A：PolarDB continues to break through in architectural innovation, not only innovating in hardware with innovations such as three-tier decoupling, but also, with the efforts of the database team, once again topping the TPC-C list, demonstrating good cost control and the ability to offer inclusive solutions to the huge computational power demand. At the same time, through a storage-compute separation architecture and built-in vector retrieval capabilities, it provides a one-stop platform supporting data storage, transaction processing, and AI inference, simplifying developers' workflow.

A：Driven by market demand, PolarDB has successfully served core systems in industries such as finance, government, and telecommunications through its integrated architecture and storage-compute separation, as well as a one-stop platform. This has proven its ability to support both TPAP and AI workloads simultaneously, making it the preferred path for enterprise intelligence. This success is attributed to the extensive trust and cooperation of customers, collectively driving improvements in product performance, cost, and usability.

A：The development path of AI is characterized by evolving from learning from and assisting humans to self-iterating and surpassing humans, leading to super artificial intelligence. In this process, the ability to collect and process data becomes crucial for expanding AI intelligence, with database systems playing a crucial role in storing and managing massive data and processing capabilities. Embedding AI large models into databases has become an inevitable choice to solve problems such as data movement, data sovereignty, and security, and to achieve an organic combination of hot data and model operators, thereby promoting real-time, high-quality data processing and intelligent interactions.

A：We need to undergo a fundamental paradigm shift, transforming data infrastructure from traditional business support systems to data engines that can maximize data value and serve as native entrances for AI. The core of the maturity of AI databases lies in deep integration, rather than simple feature addition. For example, PolarDB has launched a version with built-in AI large models, using model operator technology to embed them in the kernel. Developers can directly call AI to complete the entire closed loop, significantly improving performance and simplifying complex architectures.

A：We have done the following work: first, we have developed a multimodal database integrated storage solution for agent natives, combining hardware innovations such as CXL switch to build a unified memory pool, and upgrading polo FS to support integrated lake base. At the same time, we optimized model tuning rate and achieved the organic combination of multi-tenant elastic online inference services and transaction analysis by connecting the Hundred Chain platform and Pi.

A：We have upgraded the built-in metadata management capabilities of PolarDB, now supporting zero ETL lake synchronization service. This means that when any data source in the lake base changes, its original metadata information can be automatically synchronized in real time. Additionally, we also have the ability to handle heterogeneous data from multiple sources, and support seamless embedding, tagging, labeling, and feature extraction operations in metadata management.

A：We have enhanced the multi-mode data retrieval and processing capabilities of PolarDB, and further promoted model computation, making it possible to provide efficient inference services within the database, allowing model calls and token consumption to be more contextualized, thereby improving resource utilization efficiency.

A：We provide agent development and operation deployment support, especially focusing on key points such as long short-term memory, vectorized embedding, and feature extraction. We abstract these capabilities in PolarDB, supporting full-text indexing, graph indexing, and vector indexing, to help seamlessly support AI development in full-text data, and simplify the process of intelligent development through zero-code development.

A：PolarDB has evolved from a cloud-native database to an AI-ready cloud-native database, with key evolutionary features such as Lake base, multi-model operatorization, A-performance support and development, and hardware innovation breakthroughs. Currently, PolarDB has won the trust of over 20,000 customers, with a deployment scale of over 3 million cores, covering 86 global availability zones, and has received a series of recognitions globally, such as maintaining its leadership position in the Gartner Cloud DBMS Magic Quadrant for six consecutive years.

A：With the explosive growth of smart agents and tokens, enterprises will face challenges in managing and orchestrating smart agents, achieving synergy between human and smart agents, governance, security control, and cost control. At the same time, data volume will grow at a rate of 2% to 14% per year, with data no longer just being a byproduct but a core production factor. Enterprises need to reshape their business processes, including productivity tools, business processes, user experience, product services, ecosystem partners, decision-making patterns, employee organizations, and business models, to adapt to future development and establish intelligent data architecture to integrate information, deliver insights, and form a data-driven culture.

A：In the AI era, the role of databases is becoming more crucial and important. AI models do not store facts, but databases need to store facts, especially after the emergence of intelligent agents, databases have shifted from being read systems to action systems, and their responsibilities and roles have changed. In addition, the cost of AI failure is increasing, making the responsibility of the data layer even greater. The overall trend in the database market is growth, with relational databases taking the lead and moving towards AI-native. AI for databases will become a standard capability, enhancing the level of automation in database maintenance and management.

