Big Data & AI Product Technology Monthly [July 2026]. Covering July's technology updates, product and feature releases, and market and customer application practices, this digest helps you quickly catch up on the latest developments in Alibaba Cloud Big Data & AI.
MaxCompute - Agentic Ecosystem Access Suite Officially Released
MaxCompute has launched the Agentic ecosystem access suite, comprising three components: an AI data exploration client, an MCP Server, and a CLI. With AI Query (intelligent data Q&A) and AI Coding (assisted coding) at its core, the suite bridges desktop interactive analysis and AI Agent automation, covering the full spectrum of data development scenarios.
MaxCompute - MaxFrame EGO Data Pipeline Solution for the Embodied AI Industry Released
Targeting embodied intelligence training data production, MaxFrame has released the EGO end-to-end data pipeline solution for the embodied AI industry. Built on the MaxCompute DPE engine, it runs the complete workflow from first-person human hand videos to LeRobot datasets, orchestrates CU/GU/Token resources under unified scheduling, and covers multiple stages including segmentation and frame extraction, geometric estimation, 3D hand reconstruction, action segmentation, VLM labeling and embedding, and LeRobot export.
MaxCompute - MaxAgent Official Release
MaxAgent is the AI assistant of the MaxCompute platform, providing users with full-link self-service and intelligent O&M capabilities to lower the usage barrier and improve operational efficiency. With MaxAgent, users can complete job diagnosis, data querying, cost analysis, resource optimization, permission auditing, metadata exploration, and other tasks in natural language, handling daily development and O&M work without memorizing complex commands or searching documentation manually.
MaxCompute - Multimodal Retrieval Capability Released
Building on its existing offline batch processing and data analysis engine, MaxCompute now delivers a full-link capability chain through the native VECTOR data type, the AI_EMBEDDING function, VECTOR INDEX, and the VECTOR SEARCH function. Users can use SQL alone to complete the end-to-end pipeline from multimodal data ingestion, embedding generation and vectorized storage, and index building, through to offline batch similarity retrieval — especially suitable for scenarios involving massive amounts of data.
MaxCompute - Enhanced Nested Column Pruning for SQL
A performance optimization capability from MaxCompute for querying complex-typed data. When a table contains STRUCT, ARRAY, MAP, or multi-level composite nested types, queries typically access only a small number of fields. With enhanced nested column pruning enabled, MaxCompute scans only the subfields actually needed when reading data and skips unused nested fields, thereby reducing storage I/O, lowering memory overhead, and improving query performance.
MaxCompute - Support for Generated Columns
Generated Column is a special column type provided by MaxCompute. Its values are automatically computed from other columns in the table through a deterministic expression, requiring no manual insertion or updates from users. Generated columns significantly simplify data processing logic, ensure data consistency, and improve query performance.
MaxCompute - MaxFrame AI Function Supports Multimodal Mixed Input
MaxFrame AI Function now supports multimodal mixed input, allowing any combination of text, images, audio, and video to be passed in a single call. Integrated with the Alibaba Cloud Model Studio (Bailian) Qwen-Omni omni-modal model, it performs cross-modal joint inference and supports batch processing of mixed-modality data.
MaxCompute - MaxFrame Automatic Elastic Memory
MaxFrame provides automatic OOM detection and elastic scale-up retry at SubStep granularity, upgrading memory resource allocation from global peak-based reservation to shard-based on-demand allocation. This fundamentally resolves memory waste and job failures caused by data skew. The feature can be enabled with a single line of configuration and can reduce CU costs by up to 60%+ in specific scenarios.
Centered on Hologres instances, the AI Assistant is powered by multiple Skills distilled from seasoned data warehouse experts. It offers multi-agent collaboration covering issues across the full data warehouse lifecycle, including knowledge Q&A, instance diagnostics, and SQL tuning, helping users progress from beginner to expert — zero code, zero barriers, out of the box.
DataWorks - Data Agent Token Consumption Supports Tenant-Level and User-Level Quota Management
User quotas can be adjusted in batches or individually, allowing teams to allocate usage based on roles and task requirements and reducing the cost risk of over-consumption.
