Lazy Maintenance of Materialized Views

Materialized views can speed up query processing greatly but they have to be kept up to date to be useful. Today, database systems typically maintain views eagerly in the same transaction as the base table updates. This has the effect that updates pay for view maintenance while beneficiaries (queries) get a free ride! View maintenance overhead can be significant and it seems unfair to have updates bear the cost. We present a novel way to lazily maintain materialized views that relieves updates of this overhead. Maintenance of a view is postponed until the system has free cycles or the view is referenced by a query. View maintenance is fully or partly hidden from queries depending on the system load. Ideally, views are maintained entirely on system time at no cost to updates and queries. The efficiency of lazy maintenance is improved by combining updates from several transactions into a single maintenance operation, by condensing multiple updates of the same row into a single update, and by exploiting row versioning. Experiments using a prototype implementation in Microsoft SQL Server show much faster response times for updates and also significant reduction in maintenance cost when combining updates.