Getting started / Spark application

• Setup your local environment

• Spark on K8S

  • Setup liminal
  • Liminal YAML walkthrough
  • Evaluate the Iris Classification model
  • Debugging Kubernetes Deployments

• Closing up

In this tutorial, we will guide you through setting up Apache Liminal on your local machine and run a simple machine-learning workflow, based on the classic Iris dataset classification example including a feature engineering with Spark engine.
More details in this link.

Prerequisites

• Python 3 (3.6 and up)
• Python Virtual Environments
• Docker Desktop
• Kubernetes CLI (kubectl)

Note: Make sure kubernetes cluster is running in docker desktop

Setup your local env environment

In the dev folder, clone the example code from liminal:

git clone https://github.com/apache/incubator-liminal

Note: You just cloned the entire Liminal Project, you actually only need examples folder.

Create a python virtual environment to isolate your runs:

cd incubator-liminal/examples/spark-app-demo/
python3 -m venv env

Activate your virtual environment:

source env/bin/activate

Now we are ready to install liminal:

pip install apache-liminal

Spark On K8S

We will define the following steps and services to implement the Iris classification example including a simple feature engineering with Apache Spark:
Clean data, Train, Validate & Deploy - Cleaning the input data set and prepare it for training and validation execution is managed by Liminal Airflow extension. The training task trains a regression model using a public dataset.
We then validate the model and deploy it to a model-store in mounted volume.
Inference - online inference is done using a Python Flask service running on the local Kubernetes in docker desktop. The service exposes the /predict endpoint. It reads the model stored in the mounted drive and uses it to evaluate the request.

Setup liminal

cd incubator-liminal/examples/spark-app-demo/k8s

Liminal build

The build will create docker images based on the liminal.yml file in the images section.

liminal build

Liminal create

All tasks use a mounted volume as defined in the pipeline YAML.
In our case the mounted volume will point to the liminal Iris Classification example. The training task trains a regression model using a public dataset. We then validate the model and deploy it to a model-store in the mounted volume.

Create a kubernetes local volume:

liminal create

Liminal deploy

The deploy command deploys a liminal server and deploys any liminal.yml files in your working directory or any of its subdirectories to your liminal home directory.

liminal deploy --clean  

Note: liminal home directory is located in the path defined in LIMINAL_HOME env variable. If the LIMINAL_HOME environemnet variable is not defined, home directory defaults to ~/liminal_home directory.

Liminal start

The start command spins up 3 containers that load the Apache Airflow stack. Liminal's Airflow extension is responsible to execute the workflows defined in the liminal.yml file as standard Airflow DAGs.

liminal start

You can go to graph view to see all the tasks configured in the liminal.yml file: http://localhost:8080/admin/airflow/graph?dag_id=my_first_spark_pipeline

You should see the following dag:

Liminal YAML walkthrough

• Local archetype
• Pipeline flow

Local archetype

A superliminal for an easy local development

name: InfraSpark
owner: Bosco Albert Baracus
type: super
executors:
  - executor: k8s
    type: kubernetes
variables:
  output_root_dir: /mnt/gettingstartedvol
  input_root_dir: ''
images:
  - image: my_spark_image
    type: spark
    source: .
    no_cache: True
task_defaults:
  spark:
    executor: k8s
    executors: 2
    application_source: '{{application}}'
    mounts:
      - mount: mymount
        volume: gettingstartedvol
        path: /mnt/gettingstartedvol

We specify the MOUNT_PATH in which we store the trained model.

You can read more about superliminal variables and defaults in advanced.liminal.yml

Pipeline flow

Declaration of the pipeline tasks flow in your liminal YAML:

name: MyFirstLiminalSparkApp
super: InfraSpark
owner: Bosco Albert Baracus
variables:
  output_path: '{{output_root_dir}}/my_first_liminal_spark_app_outputs/'
  application: data_cleanup.py
task_defaults:
  python:
    mounts:
      - mount: mymount
        volume: gettingstartedvol
        path: /mnt/gettingstartedvol
pipelines:
  - pipeline: my_first_spark_pipeline
    start_date: 1970-01-01
    timeout_minutes: 45
    schedule: 0 * 1 * *
    tasks:
      - task: data_preprocessing
        type: spark
        description: prepare the data for training
        application_arguments:
          - '{{input_root_dir}}data/iris.csv'
          - '{{output_path}}'
      - task: train
        type: python
        description: train model
        image: myorg/mydatascienceapp
        cmd: python -u training.py train '{{output_path}}'
        env:
          MOUNT_PATH: /mnt/gettingstartedvol
        ...

Evaluate the iris classification model

Once the Iris Classification model trainging is completed and model is deployed (to the mounted volume), you can launch a pod of the pre-built image which contains a flask server, by applying the following Kubernetes manifest configuration:

kubectl apply -f manifests/spark-app-demo.yaml

Alternatively, create a Kubernetes pod from stdin:

cat <<EOF | kubectl apply -f -
---
apiVersion: v1
kind: Pod
metadata:
  name: spark-app-demo
spec:
  volumes:
    - name: task-pv-storage
      persistentVolumeClaim:
        claimName: gettingstartedvol-pvc
  containers:
    - name: task-pv-container
      imagePullPolicy: Never
      image: myorg/mydatascienceapp
      lifecycle:
        postStart:
          exec:
            command: [ "/bin/bash", "-c", "apt update && apt install curl -y" ]
      ports:
        - containerPort: 80
          name: "http-server"
      volumeMounts:
        - mountPath: "/mnt/gettingstartedvol"
          name: task-pv-storage
  EOF

Check that the service is running:

kubectl get pods --namespace=default

Check that the service is up:

kubectl exec -it --namespace=default spark-app-demo-- /bin/bash -c "curl localhost/healthcheck"

Check the prediction:

kubectl exec -it --namespace=default spark-app-demo-- /bin/bash -c "curl -X POST -d '{\"petal_width\": \"2.1\"}' localhost/predict"

Debugging Kubernetes Deployments

kubectl get pods will help you check your pod status:

kubectl get pods --namespace=default

kubectl logs will help you check your pods log:

kubectl logs --namespace=default spark-app-demo

kubectl exec to get a shell to a running container:

kubectl exec --namespace=default spark-app-demo -- bash

Then you can check the mounted volume df -h and to verify the result of the model.

You can go to graph view to see all the tasks configured in the liminal.yml file: [http://localhost:8080/admin/airflow/graph?dag_id=my_first_pipeline]( http://localhost:8080/admin/airflow/graph?dag_id=my_first_pipeline

Here are the entire list of commands, if you want to start from scratch:

git clone https://github.com/apache/incubator-liminal
cd examples/spark-app-demo/k8s
# cd examples/spark-app-demo/emr
python3 -m venv env
source env/bin/activate
rm -rf ~/liminal_home
pip uninstall apache-liminal
pip install apache-liminal
Liminal build
Liminal create
liminal deploy --clean
liminal start

Closing up

To make sure liminal containers are stopped use:

liminal stop

To deactivate the python virtual env use:

deactivate

To terminate the kubernetes pod:

kubectl delete pod --namespace=default spark-app-demo