Spark in Kubernetes with OzoneFS

This recipe shows how Ozone object store can be used from Spark using:

  • OzoneFS (Hadoop compatible file system)
  • Hadoop 2.7 (included in the Spark distribution)
  • Kubernetes Spark scheduler
  • Local spark client

Requirements

Download latest Spark and Ozone distribution and extract them. This method is tested with the spark-2.4.6-bin-hadoop2.7 distribution.

You also need the following:

  • A container repository to push and pull the spark+ozone images. (In this recipe we will use the dockerhub)
  • A repo/name for the custom containers (in this recipe myrepo/ozone-spark)
  • A dedicated namespace in kubernetes (we use yournamespace in this recipe)

Create the docker image for drivers

Create the base Spark driver/executor image

First of all create a docker image with the Spark image creator. Execute the following from the Spark distribution

  1. ./bin/docker-image-tool.sh -r myrepo -t 2.4.6 build

Note: if you use Minikube add the -m flag to use the docker daemon of the Minikube image:

  1. ./bin/docker-image-tool.sh -m -r myrepo -t 2.4.6 build

./bin/docker-image-tool.sh is an official Spark tool to create container images and this step will create multiple Spark container images with the name myrepo/spark. The first container will be used as a base container in the following steps.

Customize the docker image

Create a new directory for customizing the created docker image.

Copy the ozone-site.xml from the cluster:

  1. kubectl cp om-0:/opt/hadoop/etc/hadoop/ozone-site.xml .

And create a custom core-site.xml.

  1. <configuration>
  2.     <property>
  3.         <name>fs.AbstractFileSystem.o3fs.impl</name>
  4.         <value>org.apache.hadoop.fs.ozone.OzFs</value>
  5.      </property>
  6. </configuration>

Copy the ozonefs.jar file from an ozone distribution (use the hadoop2 version!)

  1. kubectl cp om-0:/opt/hadoop/share/ozone/lib/ozone-filesystem-hadoop2-VERSION.jar ozone-filesystem-hadoop2.jar

Create a new Dockerfile and build the image:

  1. FROM myrepo/spark:2.4.6
  2. ADD core-site.xml /opt/hadoop/conf/core-site.xml
  3. ADD ozone-site.xml /opt/hadoop/conf/ozone-site.xml
  4. ENV HADOOP_CONF_DIR=/opt/hadoop/conf
  5. ENV SPARK_EXTRA_CLASSPATH=/opt/hadoop/conf
  6. ADD ozone-filesystem-hadoop2.jar /opt/ozone-filesystem-hadoop2.jar
  1. docker build -t myrepo/spark-ozone

For remote Kubernetes cluster you may need to push it:

  1. docker push myrepo/spark-ozone

Create a bucket

Download any text file and put it to the /tmp/alice.txt first.

  1. kubectl port-forward s3g-0 9878:9878
  2. aws s3api --endpoint http://localhost:9878 create-bucket --bucket=test
  3. aws s3api --endpoint http://localhost:9878 put-object --bucket test --key alice.txt --body /tmp/alice.txt

Create service account to use

  1. kubectl create serviceaccount spark -n yournamespace
  2. kubectl create clusterrolebinding spark-role --clusterrole=edit --serviceaccount=yournamespace:spark --namespace=yournamespace

Execute the job

Execute the following spark-submit command, but change at least the following values:

  • the Kubernetes master url (you can check your ~/.kube/config to find the actual value)
  • the Kubernetes namespace (yournamespace in this example)
  • serviceAccountName (you can use the spark value if you followed the previous steps)
  • container.image (in this example this is myrepo/spark-ozone. This is pushed to the registry in the previous steps)
  1. bin/spark-submit \
  2.     --master k8s://https://kubernetes:6443 \
  3.     --deploy-mode cluster \
  4.     --name spark-word-count \
  5.     --class org.apache.spark.examples.JavaWordCount \
  6.     --conf spark.executor.instances=1 \
  7.     --conf spark.kubernetes.namespace=yournamespace \
  8.     --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \
  9.     --conf spark.kubernetes.container.image=myrepo/spark-ozone \
  10.     --conf spark.kubernetes.container.image.pullPolicy=Always \
  11.     --jars /opt/ozone-filesystem-hadoop2.jar \
  12.     local:///opt/spark/examples/jars/spark-examples_2.11-2.4.0.jar \
  13.     o3fs://test.s3v.ozone-om-0.ozone-om:9862/alice.txt

Check the available spark-word-count-... pods with kubectl get pod

Check the output of the calculation with
kubectl logs spark-word-count-1549973913699-driver

You should see the output of the wordcount job. For example:

  1. ...
  2. name: 8
  3. William: 3
  4. this,': 1
  5. SOUP!': 1
  6. `Silence: 1
  7. `Mine: 1
  8. ordered.: 1
  9. considering: 3
  10. muttering: 3
  11. candle: 2
  12. ...