1B row vector search in less than a second with Azure SQL Database Hyperscale | Data Exposed

This episode focuses on how vector indexing works in Azure SQL Database Hyperscale, what customer requirements drove the design, and what capabilities are now generally available in Azure SQL PaaS and Fabric SQL.

Overview

The Azure SQL team walks through the evolution of vector search requirements from real customer scenarios (for example, finding similar support cases), and explains how Azure SQL addresses scale, filtering, and data modification needs while keeping query latency low.

Key topics covered

Customer-driven requirements for vector search

• Starting point: customers want similarity search (e.g., “find related cases”).
• Scaling requirement: the same scenario needs to work as data grows significantly.
• Filtering requirement: customers want vector search combined with additional predicates (not just pure similarity search).
• Mutability requirement: customers want to insert, update, and delete rows without having to rebuild or manually maintain the vector index.

Why exact search is not enough

• The discussion contrasts exact search with approximate approaches for vector similarity.
• The goal is to keep results fast while still returning useful “nearest neighbor” matches.

DiskANN: graph-based vector indexing

• The team introduces DiskANN as the underlying approach for the vector index.
• DiskANN is described as a graph-based approximate nearest neighbor (ANN) index designed for performance at scale.

Filtering behavior: iterative filtering vs post-filtering

• The episode discusses the need to combine vector similarity with filters.
• It contrasts iterative filtering with post-filtering as approaches to applying predicates alongside vector search.

DML support with vectors

• The team demonstrates that vector scenarios support standard data modification operations:
  • INSERT
  • UPDATE
  • DELETE
• The intent is to avoid workflows where index maintenance becomes a separate manual step.

Optimizer choice: approximate vs exact vector search

• The episode notes that approximate vs exact vector search can be handled by the query optimizer.

Billion-row scale performance

• A key scenario discussed is searching over 1 billion rows.
• The episode claims vector search can still return results within milliseconds in that scale scenario.

Availability

• Vector Index in Azure SQL PaaS and Microsoft Fabric SQL is stated as generally available (GA).

People and links mentioned

• LinkedIn: Pooja Kamath
• LinkedIn: Krithika Subramanian
• Twitter: Anna Hoffman
• Twitter: AzureSQL
• Data Exposed playlist: https://aka.ms/dataexposedyt
• Microsoft Azure SQL channel: https://aka.ms/msazuresqlyt
• Microsoft SQL Server channel: https://aka.ms/mssqlserveryt
• Microsoft Developer channel: https://aka.ms/microsoftdeveloperyt