I can confirm this behaviour. To run the same job on shared cluster in "USER_ISOLATION" mode with nothing changes between the job definition or source data, the performance drop is significant. So much so that there needs to be a radical change in how we need to process data.