当模型开始理解物理世界，当 Agent 开始进入真实生产，数据平台需要提供哪些新的能力？在 2026 云栖大会上，阿里云智能集团计算平台事业部 EMR 负责人李钰以《EMR：Agentic Lake 全模态数据处理与极速分析》为题，分享了相关实践与思考。
去年 EMR 的判断是：开放让计算与存储连接起来，高效使规模化使用成为可能，智能让产品的专业能力更容易被使用。这些判断依然成立，但过去一年出现了两个值得关注的变化：Agent 从单轮、多轮对话进入真实生产；模型的应用则从数字场景拓展到物理世界，开始学习空间、时间与动作之间的关系。
李钰提到，面向这些新场景，AI-Ready Data Infra 需要具备三项关键能力：在数据生产上，让视频、点云、动作和传感信号在时间、空间与质量上保持一致，能够还原真实动作；在数据服务上，除传统在线服务外，还要通过标量、向量、全文等组合查询支持模型迭代与样本圈选；在长期生产中，让可信的成功与失败经验经过治理转化为知识，成为 Agent 下一轮工作的基础。EMR 的年度升级正是围绕这三项能力展开。
Serverless Spark：让多模态采集成为可训练资产
具身智能对数据生产提出了不同于传统分析的要求。人形机器人需要对齐视频、点云和关节轨迹；工业操作臂需要从高频力控信号中识别有效动作；四足机器人则需要融合相机、IMU 与地形信息。这些场景共同要求数据在时间、空间和质量上保持一致，能够还原真实动作。
EMR 宣布即将发布 Serverless Spark 具身智能数据处理平台，将精准重建、批量处理与可视化交付结合起来，以 Serverless 按需弹性承接端到端训练数据生产。平台面向 PB 级多源采集，支持 LeRobot / RLDS 数据集交付，原生对接 Paimon / Iceberg。
可信重建与质量验证贯穿交付
以头戴式设备采集的第一视角视频为例，看见动作并不等于准确还原动作。近距离定位误差会影响手与物体的接触关系，多人干扰可能导致操作手身份切换，遮挡和运动模糊则可能造成轨迹断裂。
平台通过双目融合、佩戴者双手关联与时序关系，恢复手部、相机与世界运动。对于轨迹缺口，仅在两端可信、身份一致且通过检查时进行短缺口补全；长缺口或身份切换不强行补全，避免将缺乏依据的推断带入训练集。
交付环节支持将 EgoAir ZIP 数据处理为 LeRobot v3 标准数据，同步提供重建结果、动作时间线和相机 / 世界双视角回放。输入检查、可信重建、标准输出、同步回放与发布质检贯穿全流程，让数据既可训练，也可回放、可核验。
精度与效率协同提升
根据团队内部测试，在 HOT3D Quest3 的 100 条数据上，EMR 双目优化相较 HandFlow 原版单目方案，将手部定位误差指标 MPJPE 从 70.03 mm 降至 28.47 mm，误差降低 59.3%。在独立的 EgoAir 具身视频测试中，两阶段优化后，折算产品处理耗时从 902.499 秒降至 536.550 秒，累计减少 40.5%。
规模化生产还需要通用引擎与异构计算支撑。Fusion 2.5 通过 SIMD JSON、多表达式复用和 scan sharing 减少重复计算，优化 Spill、Shuffle、排序与关联执行，并结合模型分池、跨阶段流水和按负载调度提升资源效率。团队完整 12 项内部测试的平均加速比为 2.08×，其中 Cache Table 代表场景达到 3.09×，比较基线为去年云栖大会发布的 Fusion 2.0 版本。
Serverless StarRocks：在同一份湖仓资产上统一分析与检索
经营分析需要多维聚合，在线服务需要高频点查，模型数据圈选则需要结构化、地理、全文与向量条件组合。若每类需求分别依赖数据副本和查询引擎，应用层就需要承担跨系统关联、结果拼接与一致性校验，训练目标变化时容易引入重复加工。
本次大会重磅发布 EMR Serverless StarRocks 全模态全场景在线分析服务，面向 Paimon / Iceberg / Lance 等开放 Lakehouse 数据形态，基于 One Data 复用共同数据资产，由一套引擎组合完成在线分析与检索，索引和缓存服务于统一查询与性能。
一条 SQL 承接复杂圈选意图
智能驾驶团队可能需要寻找指定区域内、雨天夜间、满足车速条件、出现车辆加塞，且与参考片段相似的样本。这一需求同时涉及结构化过滤、地理条件、全文匹配和向量排序。
