Almost Triggerless Writeset Extraction in Multiversioned Databases

Efficient writeset extraction and application is crucial in modern database replication protocols since they execute all transaction operations in a delegate replica, collecting and propagating the updates to all replicas when transactions request commitment. In order to ensure portability, such replication protocols can be implemented in a middleware and some of such middleware systems use triggers for managing writeset collection. MADIS is an example of such systems. It has been developed assuming PostgreSQL as its underlying DBMS. Thus, its protocols are usually oriented to enforce snapshot isolation and rely on multiversioned concurrency control. PostgreSQL provides in its catalog information regarding its multiversioned control and such information can be used by the replication middleware for improving its writeset collection and application performance, without requiring any knowledge about DBMS´s internals nor compromising the independence between the middleware and the DBMS layer. This paper describes an almost triggerless fully-user-space approach for writeset extraction in the MADIS middleware database replication system, comparing its performance with a minimal implementation based on triggers.