Optimal termination protocols for network partitioning

Abstract

Commit protocols guarantee the consistency of distributed databases in absence of any failures. A commit protocol is resilient to a class of failures if it is possible to guarantee that a) databases at all operational sites in presence of these failures are consistent and b) other sites can be recovered consistently with these sites when the failure is repaired. Such a commit protocol is called nonblocking if no operational site needs to wait on a transaction which is incomplete at the time of the failure. It is known that no nonblocking commit protocol resilient to network partitioning exists. In this paper, the possible termination protocols of commit protocols are studied in the context of network partitioning. A formal model for termination protocols is introduced and a general logical interpretation of termination protocols is presented. The model makes use of all the information that is available in a component of the partition --- namely, the constituent sites and their respective states at the time of partition. Optimality measures for the termination protocols in terms of the number of waiting components and average number of waiting sites are introduced and protocols optimal under these measures are produced for all the possible centralized and decentralized commit protocols. It is proved that quorum-based termination protocols indeed perform very well in the presence of network partitioning. If the central site(s) is reliable, we can prove that centralized commit protocols indeed perform better than all decentralized ones. Thus, the general preference for centralized commit protocols is justified.