2026 年 10 月 21 日至 22 日，由飞轮科技主办的 Apache Doris Summit 2026 线上技术峰会圆满落下帷幕。本次峰会以 “Powering Real-Time Intelligence for the Agentic Era” 为主题，成功吸引了全球超过 12 万开发者、数据架构师及 AI 技术专家的云端参与，共同见证数据基础设施从“实时分析”迈向“实时智能”的历史性变革。
本届峰会设立了主论坛、实时湖仓论坛以及数据智能论坛三大重磅板块，围绕 AI 多模湖仓、增量计算、Agentic Analytics、可观测性评估与企业级生产落地 等前沿方向，汇聚了来自互联网、金融、国央企、物流、自动驾驶、智能制造等领域的 34 名顶尖专家，为业界呈现了一场贯穿底层引擎创新与上层智能应用的科技盛宴。
升级与蜕变：Apache Doris 5.0 重磅演进
峰会在 Apache Doris PMC Chair 陈明雨题为《从实时分析到实时智能：Apache Doris 5.0 驱动 Agentic AI 时代》的主旨演讲中拉开帷幕。
回顾过去 12 个月，Apache Doris 保持着平均约 3 周迭代一个版本的高频演进节奏，共发布 17 个版本。目前社区累计拥有 760 余位贡献者、109 位 Committer 与 26 位 PMC 成员；GitHub Stars 突破 16,000，主干 PR 超过 4,400 个。开源活动覆盖从国内城市行到新加坡、印度、旧金山、悉尼及欧洲等海外阵地，Apache Doris 开源生态正在全球范围内加速繁荣。
在 AI 时代，数据消费者正从人扩展到 Agent，数据的价值链路也从传统的数据到洞察，演化为数据到上下文再到行动。面对并发量增长 10 到 100 倍且具备强 Ad-hoc 特征的 Agent 查询需求，即将于年底发布的 Apache Doris 5.0 确立了全新的三大演讲主线：
- 数据层（多模湖仓）：将结构化与非结构化数据纳入统一的分析体系。
- 计算层（增量计算）：使计算结果随数据变化实现实时更新。
- 交互层（Agent 式分析）：将 SQL、检索能力与大模型深度结合。
这三个方向的结合，构成了 Apache Doris 5.0 的全新定位：面向 Agent 的数据基础设施，从实时分析走向实时智能。
底层引擎突破与上层 Agentic AI 的深度融合
在 AI Agent 及大模型深度融入业务场景的背景下，数据基础设施面临数据多模态化、计算持续化、交互自然化、链路可观测化以及部署 Serverless 化等关键挑战。在主论坛中，多位技术专家围绕多模湖仓、增量计算以及 Agent 上下文交互三大演进主线，系统展示了 Apache Doris 与 SelectDB 的底层技术架构突破与工程实践。
多模湖仓：存储开放，计算统一
Apache Doris PMC Chair 陈明雨指出，在 AI 时代，多模态数据的统一融合，是 Agent 能够感知真实世界的基石。
在主论坛演讲中，Apache Doris 联合快手、 LanceDB，正式公告了 Apache Doris 5.0 面向 AI 的多模湖仓架构。面对 AI Agent 与大模型应用对数据层提出的“完整、统一、实时”极速上下文需求，Apache Doris 给出 “存储保持开放，查询与执行保持统一” 的全新解法。
换言之，无需将所有数据搬迁至单一存储系统，允许数据保留在各自的开放格式中，而在查询与执行层实现统一。
- 存储开放，按负载解耦：无需搬迁数据，允许 Iceberg、Paimon、Lance 等开放格式各自承载长期分级、持续更新或多模态数据集；Doris 引擎直接复用湖表原生的向量、全文及标量索引，避免冗余存储与同步开销。
- 计算统一，单条 SQL 混合检索：打破结构化与非结构化数据壁垒，通过单条 SQL 在同一个执行计划中完成“向量语义召回 + Variant 动态路径过滤 + 跨源多表 Join + 权限过滤”，从根本上解决传统架构“相似但不可用”的痛点。
- 秒级实时，闭环写回：支持通过 MERGE INTO 等语法将 Agent 产生的标注重构原子写回 Paimon 湖表；借助全新的 Fluss Union Read（流湖一体直读），无缝合成历史快照与秒级实时日志，提供最新的分析上下文。
