Cache consistency in adaptive page-server database systems

We present cache consistency (adaptive concurrency-control and adaptive replica management) protocols for page-server database system in which the database is indexed by B±tree. These protocols avoid the data contention that may occur when concurrency control and replica management are performed at the page level. Page-level X locks are needed only for splitting or merging database (index or data)) pages, while record inserts and deletes on a data (leaf) page need only a short-duration U lock on the affected page (besides the record-level X locks). Thus, a record in a U-locked data page can be updated by one client transaction while other records can simultaneously be fetched by transactions that cache a copy of the page at other clients. Page-level S locks on data pages are not held for commit duration, but are changed to a set of commit-duration record-level S locks at the server if the page needs to be updated at another client and is therefore called back.