Abstract: A method and system for recovering after a crash of an instance in a database is disclosed. According to the method, transaction information corresponding to all the transactions that existed in the failed instance is inspected. These transactions that were active at the time the instance crashed are marked as "dead." The database is made available to users after updating the transaction information and before undoing all of the changes or updates performed by dead transactions. The changes made by the dead transactions may be undone by new transactions on a on-demand basis after the database is made available. Specifically, when a new transaction is waiting to obtain a lock on a resource held by another transaction, the new transaction determines whether the owner of the lock is one of the dead transactions. If so, the new transaction rolls back some or all of the changes made by the dead transaction.

Abstract: A computer system and method are provided in which an existing database instance is able to report the status of an externally initiated distributed transaction that existed in another database instance that has crashed. The existing database instance is also able to roll back or commit the changes made by the externally initiated distributed transaction in response to instructions from the external coordinator process that initiated the transaction. Because the existing database instance is able to perform these services, the external coordinator process does not have to wait for the crashed database instance to be revived to complete the distributed transaction.