Random excessive UNDO reads long after DML committed (Exadata) -- need debug approach

Hello, having this problem again at a couple of customers, and it's getting worse. Looking for ways to further trace the issue and/or know what parameters might need to be changed.

ETL happens 1-3 A.M. All table changes committed by 3 A.M. Downstream query extracts start after 3:30 A.M. Extract queries will run fine MOST days, but even with same execution plan, most of the activity is in the UNDO tablespace, making the query run 10x longer (or more). Confirmed that there is no other contention on the tables; COMMITs on those tables happened at least an hour earlier.