Let's operationalise a Spark regression model

Lately I've been doing a lot of work with Azure DataBricks which is a superb Spark-based analytics platform fully integrated with Azure. It gives you everything that Open Source Spark does and then some. I've been especially enjoying the effortless ways to move large datasets around and the ease of MLlib for my AI-projects.

One of the questions with the simpler models like regressions and clusterings is always how to operationalise the models so that the rest of the organisation could consume the models by calling webservices. Luckily there exists model export functionality that takes us almost there. It exports Spark MLlib models so that they can be run anywhere where there is a JVM available ... no Spark needed.

So I decided to build the missing half : wrap the model inside a simple java web app and deploy to Azure to be consumed.

Some theory

First read this article to properly understand the model export feature:https://docs.databricks.com/spark/latest/mllib/model-export.html

Then you might want to take a look at some sample code on instancing the exported models and how to use them for predicting stuff:
https://github.com/databricks/databricks-ml-examples/tree/master/model-export-demo

Finally you should download my sample IntelliJ project and use it as base for you own operationalising :
https://petsablob.blob.core.windows.net/share/siirto/modeltestwebapp.zip

Important bits

The solution itself is ridiculously simple and I'll go throug it here for your pleasure.
I set up a basic maven archetype web app project with IntelliJ Idea and added one utility class and one library (read the docs).

This is my index.jsp-file that does the interfacing with the caller.

Here we basically set up the parameter block structure that goes into the model and take values from the web request that was sent to us. The request could be something like this: https://YOURSERVER.azurewebsites.net/?topic=sci.space&text=space%20the%20final%20frontier
Then the values are handed over to our utility class to be reacted upon and the result is returned to the caller ... not too complex , eh ?.

The utility class is even simpler ...
It knows how azure treats java apps and how you can refer to files contained in the war-package (our model files, that is).
Knowing all this it creates a static instance of the model and starts churning out predictions when asked to do so.

I then used Microsofts excellent IntelliJ Azure plugin to deploy this project to ... you guessed it ... to Azure.

Summary

The funny thing is that done this way your service is automagically enjoying the benefits of Azure Web App Autoscaling and Azure Authentication (Easy Auth, and yes, it works with Java too) making this a real production ready enterprise solution ... with embarrassingly few lines of code .