Almost Strong Consistency: "Good Enough" in Distributed Storage Systems

Abstract: A consistency/latency tradeoff arises as soon as a distributed storage system replicates data. For low latency, modern storage systems often settle for weak consistency conditions, which provide little, or even worse, no guarantee for data consistency. In this paper we propose the notion of almost strong consistency as a better balance option for the consistency/latency tradeoff. It provides both deterministically bounded staleness of data versions for each read and probabilistic quantification on the rate of "reading stale values", while achieving low latency. In the context of distributed storage systems, we investigate almost strong consistency in terms of 2-atomicity. Our 2AM (2-Atomicity Maintenance) algorithm completes both reads and writes in one communication round-trip, and guarantees that each read obtains the value of within the latest 2 versions. To quantify the rate of "reading stale values", we decompose the so-called "old-new inversion" phenomenon into concurrency patterns and read-write patterns, and propose a stochastic queueing model and a "timed balls-into-bins model" to analyze them, respectively. The theoretical analysis not only demonstrates that "old-new inversions" rarely occur as expected, but also reveals that the read-write pattern dominates in guaranteeing such rare data inconsistencies. These are further confirmed by the experimental results, showing that 2-atomicity is "good enough" in distributed storage systems by achieving low latency, bounded staleness, and rare data inconsistencies.