On November 17, 2026, Satya Nadella told Ignite that Azure HorizonDB, Microsoft's distributed PostgreSQL service, is now generally available. He gave two claims: up to 8X the performance of self-managed Postgres, and a plan to bring HorizonDB into Fabric so agents can "reason over live transactions alongside petabytes of historical data in real time."
Both are worth testing before you move anything. This post sets out what Microsoft said, what earlier coverage reported, and what to check.
What did Microsoft announce about HorizonDB?
From Nadella's X thread during the keynote:
HorizonDB slide from Microsoft Ignite 2026
HorizonDB, as shown in Nadella's X thread.
- HorizonDB is "now GA in Azure," delivering "up to 8X the performance of self-managed Postgres."
- Bringing HorizonDB into Fabric means agents "will be able to reason over live transactions alongside petabytes of historical data in real time."
- Microsoft is "rearchitecting the core compute, storage, and networking primitives in Azure" to run, connect and share context at agent scale.
The Ignite live blog we could read did not mention HorizonDB, and we found no GA announcement beyond Nadella's thread. Treat the details below as the claim plus background, and confirm on Azure's own pages.
What is the claim, and what is it measured against?
"Up to 8X" against "self-managed Postgres" is a best-case figure against a baseline Microsoft chose. A self-managed instance can be sized and tuned many ways, so the comparison depends on the hardware, the workload and the tuning.
There is a useful reference point. In August, InfoWorld reported that Microsoft said HorizonDB could deliver "up to three times the throughput of open-source PostgreSQL." The same piece noted that Microsoft made no comparison with Aurora or AlloyDB, and that analysts called the advantages "largely on paper." The jump from 3X to 8X may reflect GA improvements or a different test. Microsoft has not said which, in the material we reviewed.
What is HorizonDB, technically?
Per InfoWorld, HorizonDB is a cloud-native, Postgres-compatible service with disaggregated compute and storage and a database-as-log design. Native vector search and tight integration with Foundry and Fabric are its pitch for AI workloads. Microsoft first showed it as a preview at Ignite 2025.
It sits beside, not on top of, Azure Database for PostgreSQL. Microsoft's own PostgreSQL material keeps that service for lift-and-shift and standard open-source workloads.
How does it compare with Aurora and AlloyDB?
| Service | Maker | Status | Notes |
|---|---|---|---|
| HorizonDB | Microsoft | GA per Nadella, Nov 17, 2026 | Disaggregated storage; Fabric and Foundry ties |
| Aurora PostgreSQL | AWS | GA since 2017 | Vector search and Bedrock integration |
| AlloyDB | Launched 2022 | AlloyDB AI | |
| Lakebase | Databricks | GA on AWS and Azure | Postgres on the lakehouse |
| Snowflake Postgres | Snowflake | GA February 2026 | Postgres inside Snowflake |
Statuses for the competitors are from InfoWorld's August comparison. HorizonDB's own GA row rests on Nadella's announcement. We found no independent benchmark of HorizonDB against any of them.
What does "agents reason over live transactions" actually need?
The agent claim is the more interesting one, and the vaguer. Agents that answer questions about the business need two things at once: fresh operational data and long history. Today that usually means a transactional database plus a warehouse, with a pipeline between them and a delay.
Microsoft's pitch is that Fabric will query both without the delay. If it works, an agent can answer "what did this customer order this morning compared with the last three years?" from one place. Questions to put to Microsoft:
- Which of this ships at GA, and which is planned? Nadella's wording is future tense.
- Is the live path read-only for agents, and what stops an agent from writing to the transactional store?
- What is the latency from commit to visibility in Fabric?
- What does it cost to keep petabytes of history queryable?
On the safety side, a database that agents can read live is also a database an agent can leak. If coding or business agents query it, record what they ran. AgentBeam, built by the explainx.ai team, hooks into coding agents on developer machines and records their tool calls, which covers the developer side of that risk. It does not secure the database itself.