DataWorks - ChatBI Sessions Support Temporary File Uploads for Analysis
ChatBI sessions now support temporary file uploads, allowing users to upload local files directly as temporary data for ad hoc analysis. This facilitates agile exploration and improves the flexibility of data analysis.
DataWorks - ChatBI Deeply Integrates with the Semantic Layer and Data Agent
ChatBI datasets now fully link with semantic parsing tasks and connect to the DataWorks Data Agent, significantly improving the accuracy of natural language Q&A and the depth of intelligent analysis.
DataWorks - ChatBI Supports Custom Skills and Cross-Platform AI Collaboration
ChatBI opens up custom Skill registration and management and fully integrates with the DataWorks AI assistant, enabling a seamless office-collaboration data analysis experience via the IM client.
DataWorks - ChatBI Session Interaction Experience Upgraded with Permission Delegation
ChatBI adds session pinning and a dataset details sidebar, and optimizes UI interactions. It also delegates data source reference permissions, lowering the usage barrier for regular members.
DataWorks - ChatBI Deep Analysis Mode Supports Notebook Generation
In deep analysis mode, ChatBI can now automatically generate a Notebook with preview, turning the analysis process into reusable code for subsequent secondary editing.
DataWorks - New Data Agent-Based Security Management Suite and Skills Released
The data protection and Security Center expert suite, along with multiple security Skills, has been released. By empowering data security operations, permission management, and automated approvals with AI, it comprehensively enhances enterprise data security protection and governance efficiency.
DataWorks - Security Center Supports Backing Up Sensitive Data Access Logs to OSS
DataWorks Security Center now supports exporting and backing up sensitive data access logs to OSS, meeting the requirements of long-term log retention and compliance archiving.
DataWorks - Open Platform Expert Suite Released
The DataWorks Open Platform has released an expert suite focused on API observability analysis and intelligent developer services, comprehensively improving API operational efficiency and the developer onboarding experience.
Realtime Compute for Apache Flink - Flink Agents Supports Streaming AI Agent Development
Realtime Compute for Apache Flink adds Flink Agents capability, supporting the development of event-driven streaming AI Agent jobs. Developers can process continuously arriving business events within Flink jobs and feed real-time data processing results into agent workflows, building AI applications for real-time perception, analysis, and decision-making.
Kafka YAML Source now supports duplicate-key JSON and tombstone delete events, improving compatibility for complex data ingestion and the synchronization of log-compacted topics.
Realtime Compute for Apache Flink - Paimon Variant Shredding Now Available
Paimon in Realtime Compute for Apache Flink adds Variant Shredding capability, enabling tiered storage and processing of semi-structured Variant data. This improves data storage efficiency and query performance, helping customers manage and analyze semi-structured data such as JSON more efficiently.
Realtime Compute for Apache Flink - SQL Server CDC Source Enters Public Beta
SQL Server CDC Source has entered public beta, supporting real-time capture of SQL Server data changes and extending real-time synchronization capabilities for enterprise databases.
Realtime Compute for Apache Flink - Flink AI Function Supports NVIDIA Triton Inference Server
The AI Function in Realtime Compute for Apache Flink now supports NVIDIA Triton Inference Server. Users can connect Flink jobs to deployed Triton inference services and invoke remote models for real-time inference, expanding model service access options and better reusing existing NVIDIA Triton inference infrastructure.
Realtime Compute for Apache Flink - ml_predict Supports Reusing CREATE MODEL Parameters
The ml_predict function now supports reusing some of the model parameters defined in CREATE MODEL. Developers no longer need to repeat the same configuration in every inference SQL statement, keeping model invocation parameters consistent while lowering the cost of job authoring and maintenance. This capability is suitable for scenarios that require building real-time AI inference jobs at scale.
Realtime Compute for Apache Flink - Multimodal Model Invocation Parameter Adaptation Optimized
Realtime Compute for Apache Flink has optimized parameter adaptation for multimodal model invocations, making it easier for developers to call image and video understanding models in streaming SQL jobs. This reduces the complexity of multimodal inference integration and parameter configuration, helping customers bring the understanding results of unstructured content such as images and videos into their data processing pipelines in real time, and improving development efficiency for content understanding and real-time business analysis.
