All Products
Search

This Product

    Realtime Compute for Apache Flink:June 3, 2026 release

    Document Center

    Realtime Compute for Apache Flink:June 3, 2026 release

    Last Updated:Jun 22, 2026

    Major feature changes and key bug fixes in Realtime Compute for Apache Flink released on June 3, 2026.

    Important

    This upgrade is rolled out across regions in stages. For the rollout schedule, check the latest announcements on the right side of the Realtime Compute for Apache Flink console. If new features are unavailable, your account has not yet been upgraded. To request an expedited upgrade, submit a ticket and we will arrange the upgrade based on your situation.

    Overview

    Realtime Compute for Apache Flink releases engine version VVR 11.7 (based on Apache Flink 1.20.3). This release strengthens security, AI real-time inference, vector search, semi-structured data processing, and data ingestion. The SQL built-in function call chain is hardened, and AI Function now supports multimodal models, CREATE MODEL parameter passthrough, and remote inference services. DLF Paimon vector table search is available, and Variant and DECIMAL handling is improved. For data ingestion, Kafka YAML Source, MySQL CDC, SQL Server CDC, Canal JSON, and Transform expressions are refined. Core connectors including Elasticsearch, Iceberg, Hologres, MySQL, and Kafka are improved in stability, authentication, connection management, and usability. This release also incorporates Apache Flink community improvements and fixes for Async Function and AsyncScalarFunction.

    Engine

    Enhancements to security, AI inference, vector search, the type system, and APIs strengthen real-time data processing, intelligent analytics, and lakehouse vector search.

    AI Function and model inference

    • Flink Agents development support: Build event-driven streaming AI Agent jobs based on the open-source Apache Flink Agents framework.

    • NVIDIA Triton Inference Server support: Remote calls to NVIDIA Triton Inference Server are now supported, expanding integration options for online inference services.

    • Multimodal model invocation improvements: Improved parameter adaptation for image and video understanding models makes AI Function easier to use with visual understanding models.

    • ml_predict function enhancement: The ml_predict function can now reuse selected model parameters from CREATE MODEL, reducing duplicate configuration in inference SQL.

    SQL and type system

    • DECIMAL precision enhancement: The DECIMAL type now supports higher precision for numeric computation scenarios (experimental).

    • Paimon Variant Shredding support: Improves the storage and query efficiency of semi-structured data in Paimon.

    • Python DataStream API enhancements: Incorporates Apache Flink community improvements for Async Function in the Python DataStream API.

    • AsyncScalarFunction improvements: Incorporates Apache Flink community fixes and enhancements for AsyncScalarFunction, improving stability and usability of asynchronous scalar functions.

    Security hardening

    • SQL built-in function hardening: Fixed a potential code injection risk in the SQL built-in function call chain, improving job runtime robustness.

    Data ingestion (Flink CDC)

    • Kafka YAML Source enhancements:

      • JSON format now tolerates duplicate keys in Key or Value fields, improving compatibility when ingesting complex JSON data.

      • Tombstone messages (with empty Value) can now be converted to DELETE events, supporting log compaction topics and delete-semantics synchronization.

    • Canal JSON format enhancement: Kafka Canal JSON data now supports JSON Converter, providing more flexible parsing and type conversion.

    • MySQL CDC optimizations:

      • Improved field name case handling for better correctness in case-sensitive synchronization scenarios.

      • When reading RDS archived logs, the primary instance ID can now be looked up automatically, reducing configuration overhead.

    • SQL Server CDC Source: Adds CDC support for SQL Server, enabling real-time change data capture from SQL Server (public preview).

    • Transform expression enhancement: High-precision DECIMAL expression handling has been improved, increasing computation correctness in data transformation scenarios.

    • Hologres YAML Sink: Hologres now supports the legacy Put Handler, improving compatibility with existing write paths.

    • Paimon YAML Sink: Supports writing Blob fields stored via the blob-descriptor-field descriptor.

    Connectors

    • Elasticsearch: Adds the connection.keep-alive parameter for configurable connection keep-alive, improving stability for long-lived connections.

    • Kafka: The YAML Source connector now supports JSON Converter for Canal JSON data, with improved handling of duplicate keys and Tombstone messages.

    • Paimon: Added the vector_search stored procedure for querying vector tables in Paimon. For usage details, see support vector search procedure for flink.

    Bug fixes

    • Correctness fixes:

      • Fixed insufficient DECIMAL type inference precision in Transform expressions.

      • Fixed unexpected field name case handling in Flink CDC MySQL.

      • Fixed compatibility issues in Kafka YAML Source when processing JSON data containing duplicate keys.

    • Stability fixes:

      • Fixed an issue with MySQL BinlogOffset comparison logic.

      • Fixed a CREATE CATALOG DDL error for Paimon DLF 1.0.

      • Fixed a MySQL CDC error: Variable character string length must be between 1 and 2147483647.

      • Fixed a default-value handling error when Flink CDC YAML jobs write Kafka Debezium JSON data.