CrossStitch: An efficient transaction processing framework for geo-distributed system

Abstract

Current transaction system for geo-distributed datastores either have high transaction processing latencies or are unable to support general transactions with dependent operations. In this paper, we introduce CrossStitch, an efficient transaction processing framework that reduces latency by restructuring each transaction into a chain of state transitions, where each state consists of a key operation and computation.

Transaction states are processed sequentially, and the transaction code and data is sent directly to the next hop in the chain. CrossStitch transactions can be organized such that all states in a location are processed before transitioning to a state in a different location. This allows CrossStitch to significantly reduce the number of inter-location crossings compared to transaction systems that retrieve remote data to a single location for processing.

To provide transactional properties while preserving the chain communication pattern, CrossStitch introduces a pipelined commit protocol that executes in parallel with the transaction and does not require any centralized co-ordination. Our evaluation results show that CrossStitch can reduce the latency of geo-distributed transactions when compared to a traditional 2PC-based distributed transaction system. We demonstrate that CrossStitch can reduce the number of round trips by more than half for TPC-C-like transactions.