AzureFeeds

Call AI models directly from SQL, build durable vector pipelines inside the database, and deliver high-accuracy similarity search at massive scale with DiskANN and AI re-ranking, all without leaving PostgreSQL. 

Debug and optimize queries faster with the Azure HorizonDB VS Code extension. Visualize execution plans, let Copilot generate fixes, and clone production data to test environments in seconds. 

Charles Feddersen, PostgreSQL Partner Director PM, shares how to put all of it to work on Azure.

Same hardware, same zone, same PostgreSQL version.

4,200 TPS self-managed vs. 11,000+ on Azure HorizonDB. The separation of storage and compute layers make the difference. See how it works.

Chunking. Embeddings. Vector storage. 

All running as a durable background task inside PostgreSQL with AI Pipelines in Azure HorizonDB, no external orchestration needed. Check it out.

Visual query execution plans. Copilot-generated fixes. 

The Azure HorizonDB VS Code extension brings all of it into the editor. Get it now.

QUICK LINKS: 

00:00 — Azure HorizonDB features 

00:57 — Open-source PostgreSQL 

02:24 — How it works 

03:37 — Performance 

04:51 — Enterprise-ready security 

05:34 — Memory & storage work together 

06:29 — AI Model Management + AI Functions 

08:24 — AI Pipelines 

09:50 — DiskANN + AI Re-ranking 

10:50 — VS Code Extension + Data Cloning 

12:31 — Wrap up

Link References 

Check out our blog at https://aka.ms/azurepostgresblog

Unfamiliar with Microsoft Mechanics? 

As Microsoft’s official video series for IT, you can watch and share valuable content and demos of current and upcoming tech from the people who build it at Microsoft. 

• Subscribe to our YouTube: https://www.youtube.com/c/MicrosoftMechanicsSeries


• Talk with other IT Pros, join us on the Microsoft Tech Community: https://techcommunity.microsoft.com/t5/microsoft-mechanics-blog/bg-p/MicrosoftMechanicsBlog


• Watch or listen from anywhere, subscribe to our podcast: https://microsoftmechanics.libsyn.com/podcast 

Keep getting this insider knowledge, join us on social: 

• Follow us on Twitter: https://twitter.com/MSFTMechanics 


• Share knowledge on LinkedIn: https://www.linkedin.com/company/microsoft-mechanics/


• Enjoy us on Instagram: https://www.instagram.com/msftmechanics/


• Loosen up with us on TikTok: https://www.tiktok.com/@msftmechanics

Video Transcript:

– The Postgres you know and love is now even more supercharged on Azure with the latest platform optimizations and AI for enterprise apps all made possible with the new cloud-native Postgres service, Azure HorizonDB. Now, we’ll go deep on its ultra fast performance at high scale, the built-in resiliency across different availability zones, along with major new built-in AI-centric features, leveraging DiskANN, as well as integrated AI model management, where you can provision models or bring your own from Microsoft Foundry, plus a built-in AI pipeline that automatically chunks and processes data in real time as it’s ingested into the database. And finally, a brand new VS code extension with AI-powered query development and debugging as you build your apps and more. And today I’m joined once again by Charles Feddersen who leads the Postgres efforts on Azure. Welcome back to the show.

– Thanks Jeremy, it’s good to be back and we’ve got a lot to get through today.

– Yeah, so the new HorizonDB service in Azure is our newest managed Postgres service where we also have Azure database for Postgres, and we’re actively making deep contributions in the Postgres project as well.

– Yeah, so Microsoft has been actively supporting Postgres on Azure for about nine years now, and our approach has always been to keep the Postgres, you know, and love for its portability and ecosystem and make Azure the best place to run your Postgres workloads. You know, and as you mentioned, we are a major contributor to open source Postgres. In Postgres 19, Microsoft Committers modified over 64,000 lines of code, which represents about 8% of all changes in this version. And we made about 340 commits. These improvements cover both new functionality and also improvements based on our own learnings of running Postgres at a massive scale on Azure. We also host the largest virtual Postgres community conference in the world called POSETTE, which is now in its fifth year. Because we understand the core Postgres engine so well, it’s not unusual for our core committers on the team to work with our customers to help debug and resolve issues that they might be having as well.

– Right, and I remember last time you were on, we discussed how you were shipping major versions of Postgres on Azure within weeks of their releases.

– Yeah, and now we’ve effectively eliminated that delay. There is now zero cloud lag in waiting to leverage the latest features when using Postgres on Azure. We ship the new major version on the same day in Azure.