Realtime Compute for Apache Flink VVR 11.7 supports connecting to NVIDIA Triton Inference Server, expanding the options for remote online inference service access. Developers can invoke Triton inference services within streaming data processing jobs to build real-time intelligent processing pipelines.
Realtime Compute for Apache Flink - Flink Agents Streaming AI Agent Development Support
Realtime Compute for Apache Flink VVR 11.7 supports Flink Agents, enabling developers to build event-driven streaming AI Agent jobs based on the open-source Flink Agents. This capability combines real-time data processing with AI Agent development, helping developers build real-time intelligent applications more efficiently.
Realtime Compute for Apache Flink - ml_predict Supports Reusing CREATE MODEL Parameters
The ml_predict function supports reusing the model parameters defined in CREATE MODEL. After a model is created, users can reference the model configuration directly in inference SQL, reducing repetitive parameter entry and simplifying the development of model inference jobs.
Realtime Compute for Apache Flink - Paimon Variant Shredding Support Released
Realtime Compute for Apache Flink adds Paimon Variant Shredding support. This capability optimizes the organization of semi-structured data and improves the storage and query efficiency of Variant data such as JSON in Paimon, helping users build real-time data lakehouse analytics pipelines more efficiently.
Realtime Compute for Apache Flink - Paimon Vector Table Search Capability Released
Realtime Compute for Apache Flink adds Paimon vector table search capability. Users can search Paimon vector tables via the vector_search stored procedure, making it easier to store, manage, and retrieve vector data in lakehouse scenarios, and providing a data processing foundation for applications such as intelligent retrieval and RAG.
Realtime Compute for Apache Flink - Flink Intelligent Assistant Adds Skill Invocation
The console's intelligent assistant supports triggering Skills such as job operations, O&M diagnostics, resource tuning, and knowledge Q&A in natural language from the conversation interface. Execution results are returned in a structured format, along with follow-up parameter recommendations and operational suggestions.
Realtime Compute for Apache Flink - Flink Agent Skills Published on Alibaba Cloud Skill Hub
The first two Agent Skills for Realtime Compute for Apache Flink have been published on the Alibaba Cloud Skill Hub. They support integration with mainstream AI Agent platforms such as Claude Code and OpenClaw to perform instance management, namespace management, and workspace job O&M operations.
Realtime Compute for Apache Flink - New Purchases Support PrivateZone Shared VPC
Newly purchased instances of Realtime Compute for Apache Flink now support PrivateZone shared VPC. In a shared VPC environment, users can purchase instances without additional network configuration, reducing network access complexity and improving instance provisioning efficiency.
Realtime Compute for Apache Flink - Auto Tuning Supports Flink CDC YAML Data Ingestion Jobs
The auto tuning feature now supports Flink CDC YAML data ingestion jobs. In data ingestion scenarios, you can use auto tuning to obtain resource configuration optimization suggestions based on job runtime conditions and apply the tuning, improving resource utilization efficiency and job runtime stability.
Realtime Compute for Apache Flink - Fluss Officially Commercialized
Fluss is officially commercialized, providing a one-stop lake-stream integrated solution with primary-key table upserts and point queries, simplifying real-time architectures and reducing O&M costs.
Realtime Compute for Apache Flink - CDC YAML Supports Invoking Model Studio Built-in Models
VVR 11.8.0.preview.2 supports using built-in models to invoke AI Functions in CDC YAML jobs, and supports configuring system prompts. This capability is a public preview version; we recommend trying it out in a development or testing environment.
Realtime Compute for Apache Flink - AI Function Supports Multimodal Combined Input
The AI Function in VVR 11.8.0.preview.2 adds a content-types parameter, supporting combinations of multiple modalities such as text and images to be passed to models for inference, expanding multimodal intelligent processing capabilities. This capability is a public preview version; we recommend trying it out in a development or testing environment.
Realtime Compute for Apache Flink - Python DataFrame API Adds Multimodal Processing Operators
New multimodal operators have been added, covering scenarios such as image transformation, detection and recognition, face processing, and video frame extraction. Dependencies and models are built into the VVR image. Available in the VVR 11.8.0.preview.2 public preview.
