I have written a Spark application that uses a library that uses the Google protobuf library in version 3.3.1. On my computer I can run it with my local Spark and everything is fine. But now, I want to run it on an EMR cluster in on AWS. And I get this:
It seems that this is a common incopatibility issue (says stackoverflow). Java libraries are loaded from the classpath. The classpath is searched from front to back. And in the EMR cluster, Spark and Hadoop libraries are inserted before the user libraries. So the old protobuf version is found.
But there is the possibility to change this behaviour described in the Spark 2.4.4 documentation. So I added two corresponding options to
spark-submit in the EMR cluster step:
--conf spark.driver.userClassPathFirst=true --conf spark.executor.userClassPathFirst=true
And it worked!
NB: The proposed solution at stackoverflow is to create a “shaded” fat jar. That probably works as well and should be simple for Maven users. Maybe my next post is about that 😉