TLDR: OntoTune is a new platform that uses an ontology and knowledge graph to enhance machine learning for database query optimization. By providing a structured, semantic layer for database queries, execution plans, and system configurations, OntoTune helps convolutional neural networks make more informed decisions, leading to improved query performance and more stable learning outcomes compared to traditional methods. The system is a PostgreSQL-based prototype that aims to address the brittleness and lack of context in current learned optimizers.
Database systems are the backbone of countless applications, and their efficiency hinges significantly on how quickly and effectively they can process queries. This process, known as query optimization, has been a long-standing challenge. Traditional methods often rely on fixed heuristics, while more recent machine learning (ML) approaches, though promising, can be inconsistent and sensitive to various factors like data distribution or system configurations.
A new research paper, OntoTune: Ontology-Driven Learning for Query Optimization with Convolutional Models, introduces an innovative platform called OntoTune that aims to bring more stability and context to learned query optimization. Authored by Songhui Yue, Yang Shao, and Sean Hayes from Charleston Southern University, OntoTune proposes an ontology-driven approach to enhance how machine learning models understand and optimize database queries.
The Challenge with Current Query Optimization
Modern database workloads, such as those found in benchmarks like Jobs/Join-Order Benchmark (JOB) and StackOverflow, present complex scenarios with ‘long-tail queries’ – queries that are rare but can be extremely costly to execute. Existing learning-based optimizers, while helpful, often struggle with these unique queries and can be brittle, with performance varying based on training conditions, hardware, and database settings. This makes experiments hard to reproduce and outcomes difficult to explain.
OntoTune’s Ontology-Driven Solution
OntoTune tackles these issues by integrating an ontology and a concrete knowledge graph (KG) into the query optimization process. An ontology provides a structured, information-rich representation of knowledge, defining the types of objects and their relationships that are crucial for plan selection. The knowledge graph then stores actual instances of these objects and relations from real database operations.
This approach means that everything relevant to query performance – from SQL queries themselves, to database metadata, execution plans, system configurations (like PostgreSQL GUCs), runtime metrics, and even model predictions – becomes a ‘first-class, queryable entity’. By organizing this information semantically, OntoTune provides a missing layer of context and provenance that helps learning algorithms make more informed decisions.
How OntoTune Works
The platform is built as a PostgreSQL-based prototype, extending existing learned optimizer architectures. It includes a module that extracts comprehensive metadata and system configurations directly from PostgreSQL. This extracted data, which includes query metadata, database statistics, plan trees, and runtime metrics, is then used to create a feature matrix. This matrix, enriched by the ontology, fuses SQL, plan, and context information into a structured representation.
OntoTune utilizes a Convolutional Neural Network (CNN) as its primary predictor. The CNN is trained on this ontology-derived feature matrix to predict the ‘reward’ (inverse of cost) of different execution strategies. To address challenges in online learning, especially with less frequently used but potentially optimal strategies, OntoTune employs a ‘reward-cost complement transform’ trick. This helps the model avoid prematurely dismissing arms that might perform poorly initially but could be beneficial with more exploration.
Also Read:
- Unlocking Efficient Querying of Inconsistent Knowledge Bases with Cost-Based Semantics
- AI Uncovers Hidden Subclasses in Time Series Data Using Saliency Maps and LLMs
Promising Results and Future Directions
Case studies using the StackOverflow dataset demonstrated that OntoTune can outperform PostgreSQL’s default query execution in certain scenarios. The learned policy effectively leverages past executions to identify better strategies, significantly reducing completion times for batches of queries by avoiding costly stalls. However, the research also acknowledges that performance can be sensitive to environmental changes and that early exploration might sometimes lead to suboptimal choices.
The researchers emphasize that OntoTune is a platform designed for ongoing development. Future work includes expanding the ontology with more ‘tail-aware’ features (e.g., spill risk, memory pressure), implementing environment-aware transfer learning for greater stability, and exploring other advanced learning models like Graph Convolutional Networks (GCNs). There are also plans to integrate Large Language Models (LLMs) with the ontology for template discovery and configuration suggestions, aiming for a more systematic, explainable, and reproducible path to learned query optimization.
