[
https://issues.apache.org/jira/browse/FLINK-39980?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bowen Li resolved FLINK-39980.
------------------------------
Release Note:
DynamicKafkaSource under `global` enumerator mode can now rebalance splits upon
recovery.
Flink 2.2.1 provides SupportsSplitReassignmentOnRecovery so a source can report
restored reader splits to its enumerator before assigning them. This change
adopts that contract and lets DynamicKafkaSourceEnumerator redistribute the
active restored split set during recovery.
This can only work with Flink 2.2.1 or above.
Resolution: Implemented
> Implement split rebalance/reassignment support in DynamicKafkaSource after
> metadata changes
> -------------------------------------------------------------------------------------------
>
> Key: FLINK-39980
> URL: https://issues.apache.org/jira/browse/FLINK-39980
> Project: Flink
> Issue Type: New Feature
> Components: Connectors / Kafka
> Affects Versions: 2.3.0
> Reporter: Bowen Li
> Assignee: Bowen Li
> Priority: Major
> Labels: pull-request-available
>
> DynamicKafkaSource can remove clusters/topics when stream metadata changes.
> After removal, some source subtasks may become empty while other subtasks
> still own active partitions. This can happen even when {{{}parallelism <=
> total active partitions{}}}, because the current assignment is inherited from
> prior split ownership and does not rebalance active splits across the
> remaining readers.
> Flink has added a source contract for enumerators to regain a global split
> view and perform reassignment/rebalancing after recovery:
> {{apache/flink#27149}} / FLIP-537. DynamicKafkaSource should implement this
> contract once it is available in our Flink runtime.
> *Goals*
> Implement the new rebalance/reassignment interface in DynamicKafkaSource so
> that after metadata removal or recovery, active splits are redistributed
> across available readers.
> *Expected Behavior*
> When active Kafka partitions remain and {{{}parallelism <= total active
> partitions{}}}, DynamicKafkaSource should avoid empty source subtasks by
> rebalancing active splits across readers.
> When {{{}total active partitions < parallelism{}}}, empty subtasks are
> expected, but they should be handled by idleness logic and autoscaler
> calibration.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
