Enforcing Database Recoverability on Disks that Lack Write-Through

  • Robin Dhamankar ,
  • Hanuma Kodavalla ,
  • Vishal Kathuria

MSR-TR-2008-36 |

MSR Sponsor: SurajitC

Most database systems use ARIES-like logging and recovery scheme to recover from failures and guarantee transactional consistency. ARIES relies on Write-Ahead Logging (WAL) protocol which requires that log records be durably written prior to the corresponding data changes. Database systems use the write-through capability of the storage media to enforce write-ahead logging. While SCSI disks that are commonly deployed in enterprise servers support write-through, commodity hard drives do not. While database systems were mostly limited to enterprise servers in the past, today they are being heavily deployed in large-scale internet services and personal information management systems. In order to minimize costs, these systems use commodity hard drives that have controller caches but lack write-through. These drives delay and reorder the writes thereby breaking the WAL assumption on which recovery is based. Here we present a solution to enforce WAL, hence recoverability of the database on these drives. We also present performance measurements demonstrating that our approach does not cause any overhead in the path of the transaction.