加工引擎从视频、LiDAR 和车辆信号中形成标签、文本、向量与属性后，StarRocks 用一条 SQL 组合条件，完成过滤、召回和排序。团队可以先预览候选样本，再调整条件，持续收敛到目标数据。需求变化由查询承接，减少应用层反复拆分查询、跨引擎拼接结果的复杂性。
统一圈选也已用于阿里内部 AI Data for Qwen 的数据准备。Paimon on Pangu 承载共同资产，Ray / AI Function 等能力提取属性、标签、文本与向量，StarRocks 承接圈选、预览和分布检查，确认内容与质量后由平台导出训练集。加工、圈选和交付围绕同一套资产展开，使数据随训练目标变化持续复用。
四类湖仓负载的性能优化
EMR 围绕元数据解析、数据合并、索引访问和统一查询执行进行优化。大会披露的内部测试中，TPC-DS 10 TB 场景下，StarRocks + DLF Paimon 相较此前公开基线，性价比提升 50%；YCSB 点查场景下，1 KB 和 10 KB 数据大小下的性能分别达到 105,000 和 44,000 QPS。
向量测试在 VectorDBBench Cohere 10M、Top 10、50 并发、保持 99% 召回率的场景下，面向 DLF Paimon 和 Lance 湖表数据的查询性能分别达到 19,308 和 18,772 QPS。全文测试在 Amazon Reviews 1.5 亿行数据上达到 353 QPS，查询时延为 130 ms。各负载独立测试，不代表同时运行时的混合吞吐。
Managed Agent：让长期生产工作积累可信经验
具身数据生产与全模态分析主要解决 Model-ready，长期生产 Agent 则提出 Agent-ready 的要求。生产现场的数据分布、负载和软件版本持续变化，过去有效的方法需要重验。完成一次任务，不等于能够长期胜任复杂工作。
大会正式发布 EMR Agent 与 Managed Agent 平台，将长期生产 Agent 的共性能力提供给开发者。平台管控覆盖开发部署、资源管理、运行运维，以及权限、必要审批、沙箱、可观测和审计；引擎内核负责组合数据与工具、随新证据推进任务，并保留可信过程信息。
以作业诊断为例，Agent 需要关联日志、指标和配置形成方案，再结合真实结果检验方法及适用范围。失败任务尤其需要保留上下文、查询数据、方案、决策和人工更正，经验证与治理后形成领域知识，支持后续工作。持续进化的基础不是简单保存对话，而是积累有依据、可复用的专业经验。
平台通过 API 原语开放 Agent 引擎，开发者可以装配模型、专业工具与领域知识，带入任务上下文，设置执行边界并检查交付。首个实战应用 EMR Agent 覆盖 EMR on ECS、Serverless Spark 和 Serverless StarRocks，提供诊断、巡检、慢 SQL 分析，以及授权处置、验证状态和可回看报告。
李钰将这一目标类比为建设 Agent 时代的“Hadoop”：把复杂工作推进、可信信息管理与知识积累封装为共用基础设施，让业务团队聚焦领域建模与 Agent 本体，而不必重复实现底层机制。
从数据供给走向生产能力
今年的 EMR 产品能力升级以开放湖仓为共同基础：Serverless Spark 将真实采集转化为可靠训练资产，Serverless StarRocks 让同一份资产服务分析与检索，Managed Agent 则让可信经验支持持续工作。EMR 通过弹性计算、统一查询和治理复用连接 Model-ready 与 Agent-ready，协同阿里云 Agentic Cloud，支撑模型与智能体进入真实生产环境。
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。
When models begin to understand the physical world, when Agent begins to enter real production, what new capabilities do data platforms need to provide? At the 2026 Cloud Habitat Congress, the head of EMR, the Alibaba Cloud Smart Group for Computation of Platforms, Li Qian, shared practice and reflections on the topic " EMR: Analogue Lake ".
