Github user tgravescs commented on the issue: https://github.com/apache/spark/pull/21953 We have seen jobs overloading the first disk returned by Yarn. Unfortunately the details of the job have long expired. Its in general a good practice to distribute the load anyway. I remember one of the jobs was python. I think it was the case if you look in like EvalPythonExec.scala: // The queue used to buffer input rows so we can drain it to // combine input with output from Python. val queue = HybridRowQueue(context.taskMemoryManager(), new File(Utils.getLocalDir(SparkEnv.get.conf)), child.output.length) That is always going to hit the disk yarn returns first for every container on that node.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org