Decoupling the web app from Spark backend is recommended. Training the model can be launched in the background via a scheduling tool. Inferring the model with Spark in interactive mode s not a good option as it will do it for unitary data and Spark is better in using large dataset. The original purpose of inferring with Spark was to do it offline for large datasets and store the results in a KV store for instance, then any consumer like your web app would just read the KV store. I would personally store the trained model in PFA or PMML and serve it via another tool.
There are lots of tools to serve the models via API from managed solution like Amazon Sagemaker to open source solution like 
If you still want to call Spark backend from your web app, what I don't recommend, I would do it using Spark Jobserver or Livy to interact via rest API.

Le jeu. 4 oct. 2018 à 08:25, Jörn Franke <> a écrit :
Depending on your model size you can store it as PFA or PMML and run the prediction in Java. For larger models you will need a custom solution , potentially using a spark thrift Server/spark job server/Livy and a cache to store predictions that have been already calculated (eg based on previous requests to predict). Then you run also into thoughts on caching prediction results on the model version that has been used, evicting non-relevant predictions etc
Making the model available as a service is currently a topic where a lot of custom „plumbing“ is required , especially if models are a little bit larger.

Am 04.10.2018 um 06:55 schrieb Girish Vasmatkar <>:

On Mon, Oct 1, 2018 at 12:18 PM Girish Vasmatkar <> wrote:
Hi All

We are very early into our Spark days so the following may sound like a novice question :) I will try to keep this as short as possible.

We are trying to use Spark to introduce a recommendation engine that can be used to provide product recommendations and need help on some design decisions before moving forward. Ours is a web application running on Tomcat. So far, I have created a simple POC (standalone java program) that reads in a CSV file and feeds to FPGrowth and then fits the data and runs transformations. I would like to be able to do the following -

  • Scheduler runs nightly in Tomcat (which it does currently) and reads everything from the DB to train/fit the system. This can grow into really some large data and everyday we will have new data. Should I just use SparkContext here, within my scheduler, to FIT the system? Is this correct way to go about this? I am also planning to save the model on S3 which should be okay. We also thought on using HDFS. The scheduler's job will be just to create model and save the same and be done with it.
  • On the product page, we can then use the saved model to display the product recommendations for a particular product.
  • My understanding is that I should be able to use SparkContext here in my web application to just load the saved model and use it to derive the recommendations. Is this a good design? The problem I see using this approach is that the SparkContext does take time to initialize and this may cost dearly. Or should we keep SparkContext per web application to use a single instance of the same? We can initialize a SparkContext during application context initializaion phase. 

Since I am fairly new to using Spark properly, please help me take decision on whether the way I plan to use Spark is the recommended way? I have also seen use cases involving kafka tha does communication with Spark, but can we not do it directly using Spark Context? I am sure a lot of my understanding is wrong, so please feel free to correct me.

Thanks and Regards,
Girish Vasmatkar
HotWax Systems