Songsong Mo

Research direction

Heterogeneous and Multimodal Data Systems

ARCADE is an open-source system for real-time hybrid and continuous queries over relational, vector, spatial, text, and image data, built on RocksDB and MySQL. It brings multimodal indexing, cost-based optimisation, and incremental query processing into one implemented system.

ARCADE system architecture The query interface communicates with a query processing layer containing a write thread and read threads for hybrid and continuous queries. The processing layer communicates with an LSM-tree data store and an incremental materialised view. The LSM-tree contains an in-memory write buffer, block cache and global secondary index, together with per-SST unified secondary indexes and data on disk. Query interface layer Query processing layer Write thread Read threads Hybrid query processing Continuous query processing Storage layer LSM-tree data store Incremental materialised view LSM-tree organisation Memory Disk Write buffer Block cache Global secondary index L0 Unified secondary index Data L1 Unified secondary index Data Unified secondary index Data Unified secondary index Data × n
System architecture, redrawn from Figure 1 of the ARCADE paper. MySQL provides the query engine and RocksDB the storage engine.

Storage and indexing

Unified multimodal indexing

ARCADE integrates disk-based secondary indexes for vector, spatial, and text data with LSM-tree storage. This allows relational filters and specialised similarity or spatial predicates to operate over the same stored data rather than through separate systems.

Query processing

Query processing and optimisation

  • Hybrid queries

    The cost-based optimiser combines multiple access paths for hybrid search and hybrid nearest-neighbour queries, including vector, spatial, text, and relational operators. It connects system statistics and heterogeneous indexes with MySQL plan selection.

  • Continuous queries

    ARCADE supports time-based and event-driven continuous queries through incremental materialised views. Intermediate results are reused as data changes, allowing the system to maintain fresh answers without repeatedly executing each query from scratch.

As co-first author, I led the design and implementation of ARCADE's hybrid-query optimiser and its incremental processing for continuous queries. I contributed to technical discussions across the broader system; other team members held primary ownership of storage and indexing.

Public system and evidence

The ICDE 2026 evaluation reports improvements of up to 7.4× in read-heavy settings and 1.4× in write-heavy settings over the best-performing evaluated baseline.

IEEE ICDE paperarXivCode

Back to the research overview