Having your cake and eating it too: combining strong and eventual consistency

Given the limitations imposed on distributed systems that are necessary to maintain strong consistency guarantees there is a growing interest in relaxed consistency models. Such models are often sufficient for particular applications, but allow more freedom to improve scalability and availability. Eventual consistency is a particularly useful approach, where the correct state spreads throughout the system over time, so that at any point any element of the system may be inconsistent, but all elements will eventually converge upon a consistent state. On the other hand relaxing properties may be unacceptable in the general case: a slightly stale shopping cart is one thing, but inconsistent payment processing is quite another.

In this paper we try to balance strong and eventual consistency by proposing a general-purpose pessimistic distributed transactional memory that allows eventually consistent transactions to run alongside consistent ones. While the former maintain read-isolation (i.e., read from a consistent snapshot), they do not interfere with the latter's safety properties. The relaxed-consistency transactions are later followed by their consistent counterpart so that the user view and global state eventually agree. Our contribution is to show that we can significantly relax synchronization (to the point of eliminating it completely from eventually consistent transactions) while retaining useful properties, but without imposing additional constraints about system architecture or data operations, common to other relaxed consistency approaches. All this, without affecting those transactions that execute in consistent mode.