此外，快手现场分享了基于 Paimon 存储上亿条高维向量并用 Doris 直接检索的生产实践，LanceDB 也展示了多模态 AI 数据集与 Doris 引擎深度集成的最新成果。阿里云产品专家凤豪也分享了 SelectDB 多模湖仓架构在智驾场景的标杆落地，通过对 Paimon 数据湖中 1.28 亿级多模态数据直接检索，实现高达 96% 的向量召回率，性能衰减仅为内表的 1.07 至 1.36 倍。
增量计算：从批量刷新走向持续计算
Apache Doris PMC、飞轮科技存储引擎技术负责人周飞表示，数据越新鲜，Agent 的推理与决策就越贴近真实世界。未来的数据计算不再是定时的批量刷新，而是随着数据流入而发生的持续演进。
为了保证 Agent 在推理过程中获取最新的业务上下文，Apache Doris 5.0 在计算层推出了全新的增量计算架构与异步物化视图行级增量刷新能力（IVM） ，改变了以往依赖定时任务全量或微批刷新的处理逻辑。
底层创新：精准捕获与有状态流控
Doris 5.0 在底层推出了 Row Binlog 与 Table Stream 技术，不仅能精准记录更新前后（Update Before / Update After）的行级数据变更，更实现了有状态的消费位点管理与事务保障。既可通过 Flink Connector 将秒级变更实时推送到外部流计算生态，也可在引擎内部直接驱动计算。
引擎突破：单条 SQL 实现秒级行级刷新
基于 Row Binlog，新一代 IVM（行级增量物化视图） 允许用户通过声明式 SQL 自动维护复杂业务指标。测试显示，在 10 亿级基表数据下，IVM 仅扫描发生变更的数百行数据，在仅消耗传统分区刷新 1/50 资源的前提下，将数据变更至指标查询的 P50 延迟缩短至 0.26 秒，耗时与 CPU 消耗仅为全量刷新的数百分之一。
Apache Doris PMC、飞轮科技查询优化器负责人张文歆补充到，截止目前，IVM 已在电商实时数仓等场景成功落地，实现了发货 GMV 宽表自动拼装与公共维表高效统一加工。此外，Doris 正在将 IVM 与增量计算能力进一步延伸至 Iceberg、Paimon 等开放湖仓格式中，大幅降低数据流水线的维护成本，为企业级 AI Agent 与实时业务决策提供极致新鲜、低成本的数据底座。
Agentic Analytics：Capy Agent 发布，企业级分析新范式
Apache Doris 与 SelectDB 创始人马如悦以“对话你的数据（Talk to Your Data）”为题，正式宣布免费、自托管的 Web 工具 Capy Agent 将于 10 月 31 日发布，并重点阐述了企业级分析 Agent 的技术路线。
- 专用架构与企业级安全管控：区别于个人及通用 Coding Agent，Capy Agent 采用基于 Web 的私有化部署形态，提供企业级大模型/数据源接入集中管控、精细化权限隔离与全链路安全审计；同时创新性地引入 Markdown as Dashboard 与 Coding MCP，不仅实现了分析结果的动态实时刷盘，更攻克了传统 MCP 协议导致上下文窗口溢出的行业痛点。
- Markdown 文件系统上下文层：针对大模型数据分析“准确率低”的核心难题，引入了比较前沿的 Markdown 文件系统上下文架构（Context Layer）。通过将数据 Schema、指标语义（Ontology）、业务 SOP 及 Agent 规则统一抽象为结构化的 Markdown 目录树，允许大模型采用原生文件指令进行渐进式读取，将复杂分析任务的准确率从 17% 提升至 86%，并完美支持集成 MetricFlow 等传统语义层。
- 实时与 Serverless 是 AI 时代的数据底座未来：AI Agent 正在推动分析需求呈现 10 到 100 倍的爆发式增长。面对海量高并发与即时（Ad-hoc）查询，数据底座必须向亚秒级实时查询与 Serverless 极速弹性架构演进。亚秒级是实时分析的底线，而极低的查询延迟与 Serverless 的弹性伸缩，是让高并发、强 Ad-hoc 特征的 Agent 查询真正走向平民化落地的关键所在。
Litefuse：Agent 可观测性与评估平台
Apache Doris Committer、飞轮科技资深开发工程师姜凯强调，在生产环境里面，每天产出几百 TB 规模的 Agent Trace 数据会成为一种常态。而全链路的可追溯与可评估，是建立企业级 AI 信任的唯一解。
针对 Agent 深度参与决策后出现的路径偏离、幻觉及断言正常但业务逻辑错误等难题，姜凯在演讲中展示了基于 Litefuse 和 Apache Doris 打造的 Agent 可观测与评估平台。
