Hi, first of all, thank you for Con Fortitude team for providing an opportunity to present a topic on optimizing SQL server infrastructure for LLM workloads in hybrid environment. Thanks for joining. My name is Siva. I'm working as a principal database administrator in Paycor, which is now a Paychex company. In this talk, I will present about how SQL Server get ready for LLM workloads. Let's start the presentation. Yeah. The agenda of this talk is the LLM infrastructure challenges and legacy SQL server limitations, Azure SQL managed instance, and I link for real time replication, performance optimization, and global architecture patterns. Security, governance and real world implementation of EMI, future proofing your data infrastructure. This is the agenda for this presentation. Okay. The, what are the challenges we have right now for LLM infrastructure, that traditional SQL server. He's having a limitations, like a high latency data, access patterns, limited scalability for concurrent EA workloads, complex replication across environments, performance bottlenecks in real time scenarios. So the modern LLM requirements are subsecond response times per prompts, continuous data synchronization, elastic scaling per variable demand. Enterprise creates security and governance. So why legacy SQL architecture struggle with ai? In legacy deployments, we haven't designed for a unique demand of LLM workloads. Prompt engineering requires instant access to synchronized and courage should be optimized for data access across the distributed systems. So the gap between yay application velocity and database respond times create a critical bottleneck that impacts the user experience and limits of AI capabilities. To address these solutions, Microsoft provides a managed instance that is the. Built for hybrid AI in this hybrid cloud integration it helps to synchronize the data from on-premises to Azure without any latency. And this is all also have the built in governance and compliance and data protections. So this helps you. Increase your performance based on requirements or need. So to set up manage instance, the communication between your on-prem server to manage instance will be taken care through MI link that is called Managed Instance Link. This link. Take care of synchronization between Uber, Azure, and on-prem environment. The synchronization will happen based on log based synchronization, which is similar to Uber, always on ayc configuration. The latency would be near zero. So with this less latency, your Azure environment MI instance, is ready for your LLM access because it helps current and consistent data. So the advantage of MI it is have a performance optimization. You can have utilize query store insights that leverage historical query performance data to identify patterns, optimize execution plans, and eliminate bottlenecks before the impact of A LLM response times the other benefit memory control workload classification. So am I intelligently allocate resources based on workload, priority, ensuring prompts, serving queries, receive a dedicated memory and compute resources. The other benefit is concurrency optimization. Fine tune connection, polling parallelism settings and isolation levels to handle. Hundreds of simultaneous requests without degradation. So architecture pattern for a global prompt. Data sourcing organizations with a global operations need prompt data available across regions with consistent performance. To fulfill this requirement, the am I link enables multi-region replication to apology, where on premises SQL server feeds regional ministry instances. This architecture reduces cross regional latency while maintaining data consistency, allowing LLMs to serve. Users from the nearest data source, regardless of where the personal data resides, in simplest terms, you can set up multiple MI links across multiple regions from one source database. Key performance metrics for LLM workloads, target correl latency. The response time, concurrent request, it'll support multiple queries on the same time. And the availability. Microsoft will take care of H-E-A-D-R per Azure manager instance, and there is a less minimum delay to synchronize the data. These are the key performance metrics for LLM workloads. The other advantage with the manager instances is you can utilize it as a per hybrid data disaster recovery, as well as AI continuity. For example, your primary SQL server. That is your production environment. You have a setup for MI link with the MI instance so that the MI instance will help to synchronize the data and it act as a read access for your LLM workloads and as well as works as a d. Environment. So in, in case of disaster, you can utilize and you can fail over your right loads from on premises to Azure. So it acts as a both rules A, a access as well as Dr. The other advantage of MI is. It'll a hundred percent integrates with Azure synapse integration. So with this, you no need to have traditional ETLs. The integration analytics with Azure Synapse Link enables advanced augment generation scenarios. Pre aggregated and index frequently accessed the prompt context in synapse producing query complexity and improving response times. This separation of concerns allow transactional SQL workloads and analytical prompt preparation to operate independently, preventing resource contention, and optimizing each of its specific purpose. This is one of the example for a real world implementation. For example, you assume that there is a global bank presence across the world and they want to set up a LLM chat bot with the their customers. And the major data is an on-premises environment. In this case, you can use MI link and set up a managed instances across the regions so that the bot can access the local MI instance instead of everything hitting too, you source on premises data. So this helps the response times and reduces the latency. With this solution, you can reduce the prompts, prompt response latency, and it is compliance with the regular standards and it supports concurrent AA sessions, security and governance in a pipeline. This helps data classification you. You can use automated discovery labeling of since two data elements used in prom access control, role level and column level security and force across the replication. Each targets audit trials. You can set up a comprehensive logging all. AI driven data access for compliance and reporting encryption, always encrypted columns and TDE protected data in transit at rest. So these are the features you can use in your MI instance based on your requirements. So to tune over LLM, these are the things we need to follow first. Enable query store. This allows to capture all, all over queries, so it helps to fine tune if any queries taking longer time. Configure Resource Governor for AI workloads. So if you configure this one, you can separate ever prompt queries and other queries so that users can get a quick response, implement connection, polling best practices. You can size over buffer polls based on every requirement. Optimize index for read heavy patterns, like create a proper index based on your query. Which improves the response times. Monitor your MI link logs to ensure that everything is working in sync. So future programing, your AI data infrastructure. So the implementation first phase, you have to assess. Evaluate current SQL workload and identify yay integration points. Phase to optimize, implement MI replication and tune for low latency access scale. Expand multi-region architecture with integrated analytics. Phase four, innovate leverage hybrid infrastructure for the next generation AI applications. So the key takeaways is modernize without disruptions. MI link brings like CQL and modern AI requirements without requiring full migration. Optimize for real time courage, store and workload classification and tuning eliminates LLM bottlenecks. Scale intelligently. Hybrid architecture supports global workloads while maintaining enterprise governance. Thank you. Thank you for providing this opportunity.