– So that’s a big deal, there’s no lag, same day access. Why don’t we move back though to our newest managed Postgres service HorizonDB. How does that then change the game for anyone who’s running Postgres services now on Azure?

– Yeah, so HorizonDB brings cloud scale performance built-in resilience and AI ready capabilities to Postgres without changing how you build or run your apps. How this works is we’ve completely decoupled compute and storage to improve performance, scale and availability. At the compute layer, it runs the Postgres engine, fully compatible so existing apps and tools just work. At the storage layer, we built a new and highly optimized log service for transactions where we can sustain a commit latency of typically under one millisecond. That log service is paired to a new shared storage platform, which natively stores data across availability zones for resilient storage by default. We can attach multiple read replicas to the storage and the primary can fail over to any of these replicas in less than five seconds, in the event of an outage. This gives you read scale without replication latency. Ultimately, HorizonDB lets you run any Postgres workload from new AI apps to large scale enterprise systems with the performance, resilience and scale of Azure already built in.

– All right, so you said performance. Let’s dig deeper there. Everybody loves a performance story, so can you prove it?

– Yeah, so one of the big challenges with self-managed Postgres is that enabling high availability, especially across cloud zones, often comes with a performance cost. With HorizonDB, we introduced a new quorum commit protocol that let you durably commit across zones before flushing to disk, so you get both resilience and high performance. Now to show this in action, I’ve got a split screen, HorizonDB on the right and self-managed Postgres on the left. This is the same Postgres version on the same hardware with the same high availability setup in the same Azure region. And I’ll kick both off and then go ahead and let them run. And as you can see clearly HorizonDB is delivering about three times more transactions per second. This gain comes directly from the storage architecture, and now that it’s finished, we can see that self-managed Postgres on the left delivered over 4,200 transactions per second. And HorizonDB had more than 11,000 with much lower latency. Performance was a core design principle for HorizonDB from the beginning, and it’s something that you’ll see directly in your applications.

– Okay, so performance is ultra fast, and we know this is built for both enterprises and developers, and we know enterprises love security. So what’s different there?

– Yeah, so security of course is foundational. It spans everything from network isolation and identity to data protection. Just like performance, it’s been a core focus from day one. And we’re building on the proven enterprise capabilities of Azure database for Postgres to deliver it out of the box like Entra ID integration, where you can enable identity-based access, so users connect securely without managing passwords or private endpoints to lock down access. So the database is only reachable over your private network with no public exposure and of course encryption at rest where data is automatically encrypted on the disk. All of this is available from day one, so you get strong enterprise ready security without any extra setup.

– Okay, and bringing this back to our developers who are watching the built-in AI centric features with HorizonDB now also make it easier to build AI apps.

– Yeah, and this is an area that we’re really leaning into. Postgres was already popular and chat with your data use cases accelerated that adoption because it natively supports vectors for similarity search. Our focus is on making intelligent retrieval and generative AI work on a massive scale with high accuracy and efficiency. To do that, we’ve rethought how memory and storage work together in the database so that more IO traffic moves to disk letting you take advantage of much larger storage capacity using Microsoft’s disk accelerated nearest neighbor or DiskANN technology. This quantize a vector-based graph in memory and maps it to a full precision graph stored on disk, which significantly reduces memory requirements while still delivering fast higher quality similarity search across your data.

– And speaking of generative AI responses, we’ve also made it easier for the database to work directly with AI models.

– Yeah, we have, and this starts with AI model management, which automatically registers a set of models with your instance of HorizonDB. It’s really simple. So here’s my HorizonDB, and I’ll just select the enable managed models then confirm. And what this does behind the scenes is it automatically registers a few AI models for use. Once the provisioning is complete, you can see the AI model management blade in the portal with the GPT, text embedding and re-ranking models listed. And you can also bring your own models where you register them through the database directly.

– So now you can use models effectively with Postgres without extra manual provisioning configuration. So, how would you interact with them?

– Yeah, this is where AI functions come in. These are a set of SQL functions that you interact directly with AI models. The Azure AI extension in HorizonDB provides functions that you can invoke using SQL to leverage AI models. So let me show you a couple. For example, you can use the generate function to produce a response from a prompt. In this case, I’ll say summarize the following customer reviews in two to three sentences, and then I’ll run it and you can see the response is generated as a SQL result that I could use in my app. Alternatively, the extract function is designed to pull specific entities from unstructured text into a structured form that you can then store and query. You can see here that I’m going to use an array to ask for the main product features along with the customer sentiment from my database. Now, we’ll let that run for a moment and we can see the response of producers for each item in the catalog.

