Optimizing Certification-Based Database Recovery

Certification-based database replication protocols are a good basis to develop replica recovery when they provide the snapshot isolation level. For such isolation level, no readset needs to be transferred between replicas nor checked in the certification phase. Additionally, these protocols need to maintain a historic list of writesets that is used for certifying the transactions that arrive to the commit phase. Such historic list can be used to transfer the missed state of a recovering replica. We study the performance of the basic recovery approach - to transfer all missed writesets - and a version-based optimization - to transfer the latest version of each missed item, compacting thus the writeset list - and the results show that such optimization reduces a lot the recovery time.