Overview
StreamingDiskANN is a custom index type in pgvectorscale inspired by Microsoft's DiskANN algorithm, optimized for disk-based storage with efficient streaming updates.
Key Features
- Disk-based index supporting billions of vectors
- Streaming updates without full rebuilds
- SSD-optimized for sequential I/O
- Memory-efficient with configurable cache
- PostgreSQL-native implementation
Architecture
The index uses:
- Graph-based ANN structure (HNSW-like)
- Disk layout optimized for SSDs
- Incremental updates via streaming
- LRU cache for hot data
- Background compaction
On 50 million Cohere embeddings:
- 471 QPS at 99% recall
- 28x lower p95 latency vs Pinecone
- 16x higher throughput vs Qdrant
- 75% lower cost when self-hosted
Comparison to Alternatives
- In-memory HNSW: Faster but memory-limited
- Traditional DiskANN: Requires full rebuilds
- StreamingDiskANN: Best of both worlds
Integration
Seamlessly integrates with PostgreSQL:
- Standard SQL syntax
- Transaction support
- pg vector compatibility
- Familiar tooling
Pricing
Open-source as part of pgvectorscale.
