Near-Realtime Prediction with Flink

This project stands as an end-to-end example of building a machine learning model (linear regression), exporting that model with jpmml and using that model for real time prediction with Flink streaming.

Pre-Requisites

1. Docker

   • Mac

2. Gradle - You have a few options here

   • If you're using Intellij, just make sure it's enabled.
   • Run brew install gradle

3. Kaggle Api Token (Not absolutely necessary but recommended)

   • See https://www.kaggle.com/docs/api#authentication

Up & Running

First let's clone the repo and fire up our system,

git clone [email protected]:aedenj/flink-machine-learning-fish-market-example.git ~/projects/flink-ml-example
cd ~/projects/flink-ml-example;

The Model

The linear regression model in the Jupyter notebook is built using the Fish Market dataset on Kaggle If you'd like to step through the notebook yourself run,

cd ~/projects/flink-ml-example;
docker run --rm -p 8888:8888 -e JUPYTER_ENABLE_LAB=yes -e GRANT_SUDO=yes --user root -v ~/.kaggle:/home/jovyan/.kaggle -v "$PWD":/home/jovyan/work jupyter/pyspark-notebook

After the docker container is up and running you'll see output like this in the terminal

    To access the server, open this file in a browser:
        file:///home/jovyan/.local/share/jupyter/runtime/jpserver-16-open.html
    Or copy and paste one of these URLs:
        http://e2de1e96eae2:8888/lab?token=b9478a108f3b7b86b45f7131f10cb18a17bc337dbd45233d
        http://127.0.0.1:8888/lab?token=b9478a108f3b7b86b45f7131f10cb18a17bc337dbd45233d

Navigate to one of the urls listed in your terminal then open the file model/model.ipynb.

I find it helpful to have an alias for running Jupyter under docker. One possibility on unix systems is,

alias jupyterd='f(){ docker run --rm -p $1:8888 -e JUPYTER_ENABLE_LAB=yes -e GRANT_SUDO=yes --user root -v ~/.kaggle:/home/jovyan/.kaggle -v "$PWD":/home/jovyan/work jupyter/pyspark-notebook; unset -f f; }; f'

The Job

The job reads from a topic of observations named fishes and writes to a topic named weight-predictions. The json messages in the fishes topic contain the predictors/fields,

• Length - Representing the cross length in cm
• Species

In order to see the job in action run,

1. ./gradlew kafkaup
2. ./gradlew createtopics
3. ./gradlew shadowJar run
4. In a new terminal start a Kafka producer by running ./scripts/start-kafka-producer.sh
5. You'll see the prompt >. Enter the message 1:{length: 41.3, species:"Perch"}
6. Navigate to the Kafka Topics UI and inspect both the fishes and weight-predictions topics.

You should see the message { weight: 804.8438505999843, length: 41.3, species: Perch } in predictions topic.