Fast Crash Recovery for a Distributed Column-Store Database Management System

A traditional database management system (DMBS) supports both atomic transactions and crash recovery by maintaining an undo/redo log. The standard write-ahead logging protocol, which requires a system to force the undo and redo log records describing a modified page to stable storage before writing the modified page to disk, ensures that sufficient information exists to perform recovery. After a computer failure, a rebooted system can process the log with a crash recovery algorithm, such as ARIES [1], to redo committed transactions and undo uncommitted ones, thereby restoring the system to a consistent state.