Overview of Azure HorizonDB

Key Concepts

• Azure Horizon DB: A cloud-native, PostgreSQL-compatible relational database service.
• Disaggregation of Compute and Storage: The architectural separation of processing power from data storage to allow independent scaling.
• Database-as-a-Log Architecture: A design where the compute node offloads transaction logs (WAL) to a distributed storage fleet.
• ACID Compliance: Atomicity, Consistency, Isolation, and Durability; ensuring reliable database transactions.
• Write-Ahead Log (WAL): A transaction log used to ensure data integrity and durability.
• NVMe Pass-through: A technique to accelerate performance by bypassing kernel I/O stack layers.
• Zone Redundant Blob (ZRS): Azure storage service used for underlying durability.

---

1. Architecture and Design

Azure Horizon DB is a fork of PostgreSQL designed to be cloud-native. Unlike traditional PostgreSQL, which couples compute and storage on a single primary node, Horizon DB disaggregates these layers.

• Compute Layer: Stateless primary instance that handles transactions. It can scale up to 192 vCores, with 8GB of memory per vCore.
• Storage Fleets:
  • WAL Storage Fleet: Distributed across multiple Availability Zones (AZs). It uses NVMe pass-through for high-performance, append-only writes. A quorum of these instances must acknowledge a transaction for it to commit.
  • Data Storage Fleet: Stores actual data pages, transparently sharded into 1GB segments across the fleet. It reconstructs data pages by applying WAL content.
• Caching Hierarchy: The compute node utilizes local NVMe SSDs (larger than memory) as a cache. Data is fetched from the remote Data Storage Fleet only if it is not found in memory or the local SSD cache.

2. Scalability and Availability

• Horizontal Scaling: Users can add up to 15 standby/read replicas. These replicas share the same data storage fleet, eliminating the need for traditional PostgreSQL replication.
• Failover: Because the WAL storage fleet acts as the single source of truth and the compute layer is stateless, failover is significantly faster than standard PostgreSQL, requiring no reinitialization.
• High Availability: The system supports multi-AZ deployment by default. Geo-HA (High Availability) is planned for General Availability (GA), utilizing asynchronous replication to a secondary region.

3. Compatibility and Integration

• API Compatibility: Horizon DB maintains 100% PostgreSQL compatibility, meaning existing applications require zero code changes to migrate.
• Extensions: Supports 75+ popular PostgreSQL extensions, including pgvector for AI embeddings, diskANN for vector search, and BM25 for text ranking.
• AI Pipelines: Features declarative SQL-based AI workflows for chunking, embedding, and ranking, with integrated model management via Microsoft Foundry.
• Data Virtualization: Capable of mirroring data to Microsoft Fabric for enterprise-wide semantic modeling.

4. Operational Processes

• Backups: Managed via snapshots of the storage fleet.
• Storage Management: Storage auto-scales up to a current limit of 128 TB; manual provisioning of storage is not required.
• Authentication: Integrated with Microsoft Entra ID and supports private endpoints.

5. Pricing Model

Pricing is based on four primary dimensions:

1. Compute: Number of vCPUs (includes memory).
2. Standbys: Cost for optional read replicas.
3. Storage: Total size of the database.
4. Backups: Snapshot storage costs.

• Note: There is no separate charge for IOPS or throughput, as these are bundled into the storage cost. Users are advised to re-evaluate sizing, as the offloading of storage tasks to the fleet may allow for smaller compute instances compared to traditional PostgreSQL deployments.

6. Synthesis and Conclusion

Azure Horizon DB represents a shift toward cloud-native database architectures. By decoupling compute from storage and utilizing a distributed log-based architecture, it addresses the traditional bottlenecks of PostgreSQL (scaling, performance, and availability) without sacrificing the ecosystem or API compatibility. While currently in preview and not recommended for production, it offers a promising path for applications requiring high-scale, high-performance relational database capabilities with seamless integration into the modern AI-driven cloud ecosystem.