Realtime Compute for Apache Flink - Flink SQL Can Directly Call Multimodal Processing Functions
Once the built-in Python multimodal processing functions are loaded, they can be called directly in Flink SQL without writing Python code. Available in the VVR 11.8.0.preview.2 public preview; we recommend trying it out in a development or testing environment.
Realtime Compute for Apache Flink - New OSS CDC Connector
By consuming OSS event notifications in MNS queues, the connector captures file creation and deletion events in Object Storage Service (OSS) in real time, enabling incremental data reads. It also supports full scans of object metadata under specified paths. Available in the VVR 11.8.0.preview.2 public preview; we recommend trying it out in a development or testing environment.
Added system-triggered automatic inspections: scheduled daily health inspections are performed on all running jobs without manual initiation. Added inspection daily report delivery: daily inspection reports are pushed through the console message center, including a health score overview, a list of at-risk jobs, root cause analysis, and optimization suggestions. This capability is in public beta, free of charge, and does not count against the Flink AI Service free quota.
Flink AI Service has officially launched the pay-as-you-go billing mode. Inference invocations using built-in models incur model invocation fees separate from CU compute resources, billed by token usage. The first 1 million tokens per primary account per region per calendar month are free (input and output combined), and usage beyond that is billed according to pricing.
Added chain-of-thought visualization. Tool invocation processes and intermediate parameters are collapsed by default, and users can expand them as needed to view the complete reasoning process. The assistant's component layout, conversation fluency, and result display format have been optimized; response cards are visually unified and support collapsing and expanding structured content. Also fixed an issue where conversation context was lost in specific scenarios.
The Flink AI built-in model service now supports registration-free invocation. Without configuring an API Key, an endpoint, or a private connection, you can invoke the built-in model service directly in real-time compute jobs, reducing the costs of model integration configuration and network connectivity and accelerating the development and launch of real-time AI applications.
Realtime Compute for Apache Flink - DLF Paimon Vector Table Search
The vector_search and vector_search_agg functions can now directly search DLF Paimon vector tables. Customers can complete vector similarity queries and aggregated searches within real-time compute jobs without additional synchronization of vector data, reducing the development and O&M costs of real-time intelligent retrieval applications.
Realtime Compute for Apache Flink - SQL AI Function Supports Audio and Video Multimodal Processing
SQL AI Function adds multimodal processing capabilities for audio and video. Users can use SQL AI Function in real-time compute jobs to process audio and video data, combined with model inference to achieve multimodal content analysis and understanding. This expands the data types that real-time AI inference can handle, helping developers build richer real-time intelligent applications.
Realtime Compute for Apache Flink - Flink Agents Supports SQL Table-Valued Function Invocation
Realtime Compute for Apache Flink supports invoking Flink Agents as table-valued functions (TVFs) in SQL jobs, and is compatible with Flink Agents 0.3. This optimization allows developers to integrate AI Agent capabilities directly into existing SQL development workflows, reducing cross-language and interface integration costs and improving the development efficiency of real-time intelligent applications.
Realtime Compute for Apache Flink - Lookup Join Supports Custom Timeout Handling Logic
Asynchronous Lookup Join supports custom timeout handling logic in AsyncLookupFunction and AsyncTableFunction. You can handle asynchronous query timeout results according to business needs, reducing the impact of timeouts on real-time job stability and data processing continuity, and improving controllability and fault tolerance in exceptional scenarios.
The Kafka Canal JSON format adds the ability to retain field metadata such as mysqlType and sqlType, and supports DDL synchronization. This optimization helps customers who use Canal JSON to ingest Kafka data sources convey table structure and field type information more completely, reducing adaptation work in downstream parsing and schema evolution, and improving the compatibility, stability, and O&M efficiency of real-time data synchronization pipelines.
