Overview
This research paper published on arXiv (2510.25401) in December 2025 proposes a novel decoupled storage architecture for vector databases that dramatically improves update performance while maintaining query efficiency.
Key Innovation
Decoupled Storage Architecture
The paper introduces a new architectural approach that separates:
- Index structures from data storage
- Read paths from write paths
- Hot data from cold data
This decoupling enables independent optimization of different workload patterns.
- 10.05x faster insertions compared to traditional architectures
- 6.89x faster deletions with maintained query performance
- Reduced write amplification
- Better resource utilization
- Improved scalability for dynamic datasets
Technical Contributions
Architecture Components
- Dual-Layer Index: Separate mutable and immutable index structures
- Asynchronous Merging: Background consolidation of index updates
- Smart Caching: Adaptive cache management for mixed workloads
- Version Control: Multi-version concurrency control for consistent reads
Algorithms
- Novel index merging strategy
- Adaptive rebalancing mechanisms
- Optimized deletion handling
- Efficient space reclamation
Experimental Results
Benchmarked against:
- Faiss
- HNSWlib
- Traditional vector databases
Across datasets:
Applications
- Real-time recommendation systems
- Dynamic content platforms
- Continuously updated knowledge bases
- Stream processing with vector search
- Applications requiring frequent updates
Future Work
The paper identifies opportunities for:
- Distributed implementation
- GPU acceleration
- Integration with existing vector databases
- Cloud-native deployments
Availability
Free access on arXiv. Code and datasets planned for release.