A：The development trends of the database market include the global market size increasing from over 90 billion dollars in 2024 to over 170 billion dollars in 2029, with an annual growth rate of 12.8%. The growth rate in the Chinese market is approximately 18.7%, and relational databases still dominate the majority of the overall market share. In terms of trends, databases are evolving towards AI native integration, non-relational and multi-modal databases will accelerate their rise, and cloud multi-cloud and open standards have become the main directions of evolution.

A：The six major capability requirements are: federal real-time data access capability, vector and multimodal semantic processing capability, real-time decision-making capability integrating things and analytics, autonomous operation and intelligent database management capability, event-driven and streaming response capability, and zero trust and autonomous data governance capability. Federal real-time data access is one of the core requirements, predicting that by 2027, 80% of agent AI applications will require real-time, contextual, ubiquitous data access. Enterprises need to solve the problem of data sharing between different departments by enabling data to be called in real-time by intelligent agents while existing locally in various departments through methods such as federated computing, thereby improving business collaboration efficiency, reducing organizational friction costs, and shortening decision-making chains.

A：The development path of PolarDB goes from cloud-native to AI-ready and then to AI-native. In the cloud-native stage, it mainly solves high-concurrency and massive data storage issues; in the AI-ready stage, it supports unified management of vector and multimodal data; finally, in the AI-native stage, it has certain self-made abilities, providing global data management and compatibility with AI application ecosystems, such as memory management and knowledge base management.

A：In the AI era, users of databases have shifted from developers to being predominantly agent-based, demanding a higher level of elasticity, linear scalability, and read/write synchronization scalability for data platforms across generations. The Limit List is our attempt to explore these capabilities.

A：We have upgraded the three-tier decoupled architecture comprehensively using CXL technology. By replacing the RDMA interconnection method with CXL, we have reduced the latency from microseconds to hundreds of nanoseconds, and have broken through the performance bottleneck of a single machine, with remote memory 16 times larger than local memory. At the same time, CXL connects GPU, enabling PolarDB to implement KV cache capability internally, helping accelerate inference internally in the database.

A：The distributed version of PolarDB has proposed an integrated architecture, helping users smoothly migrate from monolithic centralized applications to native enterprise-level applications, and has built a high availability data solution at the financial level, such as global multi-activity and off-site backups. This solution has been successfully implemented and verified at Industrial and Commercial Bank of China.

A：We innovatively proposed the concept of model operators, enabling the ability to call on large models within the database and enjoy the technological upgrades brought by AI throughout the entire data lifecycle. Currently, the model operators have been connected with Bailing.

A：PolarDB chooses to embrace the lake base architecture to break through data barriers and adapt to the more generalized and in-depth data requirements of the AI era. We will strive in two directions, architecture evolution and support for open formats, to achieve unified management of global unstructured, semi-structured, and structured data, as well as to provide capabilities for managing domain data sets, universal data management, and multi-mode data retrieval.

A：By embedding AI nodes internally in PolarDB and providing a fully managed Ray processing framework, we have achieved data processing and inference happening inside the database, enabling the use of large models throughout the entire data processing lifecycle.

A：In the future, PolarDB is committed to three developments: 1) Continuously integrating the most advanced large models under the guidance of data gravity to fully explore the value of data; 2) Becoming the foundation intelligent base or intelligent platform for all intelligent integrated application developments; 3) Making the interaction of PolarDB more natural, supporting language processing and multi-modal processing, attracting more ordinary users to use it.

A：It is not suitable because when processing the image-text pair in the bottom right corner, using only the tranquil operator or data segmentation operator cannot solve the problem of the visual model understanding the semantic meaning of image-text.

A：We need to adopt a model similar to VL visual understanding model to project text vectors, text features, and visual features onto the same vector plane, forming a mixed vector pattern to identify and process these data in order to avoid information loss and improve model effectiveness.

A：For each type of data form in the enterprise, dynamic combinations should be made based on different operators to build unique matching logic for the enterprise. With the database capability, various pipelines should be constructed to transform all types of data assets into data assets that are suitable for large models and have application value.

