Main memory databases instant recovery

Abstract

The recovery process in main memory database systems (MMDBs) run in an offline way. Thus, MMDB only becomes available for new transactions after the complete recovery process has finished. Some MMDBs maintain database replicas for assuring high availability after systems failure. Nonetheless, a database replication mechanism is not immune to failures as well. For that reason, recovery techniques are required to repair failed systems as quickly as possible. This work proposes an instant recovery strategy for MMDBs, which makes MMDBs able to process transactions immediately after the recovery engine is triggered. The proposed approach rebuilds the database incrementally and on-demand. Besides, a novel checkpoint technique is proposed to interfere as little as possible in the system performance. The checkpoint technique can also act during the recovery process so that the next recoveries are faster in the face of successive failures. In order to validate the approach, simulations with a prototype implemented on Redis have been conducted over Memtier benchmark. Preliminary results evidence the suitability of the proposed recovery mechanism.