– And this is just one scenario that you showed with querying, but I can imagine another case where you might say, use data transformation where the results are written back into the database.

– Yeah, absolutely. And when you do that, obviously the retrieval time for those results stored in the database is incredibly fast and you’re saving on your tokens as well.

– That’s really great to see. So another challenge that we hear a lot of times is around building and maintaining AI pipeline. So we’re making that easier too.

– So this is where AI pipelines in HorizonDB extend AI model management and AI functions even further. Today, most generative AI apps rebuild the same steps, chunking, creating embeddings, backfilling data, often in fragile external pipelines. With AI pipelines, all of this is built directly into Postgres. So you can see here that I’ve created a pipeline using the create pipeline function in the AI extension. This chunks the data in a database in real time as it’s added. It will then create embeddings for those chunks using the embedding model that the AI model management enabled for us. And these are ultimately stored in a table. So I can then go ahead and run this, and then if I count the records in the output table, you can see that it’s increasing. And this is all happening asynchronously in the background so that it’s not blocking or really having any real performance impact on my transactional workload. But the best thing about these pipelines is that they’re durable. And this means that I could pause it. And you can see here that the row count stops increasing, even if the workload continues and I can resume it, think of it as a reliable background task. And if my server fails over, it’s no problem, the AI pipeline fails over as well and it just keeps running.

– So all that pipeline complexity then just moves into the database and kind of reduces the complexity and runs reliably.

– Exactly, and building on this, we can now also use the rank function to re-rank search results with an AI model. Earlier I talked about how DiskANN improved similarity search, but that’s just the first step. With rank, we can apply a model like Cohere to refine the top results and improve overall relevance. Let me show you, I’ve got two identical queries here. One just uses an index and the other wraps the same query in our rank function to improve the relevance of results. If I run the first query for the headphones with the highest playtime and good calling, you can see the results seem pretty good. And these aren’t bad with a few showing around 40 hours. But when I run the rank query that applies the AI ranking model, the top most results are definitely more relevant to my search. Here, there are headphones with 60 or more hours, and the top ranked model has a hundred.

– And this is really powerful. So you basically started out with fast similarity search, then used AI to make the results even more relevant. But why don’t we move beyond the database itself and look at what we’ve done then to help with the building experience of Postgres workloads on HorizonDB?

– Sure, and this is one of my favorite additions. We’ve built a new VS code extension that brings familiar Postgres tools into a modern AI powered experience. As someone who spent years working in database tools, this is something that I really wish I’d had. It makes debugging and development a lot more easy. So here we’re in VS code and on the left you can see I’m using the Postgres extension. Now this works with any Postgres, but when you connect to Postgres on Azure, you get even deeper integration. So let’s look at an Azure server, and if I right click, you can see a number of server management features like network configuration, backups, and even start stop built in with a single click. And if I right click on tables, one of the most recent additions is object properties. But now let’s run a query. This is a little slower than we’d expect. So I open the new visual query execution plans to debug. It’s easy to identify the Inefficient query operator, and I can click on that to debug further. But the best part is that Copilot can fix it for me. Here, I’ll just click on the Copilot button in the query plan and it gets to work. The Copilot generates the fix for me, and it’s really that simple. Now I want to test this fix in a non-prod server, and that’s easy to. For that, I’ll show you a first look at the new built-in data cloning. I’ll go back up to the server and right click and I can clone it with all my data. And that just takes a moment to validate my solution.

– And that’s a real shift. You got the same familiar Postgres experience, but now with AI and platform capabilities that really improve how you build and run apps. So what’s next?

– Yeah, so we’ve covered a lot today, but this is just the start. We’re continuing to push the boundaries of what you can do with Postgres on Azure. So here’s a lot more coming.

– So what’s the best way then for everyone watching to get started with Azure HorizonDB?

– Yeah, so, you know, you can go try it for yourself. It’s in preview now and it’s easy to get started with everything that I demoed. The VS code extension is available in the VS code marketplace. And of course, to stay current, check out our blog at aka.ms/azurepostgresblog.

– Thanks so much for joining us today, Charles, and of course, keep checking back to Microsoft Mechanics for the latest tech updates, and we’ll see you again next time.