This topic describes the major feature changes and bug fixes released for Realtime Compute for Apache Flink on August 20, 2026.
This release is being rolled out in phases across the network. For upgrade progress, check the latest announcements in the right pane of the Realtime Compute console. If the described capabilities are not yet available in your account, the grayscale upgrade has not reached your account yet. To expedite the upgrade, submit a ticket with your business requirements, and we will schedule accordingly.
Overview
This release focuses on the launch of a dedicated Flink AI Native Ops entry, a commercialization closed loop for AI capabilities, and improvements to the job health score logic. A natural-language interaction entry on par with the console now unifies intelligent diagnostics, intelligent inspection, and skill management, upgrading operations from reactive response to proactive prevention and from scattered operations to a single entry. OpenAPI and Skill capabilities for activating, querying, and closing AI products are added to support automated management and agent integration for users. Inspection policies support natural-language generation and form-based configuration, and skills are displayed as cards. In addition, the health score introduces a time-window mechanism, fixing the problem where accumulated historical risks kept scores from rising and misled troubleshooting, so the score reflects the current health of a job.
Platform capabilities
AI capabilities
Dedicated Flink AI Native Ops entry
-
Unified entry: Intelligent diagnostics, intelligent inspection, skill management, and token usage are presented together on a dedicated product page, providing a natural-language interaction entry on par with the console.
-
Natural-language driven: Single-step operations and complex cross-module tasks can be orchestrated automatically from natural-language instructions, allowing O&M operations to be executed directly in natural language.
-
Enhanced entity referencing: Catalog metadata, drafts, deployments, and session clusters can be referenced as entities, covering scenarios such as metadata query understanding and SQL code generation.
Intelligent inspection upgrade
-
Inspection switch: Automatic inspection tasks can be turned on or off.
-
Trigger methods: Manual triggers and scheduled inspections are supported. Inspections can target all workspaces, workspaces filtered by tag, or specified deployments.
-
Inspection configuration: Describe inspection requirements in natural language and let AI generate the configuration, or use forms to customize the cron frequency, execution time (hour granularity), and inspection scope (all, by tag, or specified deployments).
OpenAPI and Skill
AI product lifecycle management:
-
OpenAPI: Standard operations for activating (
CreateAiService), querying (DescribeAiService), and closing (CloseAiService) AI service products support automated management and operations orchestration for batches of customers, with no manual per-account operations in the console. -
Skill: Corresponding Skill wrappers support integration with user agents, completing the API-based closed loop of the commercialization process.
Experience improvements
Faster first-screen rendering of inspection daily reports
Fixed an issue where the first screen of an inspection daily report froze during rendering in scenarios with a massive number of deployments. Report pages now open faster and respond more smoothly with large data volumes.
Bug fixes
Health score logic optimization
Fixed an issue where health scores were calculated from accumulated risks, so scores only dropped and never recovered, and historical risks misled troubleshooting. Risk counts now use a sliding window, and scores recover automatically after the window expires, with no manual restart required.