Reducing Tail Latency in Storage-Disaggregated Database Systems

Key findings from this study

• The study found that log replay chains cause tail latency in storage-disaggregated databases due to the log-as-the-database design principle.
• The authors report that Replay-as-a-Service reduces P95 tail latency by 40.1% by decoupling log replay from the storage engine.
• The researchers demonstrate that executing log replay on idle cluster servers improves overall throughput by 75.9% in the implemented OpenAurora system.

Overview

Storage-disaggregated databases in cloud environments decouple compute and storage resources to improve resource utilization and enable independent scaling. The log-as-the-database architecture sends only transaction logs to the storage engine, requiring log replay to reconstruct pages. This design produces tail latency when certain page requests encounter lengthy log replay chains. The study introduces Replay-as-a-Service (RaaS), which decouples log replay logic from the storage engine into an independent service.

Methods and approach

RaaS executes log replay operations on idle or dedicated servers within the cluster rather than within the storage engine itself. The authors implemented RaaS in OpenAurora, an open-source storage-disaggregated database built on PostgreSQL. Experiments used SysBench workloads to evaluate latency and throughput characteristics against baseline storage-disaggregated architectures.

Results

RaaS reduced P95 tail latency by 40.1% compared to standard storage-disaggregated configurations. Overall throughput improved by 75.9% under the test workloads. These improvements stem from distributing log replay computation across available cluster resources, reducing contention at the storage engine and enabling better resource utilization of idle servers.

Implications

The findings demonstrate that architectural decoupling of log replay from the storage layer addresses a fundamental performance constraint in storage-disaggregated database systems. By treating replay as an independent service, database operators can leverage underutilized compute resources to accelerate critical operations. This approach extends the practical applicability of cloud storage disaggregation to latency-sensitive OLTP workloads.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.