Co-partitioned Vector Index

Overview

Co-partitioned Vector Index is an indexing strategy where vector indexes are stored in the same database partitions/tablets as the corresponding table rows, ensuring tight data locality.

Architecture

In a co-partitioned layout:

1. Table rows and their vector indexes live together
2. No cross-partition lookups needed
3. Updates are transactional within a single partition
4. Queries benefit from data locality

Benefits

• Data Locality: Vectors and metadata are physically close
• Performance: Reduced network hops for queries
• Consistency: Easier to maintain ACID properties
• Scalability: Horizontal scaling without coordination overhead
• Operational Simplicity: Unified backup and recovery

Comparison to Alternatives

Separate Vector Store

• Requires cross-system joins
• Complex consistency management
• Higher latency

Global Vector Index

• Single point of bottleneck
• Difficult to scale
• Higher operational complexity

Implementation

YugabyteDB uses co-partitioned vector indexes as part of its Vector LSM architecture, enabling billion-vector scale with sub-second latency.

Use Cases

• Distributed SQL databases with vector support
• Multi-tenant applications
• Geo-distributed systems
• Large-scale production deployments

Pricing

Architectural pattern, used in YugabyteDB and similar distributed databases.