System and method for consistent timestamping in distributed computer databases

摘要

A distributed database system has a plurality of databases located at distinct nodes, at least one of the databases comprising a timestamping database. Distributed transactions are committed using a two phase protocol. During the first phase, each cohort to the transaction votes to commit or abort the transaction, and also votes an earliest time and a latest time at which the transaction is to be committed. If all the cohorts vote to commit the transaction and the intersection of the voted time ranges is not empty, then the transaction is committed during the second phase of the protocol. A transaction time is selected from the intersection of the voted time ranges and is used to timestamp all updated data that is durably stored when the transaction is committed. Before the first phase of the two phase commit protocol, each transaction read or write locks data at each node for which it needs read or write access. Whenever a transaction enters the first phase of the commit protocol, read locks for that transaction can be converted into delay locks. Any transaction which obtains a write lock on delay locked data is a "delayed transaction". The delayed transaction votes a time range which guarantees that it will commit at a time which is later than the time at which the transactions with the delay locks commit. This combination of time range voting and delay locking ensures that the timestamp order of transactions is consistent throughout the distributed database and is consistent with a valid serialization order of the transactions.