When I tried to execute Beam pipeline over Spark structured streaming runner on YARN it failed. I did identify to problems:
When FilesToStage kicks in it tries to zip classpath directories and uses pipeline's temp directory. In my case it is HDFS and since PipelineResources.prepareFilesForStaging() uses Java IO it will not work as it does not speak HDFS
Since worker are provided with necessary classes beforehand, it would make sense to me to be able to bypass FilesToStage instead to always had to stage something even when not necessary.
Issue Priority
Priority: 2 (default / most normal work should be filed as P2)
What needs to happen?
When I tried to execute Beam pipeline over Spark structured streaming runner on YARN it failed. I did identify to problems:
PipelineResources.prepareFilesForStaging()
uses Java IO it will not work as it does not speak HDFSIssue Priority
Priority: 2 (default / most normal work should be filed as P2)
Issue Components