- 可观测侧（Observability） ：基于 OpenTelemetry 与专用 SDK 进行 Trace 采集，支持 100+ AI 生态集成。提供全局调用树展开（耗时、Token、成本逐层下钻）、基于 Session 用户/环境的维度查询，以及线上 Bad Case 一键标注与数据集沉淀。
- 评估侧（Evaluation） ：结合离线数据集与线上 Trace 回流，支持程序断言与大模型自动评估（涵盖准确率、检索质量与执行轨迹合理性），提供多版本对比实验与基线沉淀。
- 生态兼容与交付：全面兼容 Claude Code、Cursor、OpenClaw、Pi Agent 等 Agent 框架以及 MiniMax、Kimi、阶跃星辰等模型。提供开源版、企业版（SSO、项目级权限、SCIM 用户生命周期管理）及云托管三种交付形态。
在 1TB 规模 / 6.6 至 7 亿行同源数据基准测试中，Litefuse 展现出领先的优势，导入提升 2.3 倍，总存储节省 35%，高频查询均可亚秒级返回。
千行百业实践：从实时湖仓到 DataAgent 落地
在为期两天的实时湖仓与数据智能论坛中，来自金融、互联网、AI-Native、央国企政府、游戏、交通物流、消费和先进制造等数十个行业的头部企业技术专家，分享了他们如何借助 Apache Doris 或 SelectDB 实现实时多模湖仓、DataAgent 与 AI 应用落地，揭秘海量数据场景下的前沿实践。
在实时湖仓论坛上，来自 Lance、京东零售、华安基金、民泰银行、快手、中通快递、徐工汉云、欣旺达动力、实时未来科技的技术专家，分享了他们如何基于 Apache Doris 或 SelectDB 搭建实时数仓、湖仓一体及加速方案。在数据智能论坛上，来自网易、亚马逊、Datus、土巴兔、哥瑞利、正泰集团、白鲸社区、算秩未来、DataClaw、如祺出行、快猫星云、蓝象智联、NineData、青藤云安全、德灵的技术专家，分享了在 AI Agent 等智能场景下 Apache Doris 或 SelectDB 的落地实践及具体收益。
以实时为基，拥抱 Agentic AI 未来
为期两天的 Apache Doris Summit 2026 已圆满落幕。从实时分析到实时智能，Apache Doris 正在重新定义 AI 时代的数据基础设施，以实时为基，拥抱 Agentic AI 未来。感谢所有讲师的倾情分享、生态伙伴的无私协作，以及全球开发者的一路同行！明年 Summit，我们不见不散！
资料获取提示：
- 演讲回放及 PPT 资料将在 10 个工作日内逐步上线，您可通过 SelectDB 官方公众号或 Summit 官网获取最新动态：
- 关注 SelectDB 公众号，回复关键词 「 Summit 」 或 「 2026 」 获取演讲 PPT；
- Doris Summit 官网：doris-summit.org.cn
我们是谁
飞轮科技由 Apache Doris 原创团队创立，贡献超 95% 社区代码。旗下的 SelectDB 是基于 Apache Doris 内核打造的企业级/云原生极速分析数据库，在 Doris 极致性能基础上增强了安全稳定、可视化运维及多云托管等能力（性能快同类 3-10 倍）。面向 Agent 时代，为全球企业提供智能体时代的高性能实时分析服务。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
From 21 to 22 October 2026, the Apache Doris Summit 2026 on-line technology summit, hosted by Flying Wheel Technology, successfully ended. With the theme “Powering Real-Time Information for the Agency Era”, the Summit succeeded in engaging more than 120,000 developers, data architects and AI technologists around the globe to witness the historic change in data infrastructure from “real-time analysis” to “real-time intelligence”.
