A secure cloud database connects data centers worldwide, illustrated with a global map, locks, and a shield.
Microsoft has declared Azure HorizonDB generally available. The announcement came at Ignite on November 17, 2026. It is live in 13 Azure regions across the Americas, Europe and Asia Pacific, with more regions coming online soon. The Azure Updates entry carries the "Launched" label. That label means fully released, production-ready and open to all Azure customers. HorizonDB is Microsoft's PostgreSQL-compatible managed database, and until now it was a preview you could only test.

Microsoft's performance numbers are large. Most of them are the company's own benchmarks. Treat them as a reason to run your own tests, not as a promise.

A secure cloud database connects data centers worldwide, illustrated with a global map, locks, and a shield. What HorizonDB is​

HorizonDB is a fully managed database service built on the open-source PostgreSQL engine. Microsoft aims it at mission-critical and AI-powered applications. Microsoft says it introduced the service in preview at Ignite last year. Other coverage dates the Build 2026 public preview to June 2, 2026.

Microsoft's documentation describes the architecture. A cluster has one or more compute replicas, one writable and the rest readable standbys. All replicas share a single copy of the data in zone-resilient storage. A read-write endpoint always points to the primary. A read-only endpoint load-balances across the readable replicas. That documentation page was written during preview, so check its details against current docs before you rely on them.

Scale, speed and availability claims​

Microsoft's headline claims for the GA service:

  • Size: databases up to 128 TB.
  • Speed: up to 8x faster OLTP performance than self-managed PostgreSQL.
  • Resilience: every database is zone-resilient by default, with sub-millisecond commit latency and failover in under five seconds.

The 8x figure comes with disclosed test conditions. Microsoft ran HammerDB's TPROC-C workload with 95,000 warehouses and 512 users. The comparison system ran on Azure Edsv6 VMs with 64 vCores, PostgreSQL 17.11, in Korea Central across Availability Zones 1 and 2. Microsoft says both sides ran with high availability on comparable hardware. This is vendor benchmarking, not independent measurement. Your own queries, schema and concurrency will decide what you see.

The latency and failover figures are also vendor claims. Microsoft gives no workload qualifiers for them. Test failover against your own applications' connection-retry behavior before you count on a five-second recovery.

Security and identity​

For administrators, Microsoft names three integrations:

  • Microsoft Entra ID authentication for identity-based access and consistent governance.
  • Private endpoints to keep connectivity inside your virtual network boundary.
  • Customer-managed keys so you control the encryption keys under your own policies.

Customer-managed keys matter for a specific reason. Third-party coverage of the preview said they were not available then. Microsoft now offers them in public preview: the Ignite blog lists them among the service's security capabilities, but Microsoft Learn still labels customer-managed keys as preview, so they are not a GA capability. If you evaluated HorizonDB during preview and rejected it on encryption grounds, look again, but plan around preview terms.

AI features​

Microsoft pitches HorizonDB as a place to run retrieval, graph reasoning and AI workflows without separate specialized databases.

  • Vector and text search: pgvector support means existing vector applications run as they are. BM25 ranking comes through pg_textsearch. Hybrid search combines vector and keyword results with Reciprocal Rank Fusion. Microsoft says this surfaces context that either method alone might miss.
  • DiskANN indexing: Microsoft claims up to 1.5x higher throughput than HNSW at comparable recall, and indexes up to 5x smaller. It gives no test conditions for this claim.
  • Apache AGE graphs: Microsoft says the AGE extension API is unchanged, so existing AGE applications should run without code changes. Against community Apache AGE, it reports:
  • Simple reads up to 3,900x faster (LDBC SNB Interactive workload).
  • Complex multi-hop reads up to 512x faster (same workload).
  • Graph builds and refreshes up to 190x faster, with about 6x lower peak memory (internal Microsoft testing).
  • AI Functions and AI Pipelines: SQL-callable functions invoke models for embeddings, extraction and content generation. Pipelines chain them into durable, asynchronous workflows, such as enriching new rows or keeping a knowledge graph current.
  • Agent tooling: Microsoft's PostgreSQL MCP Server gives agents a standard way to discover and invoke HorizonDB capabilities through the Model Context Protocol. PostgreSQL Skills supply reusable guidance for common database tasks.