A：Long-term memory is crucial for solving the continuity reasoning calculation problems of context and agent in complex reasoning tasks. In the process of real-time dynamic updates and processing, long-term memory can enhance the accuracy of the model, and provide personalized recommendations based on user profiles and fragments. In the future, multimodal long-term memory will play a greater role in the interaction between intelligent hardware and the real physical world.

A：PolarDB has been committed to exploring the integration of software and hardware in the past year, including CCL pre-rehearsals, AMX technology applications, file storage improvements, etc., and has conducted in-depth research in the AI direction. Especially in the era of AI, facing the challenges of massive data processing, PolarDB is developing towards AI native direction, dedicated to solving four core issues: multi-mode data processing, efficient search, deep AI integration, and AI-assisted operation and maintenance.

A：PolarDB Lake Base is a product designed for the management of massive data, capable of supporting the storage and management of relational data and unstructured data (such as images). It consists of multiple systems, providing functions such as raw data management, object storage management, and data acceleration, to help users better manage and utilize massive data for training and inference work. This solution has been successfully applied in a large-scale data management scenario for a certain automotive enterprise.

A：The Polo Search product integrates various query capabilities, including vector queries, graph queries, and text queries, all of which are built-in and operated within the database. The benefit of this design is that data does not need to be migrated or moved externally for queries, achieving integrated operations. The built-in vector engine can utilize a large number of optimization techniques, such as CXO memory and HNSW algorithms, to improve query efficiency. Additionally, this engine can utilize hardware facilities like GPUs to accelerate index construction, which can be completed in just tens of minutes compared to potentially hours with a CPU, greatly enhancing performance.

A：Polo DB's text engine supports inverted index and efficient search, similar to electric search, also integrated as a built-in engine. The graph engine, as a type of old-generation database, used to be commonly used for risk processing in the past. With the development of AI, it has become increasingly important in many scenarios, used to support applications such as long-term memory.

A：Polo DB can query real-time data through its built-in engine, helping users to monitor and analyze public sentiment in real-time.

A：Starting from last year, Polo DB has been exploring the integration of various large and small models (such as BST models) into the database, so that customers can obtain real-time prediction results and utilize transaction and recharge data to determine user behavior trends.

A：Polo DB has the characteristics of low-level control capability, fast creation and deletion of multi-mode agencies. It can effectively deal with the challenges brought by the short life cycle of intelligent agents and high concurrency, and support KV cache to optimize reasoning efficiency.

A：Polo DB makes AI native efforts in four directions, including multimodal fusion, advanced search, built-in AI engine (pod model as Operator), and AI assistant. Among them, the AI assistant achieves functions such as cluster information query and performance analysis through natural language processing, simplifying operation and maintenance tasks.

A：Ideal car, due to the massive and rapidly growing data of intelligent electric vehicles, faced issues such as low development and operation efficiency of native databases, and doubled storage costs. After testing, Ideal car ultimately chose PolarDB. The architecture of separating computing and storage reduced costs by 60%, the native massive database architecture reduced operation costs by 50%, and supporting h-type capabilities increased operation efficiency three times. Ideal car solved data privacy and security issues in large-scale model inference services, as well as business response delays caused by GPU queueing, by integrating PolarDB's KV cache underlying technology and connecting to PolarDB AI inference model services. This resulted in a stability improvement of over 60%.

A：We have collaborated with PolarDB's super base to utilize its underlying PolarDB native database, high-performance storage, and AI data capabilities, integrating with the cloud to achieve 100% service stability, and improving development efficiency by over 100 times through web coding.

A：The main issue lies in intelligent data acquisition. In the traditional process, businesses submit requirements to the data warehouse and wait for hours to days to obtain data, which is clearly unable to meet the efficient demands of the AI era.

A：The main demands include: first, business personnel are not familiar with SQL writing and need to obtain the latest and hottest data in real-time from natural language; second, support data acquisition from various relational databases (such as MySQL and PostgreSQL) using multiple protocols.

A：We have integrated the data agent of PolarDB, which combines the advantages of small models and large models. The small model extracts key information, while the large model understands user semantics and performs NL to SQL optimization. After integrating enterprise internal knowledge data with a vector knowledge base, the accuracy can reach about 95%.

A：Combine with PolarDB's AI data services and data lake capabilities to manage, query, and use massive multimodal data; at the same time, collaborate in the field of data engineering to leverage PolarDB's technical support to facilitate internal developers in training, deploying, and fine-tuning data.

