A partial rollback of a transaction restores the state of the transaction and the database to the state in which they have been right after the savepoint to be rolled back has been established. The paper considers the correctness of savepoints and partial rollbacks. First, a savepoint sequence number is assigned to each savepoint. According to that, without generating any savepoint log record upon establishing savepoints, all the valid savepoints of a transaction can have different savepoint sequence numbers and the order of establishment of each valid savepoint can be discriminated efficiently. Second, the lock request entries of the manual duration are not deleted from the lock table as far as some savepoints which are established after the locks are requested but before the locks are released are valid. That guarantees the correctness of savepoints and partial rollbacks under the circumstances of supporting the four transaction isolation levels of SQL2. Third, the concept of the mini-savepoint that preserves the atomicity of update operations of the SQL level is extended and adopted to triggers and stored procedures. According to that, the atomicity of triggers and stored procedures is supported and the scope of savepoint names is restricted to the internal of the corresponding triggers or stored procedures.

We present a continuous deadlock detection and resolution algorithm in distributed database systems.Our algorithm maintains an augmented transaction wait-for graph at each site and uses a modified priority-based probe generation scheme in order to detect local deadlocks without transmitting any intra-site deadlock detection messages, to minimize the number of inter-site messages sent for detection of global deadlocks and also for the early detection of global deadlocks that might occur in the future without transmitting detection messages repeatedly.The augmented transaction wait-for graph contains, in addition to lock-wait information, information about message-wait relationships among agents of a transaction, probes received from other sites and transitive wait-for relationships among transactions.Global deadlocks are declared whenever a transitive wait-for relationship from an agent of a global transaction is constructed for some agent of the transaction.

The authors address the deadlock problem in sequential transaction processing where the strict two-phase locking and the multiple granularity locking protocol with five lock modes are used. The scheduling policy honors lock requests in a first-in-first-out basis except for lock conversions. As a basic tool, a direct graph model called the holder/wire-transaction waited-by graph (H/W-TWBG) is introduced to capture the precise status of systems in terms of deadlock. The properties of H/W-TWBG are presented. Based on H/W-TWBG, the identification principles of the victim candidates are established in a deadlock cycle, and a periodic deadlock detection and resolution algorithm which has a reasonable time and storage complexity is preserved. One important feature of the deadlock resolution scheme is that some deadlocks can be resolved without aborting any transaction.>

An algorithm for deadlock detection and resolution in sequential transaction processing is presented. Two-phase locking is assumed for ensuring serializability, the lock requests obey the granularity locking protocol, and each granule may be locked in one of the following lock modes: IS, IX, S, SIX and X. For each object, lock requests are honored according to a first-come-first-served basis except for lock conversions. The basic idea for the deadlock detection resolution is in the construction of a new direct graph called a holder/waiter-transaction waited-by graph. The authors establish guidelines for the identification of a victim in a deadlock cycle and propose a resolution algorithm whose time and space requirements are reasonable and whose solution is near optimal.>