The Summit created three heavy-pound segments of the main forum, the Real-Time Lakes Forum and the DataIntelligence Forum, which brought together 34 leading experts from the Internet, finance, government, logistics, auto-driving, and intelligent manufacturing, around the front lines of AI multi-mole lake silos, incremental computing, Agenic Analytics, observable assessments, and enterprise-level production.
Upgrade and Change: Apache Doris 5.0 Replay In.
The Summit opened with a keynote address by Apache Doris PMC Chair Chen Ming Rain, entitled " From real-time analysis to real-time intelligence: Apache Doris 5.0 driving the Age of Ages of Ages ".
Looking back over the past 12 months, Apache Doris maintained an average of about 3 weekly versions of a high frequency evolution rhythm of publish17. The community now has more than 760 contributors, 109 Committer and 26 PMC members; GitHub Stars has broken through 16,000, with more than 4,400 major PRs. Open-source activities cover overseas positions from domestic cities to Singapore, India, San Francisco, Sydney and Europe, and Apache Doris is accelerating global prosperity.
In the AI era, data consumers are expanding from people to Agent, and the value chain of data is moving from traditional data to insights into data to context to action. In the face of a 10- to-100-fold demand for Agent queries with strong Ad-hoc characteristics, Apache Doris 5.0 of publish is about to set up a whole new three main lines of presentation at the end of the year:
- Data layer (multi-molecular lake silo): Integration of structured and non-structured data into a unified system of analysis.
- Calculator layer (Incremental calculation): to bring the results up to date in real time with data changes.
- Interactive (Agent) analysis: Depth integration of SQL, search capability and LLM.
The combination of these three directions constitutes an entirely new positioning for Apache Doris 5.0: data infrastructure for Agent, moving from real-time analysis to real-time intelligence.
Depth integration of bottom engine break-ins with upper Agenic AI
In the context of the deep integration of AI Agent and LLM into the business landscape, the data infrastructure faces key challenges such as data multimodulation, computing continuity, interactive naturalization, chain observation and deployment of Serverles. In the main forum, a number of technical experts were able to demonstrate ground-level technological breakthroughs and engineering practices of Apache Doris and SElectDB around the multi-molecular lake silos, incremental computing, and the interface of Age Z48Z.
Polymole lake silo: storage open, computing uniform
Apace Doris PMC Chair Chen Ming Rain points out that in the AI era, the integration of multimodular data is the cornerstone of Agent's ability to sense the real world.
In a presentation at the main forum, Apache Doris, United Express, LanceDB, officially announced the Apache Doris 5.0 multi-molecular lake silo structure for AI. Faced with the "full, uniform, real-time" context demand for data layers from AI Agent and LLM applications, Apache Doris gave an entirely new solution to "storage remains open, search is consistent with execution".
In other words, there is no need to move all data to a single storage system, allowing data to be retained in their respective open formats and to be harmonized at the search and execution levels.
- Storage open, load decomposition: no relocation of data is required to allow open formats such as Iceberg, Paemon, Lance to carry long-term graded, continuously updated, or multi-modular data sets; Doris engine directly reuses original lake table vectors, full text and marker index, avoiding redundant storage and synchronized costs.
- Computation harmonization, single SQL mixed search: Breaking the structural and unstructured data barriers, completing the "Variant Revival + Variant Dynamic Path Filter" in a single SQL in the same implementation plan + Multi-source Table Join + permission filtering, fundamentally addressing the pain of the traditional architecture being "similar but not available".