A：The theme shared by Lu Changqing is "PolarDB helps AI emotional companionship interactive new paradigm", focusing on how to upgrade AI companionship from simple replies to truly understanding users and providing instant companionship, including technical exploration and practical experience.

A：The key to the success of company X is building a customized emotional companionship for each individual, defining user profiles through the emotional companionship mental model, designing roles through Foster's transactional thinking model, and achieving genuine emotional interaction that makes users feel understood and cared for.

A：In order to solve the memory problem, in the x adopts a 4096-dimensional embedding model and cooperates with PolarDB to store billions of dialogue data using PolarDB MySQL, as well as achieving vector retrieval at the level of hundreds of terabytes using PolarSearch. In addition, it also combines Polar Urgent Memory to solve memory confusion issues, and improves response speed through KV cache technology, achieving the effect of instant response.

A：Intel believes that computing power has become the core driving force behind the development of the digital economy. With the continuous high-speed growth of domestic computers and the vigorous development of large-scale model technology and native AI applications, the demand for comprehensive performance of CPUs is becoming increasingly high. For this reason, Intel launched the Xeon 6 processor last year, integrating various engines to support efficient collaboration of multiple industry workloads and meet the evolving computing power requirements.

A：The powerful Intel Xeon 6 processor provides comprehensive computing power, effectively improving database processing capabilities. Through built-in AI security, network data analysis and storage engines, as well as comprehensive support for CXL 2.0 specifications, significant performance improvements have been achieved in AI model tagging, responsive retrieval, and general database benchmarking within the database. For example, with AMX technology support, embedding performance can be improved by 256%. In addition, the Xeon 6 processor also optimizes the performance of built-in AI models in the database, such as utilizing AMX instruction sets and optimization toolkits to achieve performance improvements of 24% to 86% for WDL recommendation system models at different batch sizes. Intel and Alibaba Cloud work together to accelerate data processing and AI innovation through the comprehensive capabilities of the Xeon 6 processor. In the data preparation stage, the Xeon 6 demonstrates powerful processing capabilities and integrates a series of machine learning and artificial intelligence algorithms. By optimizing the performance of AI cloud-native databases, Intel and Alibaba Cloud are committed to the research and development of innovative technologies and solutions, such as promoting the development of open connectivity technology for CXL, and ensuring that the Xeon 6 supports the CXL memory pooling architecture solution to meet diverse AI workload requirements in databases. At the Cloud Computing Conference at the end of September 2025, both parties witnessed the release of the Polo DB database dedicated server equipped with CXL 2.0 switch technology, which significantly enhances database scalability and performance.

A：As an enterprise-level database product, Polo DB distributed database always adheres to security and stability as its core, realizing enterprise-level database requirements through kernel-level security design, data anti-tampering mechanisms, and data loss prevention. On this basis, it continuously improves cost-effectiveness, high concurrency service scalability, and intelligent innovation requirements. Especially in the Industrial and Commercial Bank database transformation project, Polo DB distributed database uses cloud-native technology to build a new DB stack architecture, achieving resource pooling, containerized services, and general database resource scheduling, ensuring the separation of data and control layers to provide extremely high elasticity. At the same time, Polo DBX is also moving towards integration, supporting multiple loads such as transaction loads, data analysis loads, and AI inference loads to run on the same distributed architecture, and ensuring data security and high availability through full-chain auditing and multi-level rollback mechanisms.

A：In the 4G disaster recovery scenario, due to conflicts between the configuration of two data centers with four replicas and the Pockets consistency protocol, in order to solve the problems of large and small data centers, we introduced an arbitration control cluster. When a data center-level failure is detected, the control cluster will perform replica quantity downgrade operations to ensure RPU is 0. In this way, in large-scale data disaster recovery switchover scenarios, the same-city RTO can reach 2.5 seconds, and the average RTO for remote locations is 16 seconds.

A：PolarDB achieves optimization of all types of workloads through a unified storage engine and ultra-low-cost in-memory engine. It has significantly improved the read performance of the caching system, achieving a 12-fold increase, and achieved sub-second real-time data transformation and synchronization latency. At the same time, to address the issue of rising storage costs, it has implemented maximum cost savings through online data compression and historical data archiving mechanisms. The archiving process is fully automatic, supporting partitioning and row-level TTL configuration.

