Run Apache Prediction IO using docker. Features include being able to select the template engine as long as it supports the default setup: (clone, build, rename engine.json, train, deploy). For custom actions please refer to Makefile and config/pio_app.sh

• Docker
• Docker-compose (if using DockerForMac this is already included)

We need to setup some environment variables first.

• Step 1. Copy .env.example to .env. In that file adjust the parameter values according to your needs.

• Step 2. Run make build to build the containers and the recommendation engine.

• Run make app to start the application.

• Run make stop to stop the application.

• Run make clean to stop the application and delete all data.

To submit data to the recommendation engine for your app make http request to http://localhost:7070/events.json?accessKey=M41jRfbEXGoHTxyLbn6jxpL2xnfoyxz_Psm0AMfSbzNBXOWphHO3Q1GfOUg3P6O5 where accessKey is the value of PIO_APP_KEY as defined in the .env file.

To query data and predictions make http requests to http://localhost:8000.

You can read more about what kind of data you can send and query on https://predictionio.apache.org/templates/similarproduct/quickstart/ specifically Step 4 and on. This repository has already implemented the previous steps.

• If you cannot access the query service at http://localhost:8000 this probably because you have not sent any events to the events service. To be able to train a model and query for predictions you must first record some events.

• You need to send your own data, but if you want to play around with some sample data run make sampledata. This will run the script ./data/importdata.sh which will do a couple of HTTP requests to feed the engine.

• The first build of a template engine is very slow. For reference building the Similar-Product engine for the first time takes ~30minutes on a late 2017 macbook.

After you have inputted data, it might take about 1 minute for the http://localhost:8000 page to be accessible.

The following services are defined:

• pio_db: A Postgres database that will store the events and the model
• elasticsearch: An elasticsearch database for the model metadata
• pio_events: The events service that records all events. Access it on http://localhost:7070
• pio_app: The recommendation service that hosts the prediction engine and allows for quering predictions on port http://localhost:8000.

You can read docker-compose.yml for more details.

This setup has been tested with the Similar-Product engine. You can find example of events you can send and more info in the official documentation: https://predictionio.apache.org/templates/similarproduct/quickstart/

More engines that use the same installation steps are supported. You can do that changing the .env file:

1. Setting PIO_ENGINE to the url of the template engine. You can find template engines on https://predictionio.apache.org/gallery/template-gallery/

2. Run make engine to build the engine. Note that you need to clean your data and installation if you already had installed a different engine before. Use make clean to do that.

1. Dockerfile: I used multistage builds for docker to take advantage of caching layers and to prevent rebuilding the whole image when updating version of a single vendor dependency.

2. Template engine: The template engine is been downloaded and built after the containers have started and not in the Dockerfile build process. This is to allow for different template engines to be used with the same platform, configurable from the environment variables.

3. Persistent storage: In order to minimize the time spent building the recommendation engine (it really took >30minutes on my machine) named volumes are used to persist the home directory. This speeds up considerably subsequent engine builds (now take ~5 seconds).