Azure HorizonDB is now generally available. Microsoft closed the preview window at Ignite on November 17, and if you have been waiting to move a production PostgreSQL workload off self-managed VMs or away from AWS, the question is no longer whether HorizonDB is ready — it is. The question is whether it is right for your specific workload.
We covered the public preview in June when Microsoft opened it at Build. This post covers what GA actually delivers and where the gaps still are.
What Changed from Preview to GA
GA is not just a marketing label change. The features that hardened to generally available status at Ignite include Microsoft Entra authentication, monitoring alerts and policy support, system-managed maintenance windows, and the Apache AGE graph extension optimizations — with up to 3,900x faster graph reads on specific query patterns. The Azure Database Migration Service now supports online migration to HorizonDB from AWS RDS, Google Cloud SQL, on-premises PostgreSQL, and Azure VMs.
Still in preview at GA: AI Model Management, Knowledge Graphs, the Database Agent for PostgreSQL, Fabric HorizonDB, and PostgreSQL 18 support. If any of those are requirements, factor preview SLAs into your planning.
The Architecture Behind the Performance Numbers
The headline claim is 8x OLTP throughput over self-managed PostgreSQL on a HammerDB benchmark that Microsoft ran itself. No independent replication exists at GA time, so treat the number as directional. The underlying architecture is sound.
HorizonDB separates compute from storage completely. Compute nodes are stateless and write only the WAL (Write Ahead Log) to the storage fleet. The storage layer absorbs that WAL and reconstructs data pages asynchronously. Transactions commit in a single step rather than the multi-node coordination standard PostgreSQL replication requires. Microsoft reports sub-millisecond multi-zone commit latency and failover in under five seconds. Storage scales to 128 TB automatically; compute scales to 3,072 vCores.
Failover does not copy data — it reroutes DNS. At the scale most applications operate, the difference between five-second and sixty-second failover is a business decision, not just a technical preference.
DiskANN for AI App Builders
If you are building AI applications on PostgreSQL, the most practically useful feature in HorizonDB is pg_diskann. DiskANN delivers 1.5x higher query throughput compared to HNSW at comparable recall, with index sizes that run 5x smaller. For large vector collections — the kind you get with document embedding at any meaningful scale — 5x smaller indexes translate directly to lower storage costs and faster cold reads.
The filter-aware traversal is the less-talked-about part that matters most in real applications: filtered queries run up to 3x faster because DiskANN tracks which vectors match your metadata filters during graph traversal rather than filtering after retrieval.
HorizonDB also ships hybrid search via the azure_ai extension’s ai.search() function, which combines vector similarity and full-text BM25 search with Reciprocal Rank Fusion. One query, one database — no Elasticsearch sidecar required.
-- Hybrid search: semantic + keyword in one HorizonDB query
SELECT id, title, content
FROM articles
ORDER BY ai.search(
query => 'PostgreSQL vector search performance',
embeddings_column => 'embedding',
full_text_column => 'content'
)
LIMIT 10;
Fabric HorizonDB: The Serverless Preview Worth Watching
Microsoft also introduced Fabric HorizonDB as a public preview — serverless PostgreSQL inside Microsoft Fabric with zero-ETL mirroring to OneLake. Applications use the standard PostgreSQL interface; Fabric handles the analytics and storage integration.
This distinction matters because Azure HorizonDB (the GA product) does not offer serverless compute. Fabric HorizonDB does — but only for teams already in the Fabric ecosystem, and only in preview.
The Gap: No Serverless Compute
This is the thing to know before committing. Amazon Aurora DSQL is serverless and scales to zero. AlloyDB Omni has a serverless option. HorizonDB requires provisioned compute — you pay for the vCores you provision even when query volume is low.
For teams running steady, production-grade transactional workloads, this is a minor concern. For teams with unpredictable traffic, test environments, or multi-tenant platforms where dozens of small databases sit idle most of the time, the cost model is less favorable than serverless alternatives. Microsoft has not announced a timeline for serverless compute in core Azure HorizonDB.
Who Should Move Now
HorizonDB makes sense today if you are already on Azure, building AI applications that need vector search, or migrating a SQL Server workload to PostgreSQL and want a managed landing zone. The Apache AGE graph support also makes it one of the few managed PostgreSQL services that handles relational, vector, and graph data in a single engine — useful if you are building knowledge graph applications without wanting to run Neo4j separately.
If your workload needs serverless compute, runs on multiple clouds, or depends on custom PGXS extensions outside the supported ~75, wait or evaluate Aurora DSQL for now.
HorizonDB is now the right default for new Azure PostgreSQL workloads. The full Ignite announcement covers the migration path from Azure Database for PostgreSQL Flexible Server. With the serverless gap noted, you know exactly what you are choosing.

