Logical recovery from single-page failures

Abstract

Modern hardware technologies and ever-increasing data sizes increase probability and frequency of local storage failures, e.g., unrecoverable read errors on individual disk sectors or pages on flash storage. Our prior work has formalized singlepage failures and outlined efficient methods for their detection and recovery. These prior techniques rely on old backup copies of individual pages, e.g., as part of a database backup or as old versions retained after a page migration. Those might not be available, however, e.g., after recent index creation in “non-logged” or “allocation-only logging” mode, which industrial database products commonly use. The present paper introduces techniques for single-page recovery without backup copies, e.g., pages of new indexes created in allocation-only logging mode. By rederiving lost contents of individual pages, these techniques enable efficient recovery of data lost due to damaged storage structures or storage devices. Recovery performance depends on the size of the failure and of the required data sources; it is independent of the sizes of device, index structure, etc.