Maybe load the model on each executor’s disk and load it from there? Depending on how you use the data/model, using something like Livy and sharing the same connection may help?
From: Naveen Swamy [mailto:firstname.lastname@example.org]
Sent: Wednesday, September 27, 2017 9:08 PM
Subject: Loading objects only once
I am a new user to Spark, please bear with me if this has been discussed earlier.
I am trying to run batch inference using DL frameworks pre-trained models and Spark. Basically, I want to download a model(which is usually ~500 MB) onto the workers and load the model and run inference on images fetched from the source like S3 something like this
rdd = sc.parallelize(load_from_s3)
I was able to get it running in local mode on Jupyter, However, I would like to load the model only once and not every map operation. A setup hook would have nice which loads the model once into the JVM, I came across this JIRA https://issues.apache.org/jira/browse/SPARK-650 which suggests that I can use Singleton and static initialization. I tried to do this using a Singleton metaclass following the thread here https://stackoverflow.com/questions/6760685/creating-a-singleton-in-python. Following this failed miserably complaining that Spark cannot serialize ctype objects with pointer references.
After a lot of trial and error, I moved the code to a separate file by creating a static method for predict that checks if a class variable is set or not and loads the model if not set. This approach does not sound thread safe to me, So I wanted to reach out and see if there are established patterns on how to achieve something like this.
Also, I would like to understand the executor->tasks->python process mapping, Does each task gets mapped to a separate python process? The reason I ask is I want to be to use mapPartition method to load a batch of files and run inference on them separately for which I need to load the object once per task. Any
Thanks for your time in answering my question.