Last year, EMR judged that openness connects computing to storage, that efficient use of scale is possible, and that intelligence makes product expertise easier to use. These judgements are still valid, but there have been two notable changes over the past year: Agent has moved from single-wheel, multi-wheel dialogue to real production; and model applications have expanded from digital scenes to the physical world, beginning to learn the relationship between space, time and action.
Li said that for these new scenarios, AI-Ready Data Infra needs three key capabilities: In the production of data, allowing video, cloud, motion and sensory signals to be consistent in time, space and quality so that they can be restored to true action; in the data services, supporting the iterative model and sample selection through combinations of queryes, vectors, full text, etc., in addition to traditional online services; and in long-term production, transforming credible success and failure experiences into knowledge into the basis for the next round of Agent ' s work. The EMR annual upgrade is built around these three capacities.
Serverless Spark: Making multi-mode collection a training asset
Smartness imposes requirements for data production that differ from traditional analyses. Human robots need to be equipped with video, cloud and joint trajectories; industrial handlers need to recognize effective action from high-frequency force-control signals; and four-foot robots need to integrate camera, IMU and terrain information. Together, these scenarios require data to be consistent in terms of time, space and quality and to be able to restore real action.
EMR announced that it was about to host publish Serverless Spark ' s smart data processing platform, which combines precision reconstruction, batch processing and visualization of delivery, with Serverless ' s flexible end-to-end training data production. The platform supports LeRobot / RLDS data delivery, primary docking Paimon / Iceberg, for multi-source collection at the PB level.
Credible reconstruction and quality certification cut across delivery
For example, the first-view video collected with head-on devices does not mean that the action is accurately restored. A close-positioning error may affect the contact of the hand with the object, multiple interference may result in a hand-to-man identity switch, and resistance and motion blurs may result in a fracture of trajectory.
The platform restores hand, camera and world movement through dual-eyed integration, hands-to-time connection. In the case of trajectories, the gap is filled only at the end of a period of credible, consistent identity and at the time of inspection; the long gap or identity switch is not forced to be filled and the lack of evidence is avoided.
The delivery chain supports the processing of the Egoair ZIP data as LeRobot v3 standard data, providing synchronized reconstruction results, action time lines and camera/world-wide double perspective. Input checks, credible reconstruction, standard output, synchronized playbacks and publish mass checking go through the entire process so that data can be both trained and replayed and validated.
Synergy and efficiency enhancement
According to a team in-house test, on 100 HOT3D Quest3 data, the EMR binary optimization compared to the original HandFlow Monument program reduced the manual position error indicator MPJPE from 70.03 mm to 28.47 mm, a 59.3 per cent reduction. In the independent Egoair body video test, after two-stage optimization, the processing time of the converted product decreased from 902.499 seconds to 536.550 seconds, a cumulative decrease of 40.5 per cent.
Scalable production also requires universal engine and isomer computational support. Fusion 2.5 Increases resource efficiency by reducing double counting through SIMD JSON, multi-expressive reuse and scan sharing, optimizing Spill, Shuffle, sorting and association execution, and combining model sub-pools, trans-phase streaming and load movement. The average acceleration ratio of 12 internal tests for team integrity was 2.08 x, of which Cache Table represented 3.09 x and compared the baseline to last year ' s version of publish, Fusion 2.0.
Servers StarRocks: Unified analysis and retrieval on the same lake silo asset
Business analysis requires multi-dimensional aggregation, online services require high-frequency spot-checking, and model data circles require a structured, geographic, full-text and vector combination. If each type of demand relies on a copy of data and a query engine, the application layer will have to assume cross-system linkages, collage of results and consistency verification, and training objectives change easily to introduce duplicate processing.
