nikhase / spark-on-gcp-intro

Lessons Learned Using Spark On Google Cloud Platform

Big Data ETL with Apache Spark

In this repo, we see

• how to create a Google Cloud Platform (GCP) Spark cluster with JupyterLab installed
• how to submit a PySpark Job to this cluster
• in notebooks/jupyter: Hands-On introduction to key concepts of Spark
  • spark-intro: Lessons learned after my first Spark application, building a recommender with cross join of customers and products involved

Starting a GCP Dataproc Cluster with JupyterLab

Please make sure that your Google Cloud SDK is >=243.0.0.

First, reate a bucket on Google Cloud Storage (GCS). For example: gs://spark-intro. Be sure to add your bucket in the YAML key configBucket of your cluster configuration YAML. There, you can also see the optionalComponents Anaconda and Jupyter.

Start a Dataproc cluster with Jupyter installed using

gcloud beta dataproc clusters import INSERT_CLUSTER_NAME \
    --source dataproc-jupyter-cluster.yaml \
    --region=europe-west1 \
    --project=YOUR_PROJECT

You can also choose other regions.

You need to upload the Jupyter notebook after the cluster initialization. Use

gsutil -m cp -r . gs://YOUR_BUCKET

Go to the Web Interfaces tab and open JupyterLab. The working dir is in your config bucket at notebooks/jupyter.

Submitting PySpark Jobs to GCP

Please follow the instructions on the GCloud SDK documentation to submit PySpark jobs via the SDK.

For example, to create the transaction data, you can do:

gcloud dataproc jobs submit pyspark datagen/create_transactions.py \
--cluster=YOUR_CLUSTER_NAME \
--region=YOUR_REGION

As said in the Spark Docs, we can provide dependencies as py, zip or egg files. From my experience, it's very convenient to build an egg using setup.py and use this as the dependency.