Management of a remote backup copy for disaster recovery

Reviewer: Peter John Trueman

Two multicomputer systems are linked by multiple communication paths, and each has a partitioned relational database. The primary system processes the database transactions, and the other acts as a hot-standby that will take over the processing in the event of the primary system failing. This paper describes the decentralized algorithm used to ensure that the backup system's database is up to date and consistent. The algorithm is intended for use in applications that have a lot of transactions and need a quick response time. So, rather than use an expensive two-phase commit protocol to ensure that a transaction atomically updates both systems, the transaction simply commits and then propagates to the backup. This means that when the primary fails, some committed transactions may be lost and others may have to be discarded to maintain consistency; for example, if the transaction that created a bank account is lost, updating that account would be wrong. This risk is deemed to be economically acceptable given the performance requirement; if the consequences of a lost transaction are high, however, it is possible to use an atomic commit over both systems. After some good background and introductory sections, the paper defines what it means for the backup system to be consistent, and then describes how the backup system is initialized and updated, and argues that this process results in a consistent backup. Although the paper is lengthy, it is not too long; it is well structured and clear. The reader who wants just as much information as can be easil y remembered can read the first ten pages, leaving the rest for the serious student. The paper is well worth reading.