A：For enterprise-level application intelligence empowerment, the key lies in providing deterministic and auditable operational processes. Currently, most agents operate in reactive mode, which may result in accumulated errors during execution. Databases in enterprise-level applications should provide precise factual data and accurate knowledge retrieval to meet the needs of the intelligent agent construction process. PolarDB's distributed computing storage layer introduces technologies such as vector indexing and full-text indexing, and enhances analytical acceleration and mixed retrieval capabilities in transaction systems to provide precise contextual information.

A：In the intelligent body construction process, PolarDB plays a role in building precise context, ensuring accurate knowledge search and ML training automation planning and deep search through the use of long short-term memory mechanisms. It integrates large-scale expert-level experience in the cloud to form an operational intelligent body, with functions such as partition index recommendation, distributed deadlock detection, availability fault diagnosis, SQL performance insight, etc. It can build knowledge graphs based on expert experience, continuously reflect and iterate, and ultimately call on large models to provide analysis or operational recommendations.

A：PolarDB x agent uses rag to implement intelligent question answering, can generate partitioning table structures and recommend the best indexes, and supports context-aware table structure evolution. For example, it can be easily transformed into a TTL table partitioned by day, and can provide advice on advanced functions such as enabling column store engines and suitable scenarios. In addition, it also has intelligent diagnostic capabilities, can locate and analyze the root cause based on error reporting, or call diagnostic tools to analyze the root cause of high availability switches.

A：Industrial and Commercial Bank of China (ICBC) faces major challenges in database transformation, including ensuring service continuity, data security (such as preventing financial losses), large-scale complex transformation, building operation and maintenance automation platforms, controlling learning costs, implementing rapid switching and fallback mechanisms, and meeting strict production operation and maintenance requirements. ICBC has adopted two main strategies to achieve its database transformation goals: first, leveraging its unique advantages such as strict production operation and maintenance requirements, large-scale operational practices, and deep application design, research and development, and operation accumulation; second, collaborating with leading technology companies such as Alibaba Cloud to focus on core technology breakthroughs, accelerate the realization of independent controllability and innovation breakthroughs.

A：Industrial and Commercial Bank of China has partnered with Aliyun to achieve driver-level high availability connection in terms of high availability, reducing the intermediate load balancing link; in terms of security, it has adjusted the password error lockout policy to only lock out the user who initiated the request, avoiding comprehensive impact, and increased the upper limit restriction on the recycle bin function to ensure rapid data recovery while preventing space constraints. In addition, they have jointly created a homemade operation and maintenance platform and DB stack center solution, improving full chain tracking, AWR performance report analysis capabilities, and automatic kill function for large transactions in MySQL.

A：By integrating the product with the industry's automation platform, we have achieved processes such as data migration, conversion, and cutting flow, and integrated the cutting flow of the business system itself. This manages the cutting flow of all business systems as a whole, enhancing automation and standardization capabilities. Currently, it is possible to achieve concurrent cutting flow of 300 nodes within 20 minutes in control.

A：We established a data replication link from the new database to the old database through reverse copying, and integrated with the industry automation platform, achieving the management of 300 shards on a single control panel, and completing the switching and rollback effects in 20 minutes.

A：Our tool system mainly focuses on two aspects: first, reducing research and development costs, such as optimizing DDL table structure, SQL writing and testing, etc.; second, enhancing automatic operation and maintenance capability, ensuring operation and maintenance work can handle a large number of batch tasks.

A：In the future, we will continue to enhance the carrying capacity of core applications, such as decoupling of layers, isolation of exceptions, elastic scaling, etc., and strengthen security construction, including storage encryption, desensitization, and vulnerability protection. At the same time, we hope to further promote the landing application of AI technology in the field of financial-grade distributed databases through cooperation, in order to achieve proactive prediction and resolution of potential issues, and help drive the high-quality transformation and development of the financial industry.

A：At Go To Group, PolarDB is used in credit products, covering a variety of internal and external scenarios within the group. With the growth of user base, the business volume is increasing at a double-digit or even triple-digit rate each year. During the migration process from overseas cloud service providers to Alibaba Cloud using PolarDB, there was a significant improvement in performance and stability, while reducing costs by around 50%. The features of PolarDB, such as storage and computation separation, automatic scaling of read nodes, greatly simplify database management and operation, allowing the team to focus more on business innovation and value creation.