StreamingDiskANN

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

Performance

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.