What Apache AGE is, and why it matters here
Apache AGE is a PostgreSQL extension that adds graph storage and openCypher querying. Microsoft's documentation describes it as graph modeling and traversal through Apache AGE inside HorizonDB. Graph data lives next to relational tables, and both can take part in the same PostgreSQL transaction. Microsoft's example is a supply-chain assistant. It walks supplier-to-part-to-order relationships in the graph and then joins the result to a relational table that holds live order status.
What changed in the GA release
Microsoft says the optimizations target two areas.
Loading and refreshing graphs
- MERGE now finds existing paths with a hash-based lookup, and improved memory management frees temporary per-row allocations as it goes.
- The gains matter most when incoming records keep referring to the same entities. In Microsoft's example, new records keep pointing at the same suppliers and parts.
Reading and querying graphs
- Property-aware selectivity estimates help PostgreSQL choose better join orders and join methods.
- Eligible anchored undirected patterns get index-friendly handling that keeps join shapes efficient.
- Microsoft says both read optimizations are on by default.
There is also a compatibility claim. The AGE extension API is unchanged, so existing applications can run on HorizonDB and just work without code changes.
Microsoft's performance numbers
Microsoft compares HorizonDB with community Apache AGE and reports these maximums:
| Workload | Claimed gain | Basis |
|---|---|---|
| Simple reads (direct dependencies, linked entities, ownership) | up to 3,900x faster | LDBC SNB Interactive workload |
| Complex multi-hop reads | up to 512x faster | LDBC SNB Interactive workload |
| Graph builds and refreshes (MERGE, batch upserts) | up to 190x faster, about 6x lower peak memory | Internal Microsoft testing |
Treat these as vendor figures. They are "up to" ceilings, not expected averages. The published passages don't give the baseline configuration, data size, hardware or query-by-query results. They also don't say which graph patterns qualify for each optimization. Don't assume every AGE query or ingestion job will see numbers like these. Benchmark your own workload before you plan around them.
Availability and what is still preview
Azure HorizonDB is generally available in 13 Azure regions across the Americas, Europe, and Asia Pacific, with more regions coming online soon. The announcement doesn't list the regions, so check the current regional documentation before you plan a deployment.
Keep the GA performance work separate from the previews launched alongside it. Microsoft says AI Model Management and Knowledge Graphs are in Public Preview, and Database Agent for PostgreSQL is in Limited Public Preview, with access by request. The AGE speedups are not part of those previews.
Enabling AGE on a HorizonDB cluster
Microsoft Learn's AGE overview gives this procedure:
- In the Azure portal, open the HorizonDB cluster. Under Settings, select Parameters.
- Create a parameter group. Enable AGE in both
azure.extensionsandshared_preload_libraries. - Associate the parameter group with the cluster. The cluster restarts automatically to apply the change.
- Connect with
psqland runCREATE EXTENSION IF NOT EXISTS AGE CASCADE;. - Optionally confirm the install with
SELECT * FROM pg_extension WHERE extname = 'age';. - Add the AGE schema to your search path:
SET search_path=ag_catalog,"$user",public;.
Two common failure points are documented:
- If AGE isn't in
shared_preload_libraries, the first Cypher query fails with "unhandled cypher(cstring) function call error on first cypher query". - Don't run
LOAD 'age';once AGE is inshared_preload_libraries. The library loads at cluster startup, and running it manually causes a privilege error.
The Learn page I reviewed is dated 2026-11-11 and its related guidance carries preview labels. Use it as setup guidance, not as a GA statement for every AGE-related setting.
Practical takeaways
- Best fit: workloads that build or refresh graphs with MERGE or batch upserts, and relationship-heavy reads, such as dependency tracing or ownership lookups, that can use the new planning.
- Migration risk: Microsoft says the API is unchanged, so code changes should be minimal. Still test query plans and load times on a representative dataset.
- Indexing still matters: Microsoft's separate AGE performance guidance notes that by default, Apache AGE doesn't create indexes for newly created graphs. That page is labelled preview and predates this GA change, so check whether it still applies. It also recommends checking plans with EXPLAIN. Planner improvements don't remove the need to design indexes for your queries.
- Don't over-rotate to graphs: faster traversal helps apps that keep connected data beside operational records. It doesn't mean every workload should move to a graph model.
The big multipliers will draw attention, but the more durable value is probably the combination. Graph and relational data can share one PostgreSQL transaction, and existing AGE code is meant to run unchanged. Verify the claims against your own data before you commit.
References
- (Launched) Generally Available: Improved Apache AGE graph query and ingestion performance on Azure HorizonDB Azure Updates · 2026-11-17T20:00:32Z
- Graph Database Capabilities with Apache AGE Extension - Azure HorizonDB | Microsoft Learn learn.microsoft.com
- Apache AGE Performance Best Practices in Azure HorizonDB - Azure HorizonDB learn.microsoft.com