What is data spill in Spark and how to avoid it if resources cannot be simply added?