D-HASH: Dynamic Hot-key Aware Scalable Hashing for Load Balancing in Distributed Cache Systems

• Hyeok Bang,
• Sanghoon Jeon,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.013, Download Paper (Free):

• Distributed Cache Systems
• Consistent Hashing
• Hot-Key Mitigation
• load balancing
• tail latency
• Read-Path Alternation

Abstract

In large-scale key-value stores, a small number of popular keys can temporarily monopolize traffic, causing a bottleneck on a single node and leading to unstable tail latency. To mitigate this hot-key bias with minimal state and computational overhead, this paper proposes Dynamic Hot-key Aware Scalable Hashing (D-HASH), a lightweight routing layer built on top of Consistent Hashing (CH). D-HASH promotes only keys with read frequencies exceeding a threshold to hot-keys. To maintain data consistency, it always fixes write requests to the primary node, while read requests for promoted keys are distributed between primary and alternate nodes using deterministic window-based switching. This mechanism incorporates a guard phase immediately after promotion to prevent 'cold start' issues and ensure system stability. Alternate nodes are derived from the CH ring using key-specific auxiliary hashes, eliminating the need for external load sensing. The effectiveness of the proposed technique is verified in a Redis-based experimental environment through comparisons with CH, Weighted Consistent Hashing (WCH), and Rendezvous (HRW). Comprehensive evaluations based on NASA HTTP logs and eBay auction logs revealed that D-HASH effectively mitigates load skew and stabilizes latency variability as access skew increases. Furthermore, D-HASH demonstrated algorithmic robustness by maintaining comparable performance even in environments with low access skew, such as the eBay dataset. The ability to be reproducibly tuned with just a few parameters confirms its suitability for real-world deployment in high-availability, read-intensive distributed systems.

Statistics

Cite this article