Running Sparkling Water in Kubernetes

Sparkling Water can be executed inside of the Kubernetes cluster. Please note that we currently support only internal backend of sparkling water with cluster deployment mode, which means that both executors and driver are running inside Kubernetes. Sparkling Water supports Kubernetes since Spark version 2.3.

Prerequisites:

• Sparkling Water Distribution 3.28.1.2-1-2.3

• Kubernetes Cluster

• Apache Spark 2.3.4

To start Sparkling Water on Kubernetes, the steps are:

0. Please make yourself familiar with how to run Spark on Kubernetes at Spark Kubernetes documentation.

1. Ensure you have SPARK_HOME set up to home of your Spark distribution.

2. Create the Sparkling Water base image for Kubernetes:

   Run ./bin/build-kubernetes-images.sh script inside of the Sparkling Water distribution, which can be downloaded from H2O Download page. The script takes one argument which can be either scala, python or r and creates a docker image for that specific Sparkling Water client.

   Note: Make sure that your Docker environment is the one managed by Kubernetes as Kubernetes needs to see the created images.

   Scala

   ./bin/build-kubernetes-images.sh scala
   

   Python

   ./bin/build-kubernetes-images.sh python
   

3. Run kubectl cluster-info to obtain Kubernetes master URL.

4. Create a custom Dockerfile with your application resources inside and build the image:

   Scala

   cat <<EOT > Dockerfile-Scala-CustomApp
   FROM sparkling-water-scala:3.28.1.2-1-2.3
   COPY ./app.jar "/opt/app.jar"
   EOT
   
   docker build -t "sparkling-water-scala-custom-app:3.28.1.2-1-2.3" -f Dockerfile-Scala-CustomApp .
   

   Python

   cat <<EOT > Dockerfile-Python-CustomApp
   FROM sparkling-water-python:3.28.1.2-1-2.3
   COPY ./app.py "/opt/app.py"
   EOT
   
   docker build -t "sparkling-water-python-custom-app:3.28.1.2-1-2.3" -f  Dockerfile-Python-CustomApp .
   

5. Start Sparkling Water with 3 worker nodes:

   Scala

   $SPARK_HOME/bin/spark-submit \
   --master k8s://IP:PORT \
   --deploy-mode cluster \
   --name CustomApplication \
   --class custom.app.class \
   --conf spark.kubernetes.container.image=sparkling-water-scala-custom-app:3.28.1.2-1-2.3 \
   --conf spark.executor.instances=3 \
   local:///opt/app.jar
   

   Python

   $SPARK_HOME/bin/spark-submit \
   --master k8s://IP:PORT \
   --deploy-mode cluster \
   --name CustomApplication \
   --conf spark.kubernetes.container.image=sparkling-water-python-custom-app:3.28.1.2-1-2.3 \
   --conf spark.executor.instances=3 \
   local:///opt/app.py
   

The IP:PORT represents the Kubernetes master obtained in step 3. It is important to mention that the application resource (the last argument) needs to be available in the docker image.

After this step, your job is submitted into Kubernetes cluster. You can see the logs by running kubectl logs pod_id, where you can get the pod id of the desired executor or driver by running kubectl get pods.