The online analysis service of the Conference, which weighs a pound of publish EMR Servers StarRocks, is open to Lakehouse data patterns, such as Paimon / Iceberg / Lance, based on One Data reuse of common data assets, completed online analysis and retrieval by a combination of engines, indexing and cache for unified querying and performance.
A SQL to carry the complex circle selection intent
An intelligent driving team may be required to search for sample, which is similar to the reference footage, within a designated area, during rainy days and nights, meets speed conditions, and has a vehicle fit. This requirement involves both structured filtering, geographical conditions, full text matching and vector sequencing.
After processing engines form labels, text, vectors and properties from video, LiDAR and vehicle signals, StarRocks uses a combination of SQL conditions to filter, recall and sort. The team is able to preview the candidate sample before adjusting the conditions to keep the data on target. Changes in demand are carried out by queries, reducing the complexity of repeated split queries and cross-engine integration results at the application level.
The unified circle selection has also been used for data preparation for AI Data for Qwen within Ali. Paimon on Pangu carries common assets, Ray / AI Function has capabilities to extract attributes, labels, text and vectors, StarRocks takes the circle selection, preview and distribution check, confirming content and quality, and exports the training set from the platform. Processing, rounding and delivery revolve around the same set of assets, allowing data to be reused continuously as training targets change.
Optimization of 4-class lake bunker loads
EMR optimizes metadata resolution, data consolidation, index access and uniform query execution. In the internal test disclosed by the Congress, the performance in the size of 1 KB and 10 KB data reached 105,000 and 44,000 QPS in the TCC-DS 10 TB scenario, with StarRocks + DLF Paimon showing a 50 per cent increase in value compared to the previously publicly available baseline.
VectorDBBeach Cohere 10M, Top 10, 50 and maintain a 99% recall rate for the vector test, the search performance for DLF Paimon and Lance Lake Table data reached 19,308 and 18,772 QPS, respectively. The full text test amounted to 353 QPS on Amazon Reviews 150 million lines, and latency was 130 ms. Each load is independently tested and does not represent a mixture of throughput while running.
Managed Agent: Building credible experience in long-term production
Data production and whole-state analysis mainly address Model-ready, while Agent's long-term production demands Agent-ready. The distribution of data, loads and software versions at the production site is constantly changing, and methods used in the past need to be retested. A mission does not amount to a long-term capacity for complex work.
The official publish EMR Agent platform of the Congress and Managed Agent provides developers with common capabilities for the long-term production of Agent. The management of the platform covers the development of deployments, resource management, operational dimensions, and permission, necessary clearance, sandboxes, detectable and audit; the core of the engine is responsible for combining data and tools, moving forward with new evidence and maintaining credible process information.
In the case of operational diagnostics, for example, Agent needs associated logs, indicators and configurations to form a program that combines real results testing methods and scope of application. In particular, failure requires the retention of context, access to data, programmes, decision-making and manual corrections, empirical and governance-based knowledge in the field to support follow-up. The basis for continuous evolution is not simply the preservation of dialogue, but the accumulation of evidence-based and reusable professional experience.
The platform is open to Agent engines through the API language, where developers can assemble models, professional tools and field knowledge, bring in task context, set up the execution boundary and check delivery. The first operational application EMR Agent covers EMR on ECS, Servers Spark and Servers StarRocks, providing diagnostics, inspections, slow SQL analysis, and authorized disposal, validation status and reversible reports.
Li compares this objective to building the “Hadoop” of the Age: to encapsulate complex work forward, credible information management and knowledge accumulation as a common infrastructure, to focus the business team on modelling the field and the Agent body without duplicating the bottom-up mechanism.
Moving from data availability to productive capacity
This year's EMR upgrade of product capacity is based on open lake silos: Servers Spark transforms real collection into a reliable training asset, Servers StarRocks allows the same asset service analysis and retrieval, Managed Agent allows credible experience to support continuous work. EMR supports the model and AI agent into the real production environment through flexible computing, uniform query and governance reuse links to Model-ready and Agent-ready, in collaboration with Alibaba Cloud Agency Club.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.