- Second-level real-time, closed-ringback: Supports the construction of atom-focused labels produced by Agent by syntax such as MERGE INTO back to Paimon Lake Table; provides up-to-date analysis of context using the new Fluss Union Read (Live-in-Log), seamless synthesis of historical snapshots and seconds real-time logs.
In addition, fast-trackers shared production practices based on Paimon ' s storage of hundreds of high-dimensional vectors and direct search using Doris, and LanceDB demonstrated the latest results of the integration of multi-modular AI data sets with the Doris engine depth. Alibaba Cloud product expert Feng Hou also shared the view that the SElectDB multi-molecular lake structure landed on the pole of an intellectual driving scene, achieving a vector recall rate of up to 96 per cent through direct retrieval of 128 million levels of multi-modular data from Paimon data lakes, with a performance decline of only 1.07 to 1.36 times the internal table.
Incremental calculation: from batch updating to continuous calculation
Apache Doris PMC, the technology manager of the wheel technology storage engine, Zhou Flying indicates that the more fresh the data are, the closer Agent's reasoning and decision-making is to the real world. Future data calculations are no longer scheduled batch updates, but continue to evolve as data flows.
In order to ensure that Agent acquires the latest business context, Apache Doris 5.0 in the reasoning process, he introduced a completely new incremental computing architecture and an anatomical view line incrementality (IVM) to the computational level, changing the full-scale or fine-shape processing logic of previous reliance on timing tasks.
Bottom innovation: precision capture and state flow control
The roll-out of Row Binlog and Table Stream technologies at the bottom not only allows accurate recording of changes in grade data prior to and after the update (Update Before / Update After) but also achieves a state-of-the-art consumption point management and service security. It is possible to calculate the ecology in real time from a change of seconds to an external stream through the Flink Contractor, or directly driven within the engine.
Engine Breakout: Single SQL achieves a second line refresh
Based on Row Binlog, a new generation of IVM allows users to automatically maintain complex business indicators through a declarative SQL. Tests show that under the 1 billion-grade base table data, IPM only scans hundreds of lines of data subject to change, and that the P50 delay in changing data to indicator queries is reduced to 0.26 seconds, with only one percent of time and CPU consumption of the full amount of refreshing.
As of now, IVM has successfully landed on scenes such as the electrician ' s real-time silos, and has achieved the efficient integration of the delivery GMV wide-band integration with the public dimensions. In addition, Doris is further extending its IMM and incremental computing capabilities into open lake warehouse formats such as Iceberg and Paemon, significantly reducing the cost of maintaining the data stream line and providing an extremely fresh and low-cost data base for enterprise-level AI Agent and real-time business decisions.
Agency Analytics: Capy Agen publish, new business-level analysis paradigm
Apache Doris, along with the founder of the SelectDB, is officially announcing a free-of-charge, self-custody Web tool, Capy Agent, under the title “Talk to Your Data”, on October 31, publish, with a focus on the technical route of enterprise-level analysis Agent.
- Specialized structures and enterprise-level security controls: Distinguishing from individual and universal Coding Agent, Capy Agent, using a Web-based privatization deployment, provides enterprise-level LLM/data source access centralized controls, refines permission isolation and full-chain security audits; and innovatively introduces Markdown as Dashboard and Coding MCP, not only achieves dynamic real-time brushes of the results of the analysis, but also tackles the industry pain points where the traditional MCP agreement has led to context window spills.
- Markdown filesystem context Layer: The more cutting-edge Markdown filesystem context architecture (Context Layer) was introduced in response to the core problem of "low accuracy" analysis of LLM data. LLM is allowed to use original file instructions for progressive reading by harmonizing the data Schema, Ontology, Operations SOP and Agent rules into a structured Markdown catalogue tree, increasing the accuracy of complex analytical tasks from 17% to 86%, and perfectly supporting the integration of traditional semantic layers such as MetricFlow.
