Principles of transaction-oriented database recovery

Reviewer: Vijay Varadachari Raghavan

The paper aims to remedy the problems stemming from a lack of adequate and precise terminology for describing and coming to grips with the variety of concepts and algorithms used in the area of database recovery. To achieve this, a systematic framework for establishing and evaluating the basic concepts for fault-tolerant database operation is formulated. The concept central to their framework, that of a transaction, is stated and is used to describe the types of failures to be anticipated and the kinds of recovery actions that would, as a result, be required. Thus, this survey, unlike that of Verhofstadt [1], assumes a sppecific notion of consistency, which derives from treating the transaction as the only unit of recovery. The main contribution is a taxonomy with which arbitrary implementations for database recovery can be classified. It is based on four criteria: (1)whether a set of modified pages are materialized as a unit (ATOMIC) or blocks are updated in-place, leaving open the possibility of the write of a set of pages to be interrupted by system crashes ( :9I ATOMIC), (2)whether the time of replacement of modified pages are solely determined by the buffer manager (STEAL) or modified pages are kept in buffer at least until the end of transaction ( :9I STEAL), (3)whether or not all modified pages are materialized during end-of-transaction processing (FORCE) and (4)the kind of checkpoint scheme employed. The first criterion defines the state in which a database is left when the system terminates abnormally. STEAL :9I STEAL dichotomy has important implications for UNDO recovery. The othr criterion converning buffer handling (FORCE :9I FORCE) is related to the necessity of partial REDO. Depending on the choices for the above, the type of log data and the database state that is logged would vary. Finally, checkpointing helps to constrain partial REDO to different extents depending on the criteria for determining when to start checkpointing activities. Some existing DBMs and implementing concepts are classified and described using a classification tree based on these options. Important conclusions concerning the behavior and performance of these recovery techniques are made.