Dynamically Allocate Spark Executors
By default, the Spark application will use a fixed number of executors. Setting
true enables Spark to add and remove executors
during execution depending on the number of Spark tasks scheduled to run. As
described in Dynamic Resource Allocation, dynamic allocation requires that
spark.shuffle.service.enabled be set to
On Nomad, this adds an additional shuffle service task to the executor task group. This results in a one-to-one mapping of executors to shuffle services.
When the executor exits, the shuffle service continues running so that it can serve any results produced by the executor. Due to the nature of resource allocation in Nomad, the resources allocated to the executor tasks are not freed until the shuffle service (and the application) has finished.
Learn how to integrate Spark with HDFS.