Doux: Decoupling Values from Keys for Real-Time Analytics

The Log-Structured Merge-tree (LSM-tree), widely adopted in real-time analytics systems, provides highperformance key-based operations such as point lookups and data ingestion, but suffers from poor efficiency on value-based queries, particularly range queries over secondary attributes. In this work, we propose Doux, a dual-tree storage architecture that combines a Key-Ordered Tree (KO-tree) and a Value-Ordered Tree (VO-tree) to accelerate value-range queries while preserving the high ingestion throughput and efficient key-based lookups of traditional LSM-trees. To further reduce maintenance overheads, Doux introduces two optimizations: a decoupled compaction mechanism that mitigates write amplification on the KO-tree, and DropMap, a compact auxiliary structure that reduces read amplification. Extensive experiments demonstrate that Doux significantly outperforms state-of-the-art LSM-tree baselines, achieving 5.04× average speedup on value-range queries, 2.86× faster average real-time ingestion in write-intensive workloads, and 2.51× faster average point lookups in read-intensive workloads.