[
https://issues.apache.org/jira/browse/FLINK-40552?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Yanquan Lv resolved FLINK-40552.
--------------------------------
Fix Version/s: cdc-3.7.0
Assignee: Mao Jiayi
Resolution: Fixed
Merged in master via 7d560004bcbbd644ff2411b693e4f87f1646080a.
> Support ordered asynchronous execution for YAML Transform
> ---------------------------------------------------------
>
> Key: FLINK-40552
> URL: https://issues.apache.org/jira/browse/FLINK-40552
> Project: Flink
> Issue Type: Improvement
> Components: Flink CDC
> Reporter: Mao Jiayi
> Assignee: Mao Jiayi
> Priority: Major
> Labels: pull-request-available
> Fix For: cdc-3.7.0
>
>
> Currently, the YAML PostTransform operator processes DataChangeEvents
> synchronously.
> Transform expressions may invoke external AI model services or blocking UDFs.
> In such cases, every record has to wait for the previous invocation to
> finish, which limits pipeline throughput.
> This issue proposes an optional ordered asynchronous execution mode for
> PostTransform.
> The implementation will:
> * Extract the common PostTransform processing logic into a processor shared
> by synchronous and asynchronous execution.
> * Use AsyncDataStream.orderedWait for asynchronous execution.
> * Process DataChangeEvents concurrently while preserving their input order.
> * Treat SchemaChangeEvents as barriers:
> ** Wait for all preceding DataChangeEvents to finish.
> ** Apply the schema change.
> ** Process subsequent DataChangeEvents with the updated schema.
> * Add the following pipeline options:
> ** transform.async-execution.enabled
> ** transform.async-execution.timeout
> ** transform.async-execution.capacity
> ** transform.async-execution.worker-threads
> * Preserve the current synchronous behavior when asynchronous execution is
> disabled.
> * Persist the PostTransform schema view for checkpoint and savepoint recovery.
> * Re-emit the latest CreateTableEvent once per table before replayed
> DataChangeEvents after recovery.
> * Propagate timeout and transform execution exceptions to the Flink job.
> The initial implementation will support checkpoint recovery and savepoint
> recovery with unchanged parallelism. Restoring an asynchronous transform from
> a savepoint with different parallelism is out of scope.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
