Don't Forget Range Delete! Enhancing LSM-based Key-Value Stores with More Compatible Lookups and Deletes

TL;DR

This paper introduces GLORAN, a novel method for range deletes in LSM-trees that improves point lookup efficiency and overall performance by using a global index and entry validity estimation.

Contribution

GLORAN provides a new range delete technique compatible with modern LSM-based systems, significantly reducing lookup latency and I/O costs while maintaining high throughput.

Findings

GLORAN achieves up to 10.6x faster point lookups.

GLORAN increases overall throughput by 2.7x.

It reduces I/O complexity from O(N/λ) to near logarithmic scales.

Abstract

LSM-trees are featured by out-of-place updates, where key deletion is handled by inserting a tombstone to mark its staleness instead of removing it in place. This defers actual removal to compactions with greatly reduced overhead. However, this classic strategy struggles with another fundamental operator--range deletes--which removes all keys within a specified range, requiring the system to insert numerous tombstones and causing severe performance issues. To address this, modern LSM-based systems introduce range tombstones that record the start and end keys to avoid per-key tombstones. Although this achieves impressive range delete efficiency, such a solution is incompatible with lookups. In particular, our experiments show that point lookup latency can increase by 30% even with just 1% range deletions in workloads. Further to our surprise, this issue has not been raised before, though the range tombstone solution has been employed for more than five years. To address this critical performance issue, we propose GLORAN, an efficient range delete method that can be integrated into modern LSM-based systems and offers desirable range deletion performance without compromising point lookup efficiency. It introduces a global index that allows point lookups to quickly locate relevant ranges without retrieving many irrelevant elements, reducing the I/O complexity from O(N/\lambda) to either O(\log^2 N/(\lambda F)) or O(\phi \log N/F), where 1/\lambda is the ratio of range deletes, and \phi is the FPR of Bloom filters in LSM-trees. Furthermore, we design an entry validity estimator to further enhance expected I/O cost to O(\epsilon \log^2 N/(\lambda F)) for looking up existing keys. Extensive evaluations indicate that GLORAN consistently outperforms baselines, while achieving up to 10.6 times faster point lookups and 2.7 times higher overall throughput compared to the SOTA method.