A novel recovery mechanism enabling fine-granularity locking and fast, REDO-only recovery

We present a series of novel techniques and algorithms for transaction commit, logging, recovery, and propagation control. In combination, they provide a recovery component that maintains the persistent state of the database (both log and data pages) always in a committed state. Recovery from system and media failures only requires only REDO operations, which can happen concurrently with the processing of new transactions. The mechanism supports fine-granularity locking, partial rollbacks, and snapshot isolation for reader transactions. Our design does not assume a specific hardware configuration such as non-volatile RAM or flash---it is designed for traditional disk environments. Nevertheless, it can exploit modern I/O devices for higher transaction throughput and reduced recovery time with a high degree of flexibility.