What are the known limits?
These come from August's preview-period reporting and may have changed at GA. Check each:
- Provisioned compute. InfoWorld said HorizonDB uses provisioned instances, not serverless scale-to-zero, so you pay while idle.
- Extensions. Only an approved set of PostgreSQL extensions is supported. An app that depends on an unsupported extension may not move cleanly.
- Regions. The preview ran in a handful of regions (Central US, West US 3, UK South, Australia East). We have not seen the GA list.
- Pricing. We found no GA price figures.
Why does Postgres keep showing up in agent plans?
Postgres is the default store behind a lot of agent products. It holds users, orders and tickets, and with a vector extension it also holds embeddings. Teams like it because the agent, the app and the analyst can all use the same SQL. That is why every large cloud now sells its own managed variant, and why each one pitches "for AI."
For a builder, the practical effect is small and concrete. A faster Postgres lowers the cost of the retrieval step that runs on every agent turn. A Postgres that Fabric can read directly could remove a copy job. Neither change alters what your agent can do. It changes what each answer costs and how stale the data is.
How should you benchmark the 8X claim?
A fair test takes a day, not a quarter. Use your own data and your own queries, because a vendor benchmark cannot know your access pattern.
- Pick the baseline honestly. Compare against the self-managed or managed Postgres you run now, at a similar price per month, not at the same instance size.
- Replay real traffic. Capture a day of production queries, including the slow ones, and replay them at your peak rate.
- Measure the tail. Report p50, p95 and p99 latency. A mean hides the queries that make agents time out.
- Test writes and reads together. Agent workloads mix lookups, inserts of memory or logs, and vector queries. A read-only test flatters any system.
- Kill an instance. Measure failover time and what the application sees. This is where disaggregated storage should help, and where you should verify it.
- Compute cost per thousand queries, not cost per hour. A system that is 3X faster on a 2X price is a 1.5X gain, not a 3X one.
If the result is a gain on your workload, it is real for you. If it only appears on Microsoft's benchmark, it is marketing.
A worked example: an agent that reads orders and history
Suppose a retailer runs a support agent. Today it calls a Postgres database for the customer's open orders and a warehouse for the last three years of purchases. The warehouse copy is refreshed overnight, so the agent sometimes tells a customer something that is a day out of date.
Under Microsoft's pitch, the transactional data lives in HorizonDB and Fabric reads it live next to the historical tables. The agent asks one question and gets one consistent answer. The gain is freshness and one fewer pipeline to run.
The risks are also concrete. The agent now has a path to live transactional data, so you need read-only access, row-level limits and a log of every query it ran. A bad prompt or a poisoned document should not be able to turn a "show my orders" request into a bulk export. Before any such pilot, write down which tables the agent may read, which columns are off limits, and who reviews the query log.
What should you do this week?
- Do not migrate on the headline. Ask Microsoft for the benchmark behind 8X: dataset, instance sizes, tuning and the self-managed baseline.
- List your extensions. Compare against HorizonDB's approved list before any pilot.
- Pilot on a copy. Replay real query traffic and measure p95 latency, failover time and cost per month.
- Treat Fabric integration as a roadmap item until a document shows what ships today.
- If you are on AWS or Google Cloud, nothing here forces a move. The mature options still work.
- If you are deep in Azure and Fabric, pilot it for a new agent workload rather than a core system of record.
What is known and what is not
| Known | Not known |
|---|---|
| Nadella announced GA on Nov 17, 2026 | Region list and pricing at GA |
| Claim: up to 8X self-managed Postgres | The benchmark behind it |
| Fabric integration is planned so agents can read live and historical data | What ships today in Fabric |
| August preview used provisioned compute and approved extensions | Whether those limits still hold |
Information on HorizonDB is from Satya Nadella's Ignite X thread on November 17, 2026, and from InfoWorld's August 2026 analysis. Details may change as Microsoft publishes documentation. AgentBeam is built by the explainx.ai team.
