In many ways, traditional undo segments are treated like table and index segments. All the classic operations must be performed. For example, changes to a segment must be recorded in the redo stream for recovery and the LRU chains and the write (dirty) lists must be updated to ensure smooth and proper data management. And don't forget about all the associated latching and pinning that must occur, not to mention all the associated CPU cycles to make this happen. Do these operations really need to be done using classic segments? Not according to Oracle.

Clearly, there exists an opportunity for increased performance. But as I mentioned, it's risky and time-consuming to develop, test, and implement a radical change. It requires everyone to change the way they think about Oracle undo management. But once you cross this threshold, you'll begin to see amazing performance improvement possibilities.

With IMU, Oracle has shifted as much undo work as possible to using optimized in-memory structures instead of traditional segments. Oracle still creates undo, because it must still provide rollback and read-consistency capabilities. This has many implications.