Distributed multi-version commitment ordering protocols for guaranteeing serializability during transaction processing

摘要

In a multi-version database, copies of prior committed versions (snapshots) are kept for access by the read-only transactions. The read- write transactions are selectively aborted to enforce an order of commitment of read-write transactions that is the same as an order of conflicts among the read-write transactions. In a preferred embodiment, the read-write transactions are serialized by maintaining and referencing a graph of conflicts among read-write transactions, and the read-only transactions are serialized by a timestamp mechanism for selection of the snapshots to be read. Each time that a read-write transaction is committed, the read-write transaction is assigned a unique timestamp that is used to timestamp all resources committed by the read-write transaction. Upon starting, each read-only transaction is also assigned a timestamp. Each read-only transaction reads only the latest committed versions of all resources, that are timestamped earlier than the timestamp of the read-only transaction. In a multiprocessing system, the timestamps are issued to global coordinators and distributed locally with atomic commit messages and global queries. Moreover, read-write transactions may selectively access a hierarchy of uncommitted versions to prepare for various possible commitment orders. The hierarchy defines a path for record access and for cascading aborts. A plurality of mutually-conflicting uncommitted versions may be prepared for each transaction to prepare for all possible commitment orders.