elaaf/spocker

Containerized Spark Deployment using Docker-Swarm.

Spocker : Spark on Docker

Containerized Spark deployment using Docker.
This repository refers to containerized Spark as Spocker.
Kubernetes should be preferred as the orchestrator for prodcution systems over Docker-Swarm.

Contents

• Single-Host Spark using Docker
  • Requirements
  • Deployment Steps
    • 1. Clone this repository
    • 2. Execute the docker-compose.yml
    • 3. Submit a job to the single-host Spark cluster
• Multi-Host Spark using Docker Swarm
  • Requirements
  • Deployment Steps
    • 1. Clone this repository
    • 2. Setup Docker Swarm
    • 3. Deploy Spark to Docker Swarm
    • 4. Submit a Spark job
  • TO-DO:

Single-Host Spark using Docker

Deploy containerized Spark (Standalone) inside a single machine (host) using Docker.

Requirements

• Docker - on (all) Linux Machine(s)

Install docker from here if required.

To deploy and run Spark (Standalone) inside a single machine(host), run the following commands. This runs the spark-master, and spark-worker containers on the same machine/docker-host.

Steps

1. Clone this repo.
2. Execute the docker compose file.
3. Submit a spark-job to the Spark cluster.

1. Clone this repository

git clone https://github.com/elaaf/spocker.git

cd spocker

2. Execute the docker-compose.yml

docker-compose -f spark/single-host-docker-compose/docker-compose.yml up

This will pull the latest images for spark-master and spark-worker. A docker bridge network named spark-net will be created and joined by the containers.

Three containers will be started on the host-machine:

• 1 x Spark-Master: WebUI @ Your-Machine-IP:8080
• 2 x Spark-Worker WEBUI @ Your-Machine-IP:8081 and Your-Machine-IP:8082

If the deployement was successful, the Spark WebUI should be accessible @ YOUR-IP:8080 and should look something like this:

3. Submit a job to the single-host Spark cluster

Start the spark-submit container, providing the SPARK_MASTER_URL (optional for single-host: defaults to spark://spark-master:7077) and JOB_PATH.
The contianer will run till the submitted spark-job's life, hosting the spark-driver inside it.

docker run -it \
--network spark-net \
--mount type=bind,src=$(pwd),dst=/spocker/ \
-e JOB_PATH=/spocker/spark/spark-test-job/spark_submit_test_pi.py \
elaaf/spark-submit

The --mount command will mount (bind-type) the current directory at the path /spocker/ inside the spark-submit docker container.
This allows the container to access the code/py-files inside the mounted directory.
The Spark WebUI will show a running application and will move it to completed on completion.

Multi-Host Spark using Docker Swarm

Deploy containerized Spark (Standalone) up-and-running on a cluster using Docker Swarm which is a multi-host container orchestartion tool.

Requirements

• Multiple nodes/VMs (atleast 2) on the same network, with internet access on atleast one VM.
• Docker installed on each node/VM.

Steps

1. Clone this repository
2. Setup a Docker swarm on the connected VMs.
3. Deploy the dockerized Spark stack from the Docker Swarm Manager/Leader.
4. Submit a spark-job to the Spark cluster.

1. Clone this repository

Designate one of the node/VM as Swarm Manager/Leader.
Be mindful that most commands will be executed ONLY on the Swarm Manager/Lead node/VM.
Docker will automatically propogate the required data to the Swarm Workers Nodes.

Clone this repository:

git clone https://github.com/elaaf/spocker.git

cd spocker

2. Setup Docker Swarm

From the node/VM designated as swarm Manager/Leader, initialize docker swarm:

docker swarm init

This will initialize the node/vm as the Swarm Leader, along with an join-token to join the swarm:

• On all the other nodes/VMs run the following command to join the Docker swarm:

# Run on all other nodes/VMs except, Swarm Leader
docker swarm join --token <swarm-join-token> <swarm-leader-ip>:<port>

• Verify that all nodes/VMs have joined the Swarm, run:

docker node ls

You should get an output similar to below, listing all the nodes in the Swarm and indicating the Swarm Leader:

3. Deploy Spark to Docker Swarm

Spark can be deployed across all nodes/VMs inside the Docker Swarm by executing the compose file with the following command: NOTE: Execute the command ONLY on the Swarm Leader/Manager

# Run this ONLY on the Swarm Leader 
docker stack deploy spocker --compose-file spark/docker-compose.yml

This will start the following containers :

• 1 X Spark-Master: WebUI @ :8080
• 1 X Spark-Worker on each Swarm Worker Node: WEBUI @ :8080

NOTE: This docker stack deployment configuration connects the containers to a new swarm-scope overlay network called spark-net. More details about Docker Overlay Networks here.

4. Submit a Spark job

Start the spark-submit container, providing the SPARK_MASTER_URL and JOB_PATH.
The contianer will run till the submitted spark-job's life, hosting the spark-driver inside it.

Run the following command on the Swarm Leader:

docker run -it \
--network spark-net \
--mount type=bind,src=$(pwd),dst=/spocker/ \
-e JOB_PATH=/spocker/spark/spark-test-job/spark_submit_test_pi.py \
-e SPARK_MASTER_URL=spark://spark-master:7077 \
-e DRIVER_CORES=1 \
-e DRIVER_MEMORY=4g \
-e NUM_EXECUTORS=1 \
-e EXECUTOR_CORES=1 \
-e EXECUTOR_MEMORY=1g \
elaaf/spark-submit

The --mount command will mount (bind-type) the current directory at the path /spocker/ inside the spark-submit docker container.
This allows the container to access the code/py-files inside the mounted directory.

The Spark WebUI will show a running application (similar to the above image) and will move it to completed on completion.

TO-DO:

• Implement Containerized Hadoop (HDFS, YARN)
• Integrate Containerized Spark (YARN) and Hadoop (HDFS)