How Multimodal Ranking Cuts Slow Query Optimization Time by 14%

Opening

Recently, a paper titled “RCRank: Multimodal Ranking of Root Causes of Slow Queries in Cloud Database Systems” was accepted at VLDB 2025. The work improves slow‑query optimization efficiency by about 14% compared with state‑of‑the‑art root‑cause analysis methods.

Background

Enterprises and individuals increasingly migrate databases to the cloud, but slow queries cause economic loss and erode trust. Slow queries may stem from internal factors such as missing indexes or poorly written SQL, as well as external factors like I/O bottlenecks or network issues. The goal is to provide a framework that identifies and ranks internal root causes so that the most impactful ones can be addressed.

Challenges

Existing approaches suffer from two main limitations. First, they focus on identifying the type of root cause without quantifying its impact, making it hard to prioritize fixes. Second, they rely on a single modality (e.g., CPU or memory usage) and ignore other valuable signals such as query text, execution plans, and logs, resulting in incomplete observability.

Breakthrough

The proposed multimodal diagnostic framework consists of two key components:

Slow‑query and root‑cause collection : Continuous monitoring of cloud database instances gathers queries that exceed a latency threshold. Rule‑based and large‑language‑model (LLM) analyses generate candidate root causes, and the impact is computed as (original execution time – corrected execution time) / original execution time.

Multimodal root‑cause diagnosis : Query statements, execution plans, execution logs, and key performance indicators are encoded, fused via a cross‑feature extractor, and used to estimate each root cause’s impact, producing a ranked list.

Application

Future work will explore integrating RCRank with existing Hologres instances to enhance their diagnostic capabilities.