- Real-time versus Serverless is the future of the data base of the AI age: AI Age is driving an explosive growth of 10 to 100 times the analytical demand. In the face of high mass and instantaneous (Ad-hoc) queries, the data base must evolve in real time with subseconds with the Serverless hyperflexible architecture. The subsecond level is the bottom line of real-time analysis, and the extremely low delay in queries and the elasticity of Serverles is the key to making the Agent query, which has a high-haired and strong Ad-hoc character, truly move towards civilianization.
Litefuse: Agent Observation and Assessment Platform
Apache Doris Committer and Kang Kay, a senior development engineer in Flying Wheel Technology, stressed that it would become normal to produce hundreds of TB-sized Agent Trace data per day in a production environment. The traceability and assessability of the whole chain is the only way to build corporate-level AI trust.
In response to the problems of the path deviations, hallucinations and assertions of normal business logic following Agent’s deep participation in decision-making, Kang Kai presented a platform for observation and assessment of Agent based on Litefuse and Apache Doris.
- Observable side (Observability): Trace collection based on OpenTelemetry and SDK, supporting 100+AI ecological integration. Provides global call tree roll-out (time-consuming, Token, cost-drive-down), session user/environment-based dimension queries, and a web-based Bad Case note and data set deposition.
- Evaluation: Supporting program assertion with automatic assessment of LLM (covering accuracy, search quality and rationality of execution trajectory) in combination with offline data sets and online Trace flow, providing multiple versions of comparative experiments and baseline deposition.
- Ecocompatibility and delivery: full compatibility with the Agent framework of Claude Code, Cursor, OpenClaw, Pi Agent and the MiniMax, Kimi, Stepstar, etc. Three delivery modes are available: open source, enterprise (SSO, project level permission, SCIM user life cycle management) and cloud hosting.
In the 1TB scale / 6.6 to 700 million line-source data baseline tests, Litefuse shows a leading advantage, import up 2.3 times, total storage saving 35%, and high frequency queries can be returned in subseconds.
A thousand practices: from real-time lake silo to DataAgent landing
In a two-day real-time lake silos and data intelligence forum, leading technical experts from dozens of industries, including finance, the Internet, AI-Native, the Zong State Enterprise Government, games, transport logistics, consumption and advanced manufacturing, shared how they were able to achieve real-time multi-molecular lake silos, DataAgent and AI applications in order to uncover the cutting-edge practices of the sea mass data scene.
At the Real-Time Lakes Forum, technical experts from Lance, Kyeongdong Retail, the Huaan Foundation, Min Tai Bank, Express, China Express, Xu Jie Han Yun, Shinwanda Power, and Real-Time Future Technologies shared how they built real-time and digital storage, integration and acceleration programmes based on Apache Doris or SelectDB. At the data intelligence forum, technical experts from Internet-based, Amazonian, Datus, Tuba Rabbit, Garrette Group, White Whale Community, Counting the Future, DataClaw, e.g., Traveling, Fast Cat Nebula, Blue Elephant, NineData, Endo Cloud Safety, Drum, etc. shared the localization and concrete benefits of Apace Doris or SelectDB in smart settings such as AI Agent.
Embrace Agenic AI for the future based on real time
Two days ago, Apache Doris Summit 2026 came to a successful conclusion. From real-time analysis to real-time intelligence, Apache Doris is redefining the data infrastructure of the AI age, based on real-time embraces the Agentic AI future. Thank all the lecturers for sharing their words, the selfless collaboration of the eco-partners, and the fellow global developers! Next year Summit, we're not going anywhere!
Information access tips:
The presentation and PPT material will be available in 10 working days, and you can access the latest developments through the SElectDB official WeChat Official Account or Summit website:
- Focus on SelectDB WeChat Official Account, replying to the keyword "Summit" or "2026" for speech PPT;
- Doris Summit: doris-summit.org.cn
Who are we?
Flying Wheel Technology was created by the original team of Apache Doris, which contributed more than 95% of the community code. SelectDB under the flag is based on the Apache Doris kernel's Enterprise/Cyclastic Rapid Analysis Database, which enhances security stability, visualization and cloud hosting capabilities based on Doris's extreme performance (3-10 times faster). For the Age of Agent, high-performance real-time analysis of the AI agent era was provided to global enterprises.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.