akoshel

Posted on Oct 20, 2022 • Updated on Apr 7, 2023

How to run Spark on kubernetes in jupyterhub

#spark #jupyterhub #kubernetes #tutorial

This is a basic tutorial on how to run Spark in client mode from jupyterhub notebook.
All required files are presented here https://github.com/akoshel/spark-k8s-jupyterhub

Final architecture

Motivation

I found a lot of tutorials on this topic and almost all of them have custom spark and jupyterhub deployment. So I decided to minimize custom configuration and use raw open-source solutions as it is possible.

Install minikube & helm

Firstly we should create k8s infrastructure
Minikube installation instruction https://minikube.sigs.k8s.io/docs/start/
Helm installation instruction https://helm.sh/docs/intro/install/

Make local docker images available from minikube:



eval $(minikube docker-env)

Install spark

Let's install spark locally.
Further, we will build a spark image and run the spark-pi example with spark-submit



sudo apt-get -y install openjdk-8-jdk-headless
wget https://downloads.apache.org/spark/spark-3.2.2/spark-3.2.2-bin-hadoop3.2.tgz
tar xvf spark-3.2.2-bin-hadoop3.2.tgz
sudo mv spark-3.2.2-bin-hadoop3.2 /opt/spark

Build spark image

Spark has kubernetes dockerfile. Let's build spark image



cat /opt/spark/kubernetes/dockerfiles/spark/Dockerfile
cd /opt/spark
docker build -t spark:latest -f kubernetes/dockerfiles/spark/Dockerfile .

Spark base image does not support python. So we should build pyspark image (opt/spark/spark-3.2.2-bin-hadoop3.2/kubernetes/dockerfiles/spark/bindings/python/Dockerfile).
The basic image does not support s3a and postgres. That is why maven jars should be added.
See modified image here https://github.com/akoshel/spark-k8s-jupyterhub/blob/main/pyspark.Dockerfile

Build pyspark image



cd /opt/spark
docker build -t pyspark:latest -f kubernetes/dockerfiles/spark/bindings/python/Dockerfile .

Run spark-pi

Before running examples namespace, service account, role and rolebinding should be deployed.



kubectl apply -f spark_namespace.yaml
kubectl apply -f spark_sa.yaml
kubectl apply -f spark_sa_role.yaml

Now we are ready to check the spark-pi example using spark-submit
(Use kubectl cluster-info to find your master address)



/opt/spark/bin/spark-submit \
  --master k8s://https://192.168.49.2:8443 \
  --deploy-mode cluster \
  --driver-memory 1g \
  --conf spark.kubernetes.memoryOverheadFactor=0.5 \
  --name sparkpi-test1 \
  --class org.apache.spark.examples.SparkPi \
  --conf spark.kubernetes.container.image=spark:latest \
  --conf spark.kubernetes.driver.pod.name=spark-test1-pi \
  --conf spark.kubernetes.namespace=spark \
  --conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \
  --verbose \
  local:///opt/spark/examples/jars/spark-examples_2.12-3.2.1.jar 1000

Check logs



kubectl logs -n spark spark-test1-pi | grep "Pi is roughly"
Pi is roughly 3.1416600314166003

Great! spark is running on k8s.

Install jupyterhub

Before jupyterhub installation service account, role and rolebinding should be deployed in jupyterhub namespace



kubectl apply -f jupyterhub_sa.yaml
kubectl apply -f jupyterhub_sa_role.yaml

Spark executors have to be deployed in spark namespace from a notebook which is deployed in jupyterhub.
That is why we have to deploy driver service. (driver_service.yaml)



kubectl apply -f driver_service.yaml

To get access to spark UI ingress should be deployed



kubectl apply -f driver_ingress.yaml

Java is not installed in the default jupyterhub singleuser image.
Build modified singleuser image.



docker build -f singleuser.Dockerfile -t singleuser:v1 .

See jhub_values.yaml. There are the following modifications: new image, service account and resources.
Now we are ready to deploy jupyterhub



helm upgrade --cleanup-on-fail \
--install jupyterhub jupyterhub/jupyterhub \
--namespace jupyterhub \
--create-namespace \
--version=2.0.0 \
--values jhub_values.yaml

The easiest way to get access to jupyterhub is port-forwarding from the proxy pod. Alternatively you can configure ingress in jhub_values.yaml



kubectl port-forward proxy-dd5964d5b-6lkwp  -n jupyterhub  8000:8000 # Set your pod name

Pyspark from jupyterhub

Open jupyterhub in your browser http://localhost:8000/
Create a jupyterhub terminal and install pyspark version that matches spark version in the image



pip install pyspark==3.2.2

Create notebook

Create SparkContext



from pyspark import SparkConf, SparkContext
conf = (SparkConf().setMaster("k8s://https://192.168.49.2:8443") # Your master address name
        .set("spark.kubernetes.container.image", "pyspark:latest") # Spark image name
        .set("spark.driver.port", "2222") # Needs to match svc
        .set("spark.driver.blockManager.port", "7777")
        .set("spark.driver.host", "driver-service.jupyterhub.svc.cluster.local") # Needs to match svc
        .set("spark.driver.bindAddress", "0.0.0.0")
        .set("spark.kubernetes.namespace", "spark")
        .set("spark.kubernetes.authenticate.driver.serviceAccountName", "spark")
        .set("spark.kubernetes.authenticate.serviceAccountName", "spark")
        .set("spark.executor.instances", "2")
        .set("spark.kubernetes.container.image.pullPolicy", "IfNotPresent")
       .set("spark.app.name", "tutorial_app"))

Run spark application



# Calculate the approximate sum of values in the dataset
t = sc.parallelize(range(10))
r = t.sumApprox(3)
print('Approximate sum: %s' % r)

Approximate sum: 45.0

See executor pods



kubectl get pods -n spark
NAME                                   READY   STATUS    RESTARTS   AGE
tutorial-app-d63d4c83e68ed465-exec-1   1/1     Running   0          16s
tutorial-app-d63d4c83e68ed465-exec-2   1/1     Running   0          15s

Congratulations! Pyspark in client mode is running from jupyterhub

Further steps: