[Paper Review] Key-Value Stores on Flash Storage Devices: A Survey

Krijn Doekemeijer, Animesh Trivedi|arXiv (Cornell University)|May 11, 2022

Advanced Data Storage Technologies4 citations

TL;DR

This survey synthesizes a decade of research on optimizing key-value stores for flash storage, analyzing how flash-specific characteristics like write amplification, wear leveling, and asymmetric I/O necessitate architectural changes. It identifies core design trade-offs and forecasts future trends in flash-optimized key-value systems.

ABSTRACT

Key-value stores (KV) have become one of the main components of the modern storage and data processing system stack. With the increasing need for timely data analysis, performance becomes more and more critical. In the past, these stores were frequently optimised to run on HDD and DRAM devices. However, the last decade saw an increased interest in the use of flash devices because of their attractive properties. Flash is cheaper than DRAM and yet has a lower latency and higher throughput than HDDs. This literature survey aims to highlight the changes proposed in the last decade to optimise key-value stores for flash devices and predict what role these devices might play for key-value stores in the future.

Motivation & Objective

• To analyze how flash storage's unique properties—such as low latency, asymmetric I/O, and finite write endurance—have reshaped key-value store design over the past decade.
• To identify and categorize flash-specific optimizations in key-value stores, including data structures, storage interfaces, and system-level software techniques.
• To evaluate the trade-offs between performance, durability, and resource overhead in flash-optimized key-value stores.
• To explore the evolving role of flash in key-value store workloads, from primary storage to caching and cold-tier storage.
• To identify open research challenges and future directions for key-value stores on flash, including benchmarking, integration with broader data ecosystems, and support for emerging flash technologies.

Proposed method

• Systematically surveying academic literature from the past decade on key-value stores and flash storage, focusing on optimization strategies for flash-specific constraints.
• Classifying key-value store architectures based on data structures (e.g., LSM-tree, B-tree, hash tables) and their adaptation to flash I/O patterns and wear constraints.
• Analyzing communication layers between key-value stores and flash devices, including NVMe, FTL, and host-managed SSDs (e.g., ZNS, OCSSD).
• Evaluating software-level optimizations such as write-ahead logging, value logging, and memory mapping to reduce I/O and CPU overhead.
• Surveying techniques to mitigate key flash-specific issues: write amplification, read amplification, space amplification, and garbage collection overhead.
• Assessing the role of emerging flash technologies (e.g., ZNS, QLC, SCM) and their implications for future key-value store design.

Experimental results

Research questions

• RQ1 What is the current role of flash storage in key-value store deployments across different workloads and systems?
• RQ2 How have key-value store designs evolved in response to flash storage characteristics over the past decade?
• RQ3 What are the primary challenges in deploying key-value stores on flash, and which optimization strategies effectively mitigate them?
• RQ4 How will flash storage continue to influence the future design and performance of key-value stores?

Key findings

• Flash storage has become the dominant medium for key-value stores, used across data centers, gaming services, and web applications due to its balance of cost, performance, and durability.
• Key-value stores have undergone significant architectural adaptation to flash, including the widespread adoption of LSM-trees and value logging to reduce write amplification and improve write throughput.
• Major challenges remain in balancing trade-offs between write amplification, read amplification, space amplification, and garbage collection overhead, with no single solution addressing all simultaneously.
• Flash-specific optimizations such as host-managed FTL (e.g., ZNS SSDs) and NVMe-aware data structures are increasingly being integrated into production systems to reduce CPU and I/O overhead.
• Future work is needed to standardize benchmarking and evaluation frameworks for flash-optimized key-value stores to support informed deployment decisions.
• Emerging flash technologies like QLC, ZNS, and storage class memory (SCM) are expected to further influence key-value store design, enabling new performance and cost trade-offs.