For decades, the strongest guarantee that database systems users could expect from their system was "serializability." This guarantee ensured that even though the system would process many transactions concurrently, the final state of the database system would be equivalent to as if it had processed all transactions serially, one after another. However, the database system would make no guarantee about which serial order that processed transactions would be equivalent to—arbitrary nondeterministic events such as operating system thread scheduling, deadlock, or server failure could change this equivalent serial order.
Research from the past decade has discovered a number of advantages to architecting database systems with a stronger set of guarantees. Instead of promising equivalence to any arbitrary serial order, the system instead guarantees equivalence to processing transactions in a single predetermined serial order. Furthermore, there is only one possible state the system may end up in, despite the presence of potential non-deterministic code in transaction logic. Research has shown many benefits from this stronger set of guarantees, from simpler and higher performance database replication, to improved system scalability, to removing distributed commit protocols. Furthermore, while the early work assumed the increased set of guarantees would decrease the ability of the system to process transactions concurrently, more recent research has shown the total opposite result: transaction concurrency has increased. This article describes these benefits of deterministic database systems in more detail, along with a discussion of the primary disadvantages: the lack of support of interactive transactions in the system, and the need for transaction pre-processing prior to execution.
No entries found