Realtime Compute for Apache Flink - OSS Supports Custom Roles via AssumeRole
The OSS connector adds the ability to use custom RAM roles via AssumeRole. Users can configure custom roles for jobs, and the system obtains temporary access credentials to access OSS, eliminating the need to store AK/SK in job configurations long term. This capability reduces the exposure risk of long-term keys and supports more flexible least-privilege control by business, environment, and data access scope.
E-MapReduce - EMR Agent Supports API Invocation
Provides APIs for basic conversations, alert diagnosis, scheduled inspections, and more, helping customers integrate intelligent O&M into their business platforms and apply it to a wider range of scenarios.
Elasticsearch - International Site Supports Single-AZ to Multi-AZ Configuration Changes
The international site now supports changing Elasticsearch instances from a single availability zone to multiple availability zones, helping customers improve cross-AZ disaster recovery capabilities and business availability.
Elasticsearch - ES Agent Intelligent Assistant GA Release
ES Agent is the intelligent O&M assistant provided by Alibaba Cloud Elasticsearch. After service capability and stability verification during the public beta phase, it is officially launched for commercial release.
OpenSearch - Index Data Recovery Now Available for Vector Search Edition
Vector Search Edition now supports recovering data from indexes and rebuilding indexes from recovery records, resolving the previous gap where customers who needed to rebuild an index could only rebuild from empty data.
OpenSearch - Industry Algorithm Edition Supports Time-Based Elastic Scaling
Users preconfigure time rules and target LCUs, and the system automatically scales out or scales in the compute resources (measured in LCUs) of OpenSearch instances at the scheduled times, then automatically restores them after the task completes.
OpenSearch - Industry Algorithm Edition Adds a Shared Cluster in the China (Hong Kong) Region
Industry Algorithm Edition adds a shared cluster in the China (Hong Kong) region, supporting both subscription and pay-as-you-go billing.
OpenSearch - Industry Algorithm Edition Supports the Drop-Down Suggestion Model
OpenSearch - Vector Search Edition Supports the Video Search Pipeline Feature
Vector Retrieval Service for Milvus - AI Center Fully Upgraded
The AI Center of Vector Retrieval Service for Milvus has been upgraded into a unified AI capability portal. In addition to the existing embedding models, it now supports reranking, text generation, and multimodal understanding models, with key parameter descriptions and code examples provided to help you complete model selection and integration more efficiently.
Vector Retrieval Service for Milvus - Managed Auto Scaling
Vector Retrieval Service for Milvus adds a managed auto scaling policy. Simply set the minimum and maximum node counts for supported node types, and the system automatically scales out or in within the configured range based on real-time resource load — no need to configure scaling schedules, load thresholds, or cooldown periods yourself. This policy is suitable for business scenarios with significant and hard-to-predict traffic fluctuations, such as online RAG and multi-tenant search.
Vector Retrieval Service for Milvus - Auto Scaling Expands Node Support
Auto scaling for Vector Retrieval Service for Milvus now supports elastic scale-out and scale-in for Proxy, Streaming Node, and Query Node. You can adjust the node counts for access traffic, streaming data processing, and query loads respectively, allowing different processing pipelines to scale independently according to their own business pressure.
Alibaba Cloud Hologres Tops the Official TPC-H 3,000GB Benchmark, with the Best Price/Performance
27 posts | 0 followers
FollowAlibaba Cloud Big Data and AI - May 15, 2026
Alibaba Cloud Big Data and AI - June 11, 2026
Alibaba Cloud Big Data and AI - July 17, 2026
Alibaba Cloud Big Data and AI - April 13, 2026
CloudSecurity - August 10, 2026
CloudSecurity - June 15, 2026
27 posts | 0 followers
Follow
Big Data Consulting for Data Technology Solution
Alibaba Cloud provides big data consulting services to help enterprises leverage advanced data technology.
Learn More
MaxCompute
Conduct large-scale data warehousing with MaxCompute
Learn More
Big Data Consulting Services for Retail Solution
Alibaba Cloud experts provide retailers with a lightweight and customized big data consulting service to help you assess your big data maturity and plan your big data journey.
Learn More
E-MapReduce Service
A Big Data service that uses Apache Hadoop and Spark to process and analyze data
Learn MoreMore Posts by Alibaba Cloud Big Data and AI
