I'm using Typesafe Config, https://github.com/typesafehub/config, to parameterize a Spark job running in yarn-cluster mode with a configuration file. The default behavior of Typesafe Config is to search the classpath for resources with names matching a regex and to load them into your configuration class automatically with ConfigFactory.load()
(for our purposes, assume the file it looks for is called application.conf
).
I am able to load the configuration file into the driver using --driver-class-path <directory containing configuration file>
, but using --conf spark.executor.extraClassPath=<directory containing configuration file>
does not put the resource on the classpath of all executors like it should. The executors report that they can not find a certain configuration setting for a key that does exist in the configuration file that I'm attempting to add to their classpaths.
What is the correct way to add a file to the classpaths of all executor JVMs using Spark?
It looks like the value of the
spark.executor.extraClassPath
property is relative to the working directory of the application ON THE EXECUTOR.So, to use this property correctly, one should use
--files <configuration file>
to first direct Spark to copy the file to the working directory of all executors, then usespark.executor.extraClassPath=./
to add the executor's working directory to its classpath. This combination results in the executor being able to read values from the configuration file.I use the SparkContext addFile method