- Subject
- Microsoft 365 & SharePoint
- Filed
- Source
- Robert Davis
- Loans
- 355
Microsoft Fabric HorizonDB Adds Serverless PostgreSQL for Intermittent Workloads
Fabric HorizonDB pauses compute after eligible inactivity and joins operational data with OneLake analytics. Teams should assess latency, freshness, cost and portability.

Microsoft Fabric HorizonDB is aimed at a practical problem for application teams: small or intermittent workloads can need a database without needing dedicated compute running all the time. The new serverless PostgreSQL offering, launched in public preview at this year’s Ignite conference, provisions and scales databases on demand, then pauses them when idle. It also gives applications a way to query transactional data alongside analytical data in OneLake through a PostgreSQL interface.
That combination could reduce the infrastructure and integration work behind internal tools and agent-driven applications. It also comes with trade-offs that teams should test before using it for systems that need predictable response times or strictly current analytical data.
Serverless PostgreSQL inside Fabric
Microsoft first introduced HorizonDB for Azure at its Ignite conference in November 2025. That offering was positioned for demanding enterprise workloads, including high-performance, transaction-heavy applications. Microsoft describes its architecture as shared storage paired with elastic, scale-out compute.
Fabric HorizonDB targets a different workload profile: applications that need a managed database but don’t require continuously running infrastructure. Microsoft says the service automatically provisions and scales databases as demand changes, pausing them when they’re idle and resuming them when needed. The company describes the new Fabric version as serverless PostgreSQL.
For teams already building on Fabric, the data path is a key part of the offer. Applications can query live transactional data and analytical data in OneLake through one PostgreSQL interface. Microsoft says Fabric Data Warehouse’s distributed compute handles supported analytical parts of PostgreSQL queries across CPU and GPU resources, while the PostgreSQL engine continues to process transactions.
That design is intended to bring operational and analytical access closer together. A team might otherwise have to create and maintain a pipeline or connector to combine those data sources for an application. With the Fabric approach, Microsoft says, supported queries can draw on both within the same interface. The scope matters: analytical processing applies to supported portions of queries, and teams will need to establish whether their application’s query patterns fit.
Where teams could save engineering time
Independent technology consultant David Linthicum says the serverless model could help developers and AI agents create databases quickly for specific tasks and workflows. That can cut setup work when an application is short-lived or used only occasionally, reducing the time between identifying a need and building a working application.
HFS Research executive research leader Ashish Chaturvedi points to the integration work involved in combining operational and historical information. Previously, teams often had to build, secure and maintain a pipeline or connector to retrieve both. With the built-in capability, he says, an application could retrieve a customer’s current order and purchase history in one request.
For CIOs, fewer integration components could mean less engineering work to maintain and a simpler route from operational systems to analytics and AI, Linthicum says. Chaturvedi argues that developers could spend more time on applications and less on data pipelines, potentially shortening development cycles and helping organizations extend applications across business functions.
Combining current activity with historical patterns can also shape what an application or agent recommends. Chaturvedi gives the example of a sales application showing open customer issues alongside lifetime value. An agent could likewise consider current business information and past trends when suggesting an action. Those use cases depend on the data available to the query and, for analytical information, how fresh the OneLake copy is.
Idle compute can help with cost, but check the details
The serverless model may suit prototypes and internal tools that see sporadic use. Chaturvedi says Fabric HorizonDB pauses its compute after five consecutive minutes of eligible inactivity, which could reduce the cost of compute sitting unused. The five-minute condition is specific to eligible inactivity; teams should confirm how their usage patterns and billing work before estimating savings.
Database compute is only one part of operating an application. Stephanie Walter, practice leader of AI stack at HyperFrame Research, says CIOs will need visibility into costs for database compute, storage and analytical queries, including any charges that continue while database compute is paused. For estates with many small applications, understanding those separate cost areas will matter as much as the pause behavior.
Test resume time, freshness and portability
Pausing compute introduces a possible delay when a database resumes. Linthicum cautions that serverless services can have cold-start latency, scaling lag and less predictable performance than provisioned infrastructure. He recommends testing pause and resume behavior, concurrent requests and transaction-heavy workloads.
That may make the service a better fit for occasional workloads than for customer-facing systems with consistent response-time requirements. Chaturvedi says a brief wait after an idle period may be acceptable for an internal tool, while applications that must respond predictably need closer performance checks.
Analytical freshness is another design consideration. Chaturvedi says the analytical copy in OneLake updates with a slight delay. Queries that combine operational and analytical data may therefore suit applications that can tolerate information that is a few moments old; applications requiring strictly real-time consistency need to account for that delay.
PostgreSQL compatibility provides a familiar interface, but it doesn’t guarantee that an application will move easily to another platform. Linthicum cautions that dependencies on Fabric services, OneLake or translytical query capabilities could make migration harder. Before adopting the service, teams should identify which parts of the application rely on those Fabric-specific features and test the workload’s latency, concurrency, data freshness and costs. That will show whether the reduced setup work fits the application’s operating requirements.