Realtime Compute for Apache Flink:2026-07-27 release

Multimodal AI Function support: SQL AI Function now supports multimodal data processing scenarios such as audio and video, with improved capabilities and end-to-end test coverage for image, video, and audio operators.

Built-in model registration exemption: When using built-in models provided by the platform's master account, model registration is no longer required, lowering the barrier to model invocation.

Token usage governance: SQL AI Function and PyFlink support the max-total-tokens parameter to limit the total tokens consumed in a single job run, helping control model invocation cost.

External cache support: AI Function supports external storage as a cache, improving processing efficiency for repeated inference requests.

DataFrame API enhancements: DataFrame API (Stage 3) is enhanced to support multimodal data sources, large model invocation, and multimodal operators. New capabilities include namespace-style accessor and read_video_frames, improving the video data processing experience.

Flink Agents integration: Flink Agents can now be invoked as table-valued functions (TVF) in SQL jobs, adapted to Flink Agents 0.3.

DLF Paimon vector search: vector_search and vector_search_agg support searching DLF Paimon vector tables, enhancing the integration between lakehouse data and vector search scenarios.

Lookup Join timeout handling: Lookup Join based on AsyncLookupFunction and AsyncTableFunction supports user-defined timeout handling logic, providing better control over async query scenarios.

Improvements and bug fixes from Apache Flink 1.20.4 are incorporated to continuously improve job correctness, stability, and development experience.

Canal JSON metadata and DDL: Metadata such as mysqlType and sqlType is preserved, and DDL change synchronization is supported, improving compatibility with the Canal protocol.

Emit changed fields only: Kafka Canal JSON output supports keeping only the changed fields, reducing downstream processing overhead.

Custom formats and type inference optimization: The Kafka connector supports custom formats. For Canal JSON, Debezium JSON, and JSON data, type inference avoids automatically attempting to convert strings to time types.

Kafka idempotence configuration validation: For specific cloud Kafka scenarios, jobs detect incompatible idempotence settings and provide clear prompts, reducing configuration errors.

MySQL CDC parameter validation: Debezium parameter validation is standardized to improve visibility of configuration issues.

Binlog downstream propagation control: MySQL Binlog propagation to downstream operators can now be controlled on demand through configuration.

Source rate limiting: MySQL Source supports rate limiting based on metrics such as numRecordsInOfSourcePerSecond, helping control downstream processing pressure.

SQL Server CDC: Data ingestion now supports SQL Server Source with the op_type metadata column for identifying change operation types.

StarRocks Sink enhancements: Supports the sink.ignore-update-before parameter.

Paimon skips commit on empty tables: When a table has no data written, its commit can be skipped to reduce unnecessary commit overhead.

Paimon DLS Catalog: The Filesystem Catalog now supports DLS, extending lakehouse deployment and access scenarios.

Source-to-Sink distribution strategy: Data distribution from Source to Sink can be configured in strategy mode, improving configurability for complex topologies.

YAML job Autopilot: YAML jobs support parallelism updates through Autopilot, improving elastic tuning.

Schema change compatibility: The compatibility of various Sinks with schema changes is reviewed and improved, facilitating job design and change assessment.

KMS integration framework: KMS integration capabilities are added or improved on the connector side, making it easier to centrally manage credentials for databases and other external systems. The PostgreSQL CDC connector supports KMS integration.

Credential-free connectors: Framework-level credential-free access is improved, with support for scenarios such as credential-free access via the Fluss connector.

OSS AssumeRole: OSS supports using a custom role via AssumeRole instead of AccessKey/SecretKey, improving the security and flexibility of cloud access.

SSL/TLS certificate support: Connectors support configuring custom SSL/TLS certificate file paths through parameters. For scenarios where independent certificate parameters are not supported, merged certificates can be used.

MySQL Sink batch write: The MySQL connector supports batch inserts when writing to tables without a primary key, improving write throughput.

MySQL delete handling: MySQL Sink supports a mode that ignores DELETE events, meeting specific synchronization strategies.

Hologres Sink optimization: When ignore-null is used via expressions, client-side deduplication is supported, improving write pipeline performance and result consistency.

Redis connection pool optimization: The Redis connector connection pool implementation is optimized, improving stability in high-concurrency connection scenarios.

Iceberg OSS client parameters: The Iceberg connector supports tuning OSS Client parameters, providing more configurability for object storage access.