Research
From query optimisation to complete data systems
My research asks where learned components can improve a database system, and which database guarantees the surrounding system must retain. The work spans algorithms, database internals, and end-to-end implementation.
Current direction
Semantic Query Processing
Semantic workloads place expensive model-backed operators in the same plans as conventional relational operators. I am studying how a database system should execute and optimise these mixed workloads while retaining control over latency, resource use, and result quality.
Research questions
- How should semantic operators interact with relational execution inside a database engine?
- How should an optimiser compare plans when model calls are expensive and several objectives matter?
- Which parts of execution and optimisation should remain explicit and controllable at the database layer?
Status: ongoing research. Project names, manuscript details, and experimental results will be added when they are public.
System building · ICDE 2026
Real-Time Multimodal Data Systems
ARCADE is an open-source system for real-time hybrid and continuous query processing across vector, spatial, text, image, and relational data. Built on RocksDB and MySQL, it connects storage, indexing, optimisation, and incremental query processing in a working database system.
My contribution
I led the design and implementation of the query-optimisation component and its end-to-end integration. This included connecting system statistics, access paths, and heterogeneous operators in a cost-based optimiser, while working with collaborators responsible for storage, indexing, and continuous-query support.
Evidence: IEEE ICDE 2026; open-source implementation. PaperCode
Research foundations · SIGMOD and PVLDB
AI for Database Optimisation
This programme studies three ways to use learning in query optimisation: replacing selected optimisation logic, enhancing an existing optimiser while preserving native database guarantees, and transferring reusable query-plan representations across tasks.
Replacement
lemo studies learned optimisation for concurrent-query workloads, combining plan search with reuse of intermediate results. I led this work and implemented its PostgreSQL prototype.
Enhancement
RankPQO combines plan enumeration with learning-to-rank for parameterised plan selection. The work was developed and evaluated in PostgreSQL.
Transfer
TATA develops transferable query-plan representations so that knowledge learned for one database task can reduce data collection for another.
Industry translation
In an industry-funded NTU collaboration with OceanBase, an enterprise-grade distributed database originating from Ant Group, I led the day-to-day technical development of the parametric query-optimisation work. The resulting ScalePQO techniques were integrated into OceanBase and evaluated on industry workloads.
Outputs: lemo (ACM SIGMOD 2024), RankPQO (PVLDB / VLDB 2025), TATA (PVLDB / VLDB 2026), and ScalePQO (PVLDB Industry Track 2026).
Earlier research
Algorithms for Data-Intensive Applications
My earlier work developed algorithms for graph analytics and urban data applications. It includes weighted random-walk domination, public-transport scheduling, and outdoor-advertising placement over large trajectory and graph datasets.
These projects provided the algorithmic foundation for my later systems work: formal problem definition, approximation and pruning techniques, and empirical evaluation on real datasets.
Selected outputs: PVLDB 2021, IEEE TKDE, and ACM KDD 2019 Best Paper Runner-up.
Explore