* To allow a transaction to roll back so the database looks just as it did before the transaction started.

* By default, whenever a query starts and ends, the results will match what the database looked like when the query started.

Segment management is resource-intensive. Needing to manage block-based segments in memory is clearly not optimal. It makes more sense to do all the in-memory work using highly optimized structures designed for in-memory operations, and then, when necessary, transform the in-memory structures into traditional segment structures. As an example, buffer headers are a highly optimized in-memory structures directly related to cached blocks.