Running Spark on Alluxio in Kubernetes
Alluxio can be run on Kubernetes. This guide demonstrates how to run a Spark job on Alluxio running in a Kubernetes environment.
- Basic Setup
Spark running on Kubernetes can use Alluxio as the data access layer.
This guide walks through an example Spark job on Alluxio in Kubernetes.
The example used in this tutorial is a job to count the number of lines in a file.
We refer to this job as
count in the following text.
- A Kubernetes cluster (version >= 1.8).
- Alluxio is deployed on the Kubernetes cluster. For instructions on how to deploy Alluxio, refer to this page
First, we’ll prepare a Spark Docker image including the Alluxio client and any other required jars. This image should be made available on all Kubernetes nodes.
Download the desired Spark version.
We use the pre-built binary for the
spark-submit command as well as building the Docker image
using Dockerfile included with Alluxio.
Note: Download the package prebuilt for hadoop
$ tar -xf spark-2.4.4-bin-hadoop2.7.tgz $ cd spark-2.4.4-bin-hadoop2.7
Build the Spark Docker Image
Extract the Alluxio client jar from the Alluxio Docker image:
$ id=$(docker create alluxio/alluxio:2.2.0) $ docker cp $id:/opt/alluxio/client/alluxio-2.2.0-client.jar \ - > alluxio-2.2.0-client.jar $ docker rm -v $id 1>/dev/null
Add the required Alluxio client jar and build a Docker image used for the Spark driver and executor pods. Run the following from the Spark distribution directory to add the Alluxio client jar.
$ cp <path_to_alluxio_client>/alluxio-2.2.0-client.jar jars/
Note: Any jar copied to the
jarsdirectory is included in the Spark Docker image when built.
Build the Spark Docker image
$ docker build -t spark-alluxio -f kubernetes/dockerfiles/spark/Dockerfile .
Note: Make sure all your nodes (where the spark-driver and spark-executor pods will run) have this image.
This section shows how to use the built Docker image to launch a Spark job with Alluxio as the data source.
Short-circuit access enables an Alluxio client in a Spark executor to access the Alluxio worker storage on the host machine directly. This improves performance by not communicating with the Alluxio worker using the networking stack.
If domain sockets were not setup when deploying Alluxio as per instructions on
you can skip mounting the
hostPath volumes to the Spark executors.
If a domain socket location was setup on hosts running the Alluxio worker process at location
/tmp/alluxio-domain with the Alluxio configuration as
alluxio.worker.data.server.domain.socket.address=/opt/domain, use the following Spark
configuration to mount
/opt/domain in the Spark executor pod.
spark-submit command in the following section includes these properties.
spark.kubernetes.executor.volumes.hostPath.alluxio-domain.mount.path=/opt/domain spark.kubernetes.executor.volumes.hostPath.alluxio-domain.mount.readOnly=true spark.kubernetes.executor.volumes.hostPath.alluxio-domain.options.path=/tmp/alluxio-domain spark.kubernetes.executor.volumes.hostPath.alluxio-domain.options.type=Directory
- Volume support in Spark was added in version 2.4.0.
- You may observe a performance hit when not using short-circuit access via a domain socket.
Run a Spark job
The following command runs an example word count job in the Alluxio location
The output and time taken can be seen in the logs for Spark driver pod. Refer to Spark
documentation for further instructions.
Create the service account (optional)
You can create one service account for running the spark job with the required access as below
if you do not have one to use.
$ kubectl create serviceaccount spark $ kubectl create clusterrolebinding spark-role --clusterrole=edit \ --serviceaccount=default:spark --namespace=default
Submit a Spark job
Run the job from the Spark distribution directory
$ ./bin/spark-submit --master k8s://https://<kubernetes-api-server>:8443 \ --deploy-mode cluster --name spark-alluxio --conf spark.executor.instances=1 \ --class org.apache.spark.examples.JavaWordCount \ --driver-memory 500m --executor-memory 1g \ --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \ --conf spark.kubernetes.container.image=spark-alluxio \ --conf spark.kubernetes.executor.volumes.hostPath.alluxio-domain.mount.path=/opt/domain \ --conf spark.kubernetes.executor.volumes.hostPath.alluxio-domain.mount.readOnly=true \ --conf spark.kubernetes.executor.volumes.hostPath.alluxio-domain.options.path=/tmp/alluxio-domain \ --conf spark.kubernetes.executor.volumes.hostPath.alluxio-domain.options.type=Directory \ local:///opt/spark/examples/jars/spark-examples_2.11-2.4.4.jar \ alluxio://<alluxio-master>:19998/LICENSE
Note: You can find the address of the Kubernetes API server by running
kubectl cluster-info. You can find more details in Spark documentation.
Accessing Alluxio Client Logs
The Alluxio client logs can be found in the Spark driver and executor logs. Refer to Spark documentation for further instructions.
HTTP 403 on Kubernetes client
If your spark job failed due to failure in the Kubernetes client like the following:
WARN ExecutorPodsWatchSnapshotSource: Kubernetes client has been closed ... ERROR SparkContext: Error initializing SparkContext. io.fabric8.kubernetes.client.KubernetesClientException
This is probably due to a known issue
that can be resolved by upgrading
kubernetes-client.jar to 4.4.x.
You can patch the docker image by updating the
kubernetes-client-x.x.jar before building the
rm spark-2.4.4-bin-hadoop2.7/jars/kubernetes-client-*.jar wget https://repo1.maven.org/maven2/io/fabric8/kubernetes-client/4.4.2/kubernetes-client-4.4.2.jar cp kubernetes-client-4.4.2.jar spark-2.4.4-bin-hadoop2.7/jars
Then build the
spark-alluxio image and distribute to all your nodes.
Service account does not have access
If you see errors like below complaining some operations are forbidden, that is because the service account you use for the spark job does not have enough access to perform the action.
ERROR Utils: Uncaught exception in thread main io.fabric8.kubernetes.client.KubernetesClientException: Failure executing: DELETE at: \ https://kubernetes.default.svc/api/v1/namespaces/default/pods/spark-alluxiolatest-exec-1. \ Message: Forbidden!Configured service account doesn't have access. Service account may have been revoked. \ pods "spark-alluxiolatest-exec-1" is forbidden: User "system:serviceaccount:default:default" \ cannot delete resource "pods" in API group "" in the namespace "default".
You should ensure you have the correct access by creating a service account.