
- Use in CI/CD pipelines for packaging, deployment of artifacts, and integration testing.
Warning: This tool is still under active development, so commands may change until a stable 1.0 release is made.
You can use the EMR CLI to take a project from nothing to running in EMR Serverless is 2 steps.
First, let's install the emr
command.
python3 -m pip install -U emr-cli
Note This tutorial assumes you have already setup EMR Serverless and have an EMR Serverless application, job role, and S3 bucket you can use. You can also use the
emr bootstrap
command.
- Create a sample project
emr init scratch
๐ Tip: Use
--project-type poetry
to create a Poetry project!
You should now have a sample PySpark project in your scratch directory.
scratch
โโโ Dockerfile
โโโ entrypoint.py
โโโ jobs
โย ย โโโ extreme_weather.py
โโโ pyproject.toml
1 directory, 4 files
- Now deploy and run on an EMR Serverless application!
emr run \
--entry-point entrypoint.py \
--application-id ${APPLICATION_ID} \
--job-role ${JOB_ROLE_ARN} \
--s3-code-uri s3://${S3_BUCKET}/tmp/emr-cli-demo/ \
--build \
--wait
This command performs the following actions:
- Packages your project dependencies into a python virtual environment
- Uploads the Spark entrypoint and packaged dependencies to S3
- Starts an EMR Serverless job
- Waits for the job to run to a successful completion!
And you're done. Feel free to modify the project to experiment with different things. You can simply re-run the command above to re-package and re-deploy your job.
In many organizations, PySpark is the primary language for writing Spark jobs. But Python projects can be structured in a variety of ways โย a single .py
file, requirements.txt
, setup.py
files, or even poetry
configurations. EMR CLI aims to bundle your PySpark code the same way regardless of which system you use.
While Spark Scala or Java code will be more standard from a packaging perspective, it's still useful to able to easily deploy and run your jobs across multiple EMR environments.
Want to just write some .sql
files and have those deployed? No problem.
- Create a new PySpark project (other frameworks TBD)
emr init project-dir
- Package your project into a virtual environment archive
emr package --entry-point main.py
The EMR CLI auto-detects the project type and will change the packaging method appropriately.
If you have additional .py
files, those will be included in the archive.
- Deploy an existing package artifact to S3.
emr deploy --entry-point main.py --s3-code-uri s3://<BUCKET>/code/
- Deploy a PySpark package to S3 and trigger an EMR Serverless job
emr run --entry-point main.py \
--s3-code-uri s3://<BUCKET>/code/ \
--application-id <EMR_SERVERLESS_APP> \
--job-role <JOB_ROLE_ARN>
- Build, deploy, and run an EMR Serverless job and wait for it to finish.
emr run --entry-point main.py \
--s3-code-uri s3://<BUCKET>/code/ \
--application-id <EMR_SERVERLESS_APP> \
--job-role <JOB_ROLE_ARN> \
--build \
--wait
Note: If the job fails, the command will exit with an error code.
In the future, you'll also be able to do the following:
- Utilize the same code against an EMR on EC2 cluster
emr run --cluster-id j-8675309
- Or an EMR on EKS virtual cluster.
emr run --virtual-cluster-id 654abacdefgh1uziuyackhrs1
See CONTRIBUTING for more information.
This project is licensed under the Apache-2.0 License.