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

./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:

./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:

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

And create a custom core-site.xml.

<configuration>
    <property>
        <name>fs.AbstractFileSystem.o3fs.impl</name>
        <value>org.apache.hadoop.fs.ozone.OzFs</value>
     </property>
</configuration>

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

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

Create a new Dockerfile and build the image:

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

For remote Kubernetes cluster you may need to push it:

docker push myrepo/spark-ozone

Create a bucket

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

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

Create service account to use

kubectl create serviceaccount spark -n yournamespace
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)
bin/spark-submit \
    --master k8s://https://kubernetes:6443 \
    --deploy-mode cluster \
    --name spark-word-count \
    --class org.apache.spark.examples.JavaWordCount \
    --conf spark.executor.instances=1 \
    --conf spark.kubernetes.namespace=yournamespace \
    --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \
    --conf spark.kubernetes.container.image=myrepo/spark-ozone \
    --conf spark.kubernetes.container.image.pullPolicy=Always \
    --jars /opt/hadoop-ozone-filesystem-hadoop2.jar \
    local:///opt/spark/examples/jars/spark-examples_2.11-2.4.0.jar \
    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:

...
name: 8
William: 3
this,': 1
SOUP!': 1
`Silence: 1
`Mine: 1
ordered.: 1
considering: 3
muttering: 3
candle: 2
...