[Paper Review] Continuous Partial Quorums for Consistency-Latency Tuning in Distributed NoSQL Storage Systems

Marlon McKenzie, Hua Fan+1 more|arXiv (Cornell University)|Jul 11, 2015

Cloud Computing and Resource Management21 references3 citations

TL;DR

This paper proposes Continuous Partial Quorums (CPQ), a technique for fine-grained consistency-latency tuning in NoSQL systems by stochastically selecting between two consistency levels (e.g., read-one and majority) using a tunable probability. Evaluated on Apache Cassandra, CPQ achieves the same consistency as artificial delays with significantly lower latency, outperforming fixed consistency levels and delay-based methods.

ABSTRACT

NoSQL storage systems are used extensively by web applications and provide an attractive alternative to conventional databases when the need for scalability outweighs the need for transactions. Several of these systems provide quorum-based replication and present the application developer with a choice of multiple client-side "consistency levels" that determine the number of replicas accessed by reads and writes, which in turn affects both latency and the consistency observed by the client application. Since using a fixed combination of read and write consistency levels for a given application provides only a limited number of discrete options, we investigate techniques that allow more fine-grained tuning of the consistency-latency trade-off, as may be required to support consistency-based service level agreements (SLAs). We propose a novel technique called \emph{continuous partial quorums} (CPQ) that assigns the consistency level on a per-operation basis by choosing randomly between two options, such as eventual and strong consistency, with a tunable probability. We evaluate our technique experimentally using Apache Cassandra and demonstrate that it outperforms an alternative tuning technique that delays operations artificially.

Motivation & Objective

• To address the limitation of discrete consistency levels in NoSQL systems, which offer only a small number of fixed trade-offs between consistency and latency.
• To enable continuous, per-operation tuning of consistency and latency to support service-level agreements (SLAs) with precise latency and staleness requirements.
• To evaluate and compare CPQ against artificial delays (AD), a competing technique that injects client-side delays to improve consistency at the cost of higher latency.
• To demonstrate that CPQ provides a superior consistency-latency trade-off compared to artificial delays in real-world NoSQL deployments.
• To lay the foundation for a middleware framework supporting probabilistic consistency and latency guarantees in distributed storage systems.

Proposed method

• CPQ dynamically selects between two discrete consistency levels (e.g., read-one and majority quorum) for each operation based on a tunable probability p ∈ [0,1], enabling continuous control over the consistency-latency trade-off.
• The technique operates entirely at the client side, using probabilistic selection to balance consistency and latency without modifying the underlying storage system.
• For each read or write operation, the client randomly chooses one of two consistency levels with probability p and 1−p, respectively, allowing smooth tuning across the consistency-latency spectrum.
• The method is evaluated experimentally using Apache Cassandra deployed in Amazon EC2, with workloads generated via the Yahoo Cloud Serving Benchmark (YCSB).
• The performance of CPQ is compared against artificial delays (AD), where delays are injected before or after operations to increase the likelihood of reading fresh data.
• Consistency is measured by the fraction of stale reads, and latency is measured as the 95th percentile response time, enabling direct comparison of trade-offs under SLA constraints.

Experimental results

Research questions

• RQ1 Can a continuous, per-operation consistency-latency trade-off be achieved in quorum-replicated NoSQL systems using only existing discrete consistency levels?
• RQ2 How does CPQ compare to artificial delays in achieving a desired consistency-latency trade-off, particularly in terms of latency and staleness?
• RQ3 Can CPQ achieve the same level of consistency as artificial delays while significantly reducing latency in real-world deployments?
• RQ4 Does CPQ outperform fixed consistency levels (e.g., read-one, majority) in supporting service-level agreements with specific latency and staleness constraints?
• RQ5 Is the CPQ technique generalizable to other NoSQL systems that support multiple client-side consistency levels?

Key findings

• CPQ achieves the same level of consistency (measured as fraction of stale reads) as artificial delays but with severalfold lower latency in multiple experimental scenarios.
• The technique enables continuous tuning of the consistency-latency trade-off using only two discrete consistency levels, overcoming the limitations of fixed, discrete combinations.
• In some cases, CPQ reduced latency by a factor of several compared to artificial delays while maintaining equivalent consistency guarantees.
• The experimental results confirm that injecting artificial delays into the critical path of every read operation is detrimental to latency, supporting prior informal claims.
• CPQ outperforms both fixed consistency levels and artificial delays in achieving optimal trade-offs under SLA constraints, particularly for latency-sensitive workloads.
• The method is applicable to any NoSQL system supporting multiple client-side consistency levels, suggesting broad practical relevance.