My team recently began writing Spark jobs to be deployed to a Spark cluster
in the form of a jar. Previously, my team interacted with Spark via the
REPL. The job in question works within the REPL, but fails when executed
non-interactively (i.e. packaged as a jar).
The job code looks similar to:
// imports
object Runner extends App {
val end = new DateTime()
// additional setup
someRdd.filter(f => end.isAfter(f.date)
}
The point of this example is that a value, end, is defined local to the
driver. Later in the program's execution, the locally defined value, end,
is referenced in the filter predicate of an RDD. When running
non-interactively, a NPE occurs when 'end' is referenced in the filter
predicate. However, running the exact same code via the REPL executes
successfully.
Spark environment details are:
Spark version: v0.9 using commit SHA
e2ebc3a9d8bca83bf842b134f2f056c1af0ad2be
Scala version: v2.9.3
I appreciate any help in identifying bugs/mistakes made.
Thank you,
Michael
