Over the last year there's been a significant amount of low level changes in the proprietary versions of spark (aka EMR and Databricks) designed to address reliability and stability. Out of curiosity what exceptions did you run into?