Abstract: A method and an apparatus for tracking of the dependencies between transactions is provided. Every time a data item is updated, a record is made of the transaction that updated the data item. Before another transaction locks a data item previously locked by the transaction, the entry is updated to indicate that the transaction committed and the commit time of the transaction. These entries are contained in a list head that is maintained on the same block as the data item, and a list tail that is stored separate from the data block that contains the data item. A depends-on time is maintained for each transaction. Whenever the transaction updates a data item, the depends-on time is set to the greater of the current depends-on time and the commit time of the most recently committed transaction that updated the version of the data item.

Abstract: The present invention provides a distributed log system where the logs are distributed on a per client basis. That is, the clients are partitioned into subsets. Each subset of clients is assigned a log, and the the redo records of those clients are maintained in that log. This reduces contention bottlenecks, because the number of clients writing to any one log is limited. Also, all of a client's changes are stored in a single log. During a recovery process, the changes for a block of data may be stored in different logs. For proper recovery, the changes must be applied to the database in the order in which they were originally made. The present invention provides a system for ordering the log entries in the correct order by using a "sequence number". Each page of data in the database has an associated sequence number. When a change is made to that page, the sequence number is incremented. The new sequence number is written to the log along with the change information.