Not every feature is GA just because the service is. Microsoft's database roundup describes built-in AI model management as in preview. A same-day community report says Knowledge Graphs are also in public preview and the Database Agent for PostgreSQL is in limited public preview. That report comes from another WindowsForum article, not Microsoft. Check each feature's status before you design around it.

The "up to" multipliers for AGE are maximums. Microsoft's own tests used specific workloads, so benchmark with your own graph shapes.

Developer and operations tooling​

The PostgreSQL extension for Visual Studio Code is the main management surface. Microsoft lists these features:

  • Cluster topology views showing pools, nodes, writer and read-only roles, availability zones, compute configuration, high-availability status and parameter-group sync state.
  • Metrics, wait events, top queries and active sessions for chasing slow queries and blocking.
  • Query plans rendered as interactive graphs, with GitHub Copilot analyzing plans using live schema context.
  • Apache AGE Cypher results rendered as node-edge graphs, and AI pipeline stages visualized from chunking through ranking.

Microsoft's documentation also covers quickstarts for VS Code, Python and Java, backup and restore, high availability, security, vector and hybrid search, graph databases, scaling, replicas and migration by dump and restore.

One oddity: the Microsoft Learn landing page still carries a "Preview" label on its title, even though Microsoft's announcement says the service is GA. Several Learn pages seem to lag the launch. Trust the announcement for release status. Verify specifics, such as supported regions and limits, in the portal.

Migration​

Microsoft says the new Azure Database Migration Service moves PostgreSQL databases to HorizonDB from AWS, Google Cloud, on-premises and Azure virtual machines. It supports online migration of schema and data, with built-in validation and monitoring.

Oracle is a heavier lift. Microsoft treats it as a software engineering project, not just a database task:

  • Schema conversion: the VS Code extension reads the Oracle data dictionary and uses models in Microsoft Foundry to generate PostgreSQL DDL. It compiles the result against a PostgreSQL database and flags objects needing manual work in a conversion report.
  • Application conversion: AI-assisted tooling finds Oracle-specific dependencies in SQL scripts, stored procedure calls and Java code, and proposes changes. Developers review each change in VS Code, GitHub Copilot CLI or Modernize CLI before accepting it.
  • Data movement: offline Oracle migration is in preview, and online migration is described as coming soon. It is not a GA path yet.

What Fabric HorizonDB is​

Don't confuse this launch with Microsoft Fabric HorizonDB. That is a separate offering, in public preview across Fabric regions worldwide. Microsoft describes it as serverless:

  • Databases provision in seconds.
  • Compute pauses after five minutes of inactivity.
  • Compute resumes on the next query.

It also adds translytical queries, mirroring into OneLake and synced tables. None of that is part of the Azure HorizonDB GA.

What to do now​

  • Check your region. Thirteen regions is a start, and Microsoft hasn't published a date for more. Confirm your target region in the portal. During preview, region lists differed between sources.
  • Test against your own workload. The 8x, 1.5x, 3,900x and 512x figures all carry vendor test conditions.
  • Audit feature status. AI model management and the Oracle migration paths are still preview or coming soon.
  • Plan your security setup. Decide on Entra ID authentication, private endpoints and customer-managed keys up front.
  • Don't assume zero code changes. Microsoft's no-change claim applies to AGE API compatibility. Microsoft's own Oracle tooling expects review and regression testing.

Bottom line​

HorizonDB's move to GA gives Azure customers a production-supported option for PostgreSQL at scale, with identity, networking and key controls built in. The AI and graph features are the main selling point. Some of them are still preview, and the benchmark numbers come from Microsoft's own tests. Treat HorizonDB as a serious candidate and check each claim against your own workload.


Correction (November 17, 2026): An earlier version of this article said Microsoft lists customer-managed keys as a GA capability of Azure HorizonDB. That was wrong: Microsoft Learn labels customer-managed keys for HorizonDB as public preview, so they are not generally available. The article now says so.

 

References

  1. (Launched) Generally Available: Azure HorizonDB Azure Updates 2026-11-17T20:00:32Z
  2. azure-databases-docs/articles/horizondb/overview.md at main · MicrosoftDocs/azure-databases-docs · GitHub github.com
  3. HorizonDB: PostgreSQL for enterprises, developers, and agents on Azure and Fabric techcommunity.microsoft.com