Sylwester Lachiewicz created FLINK-40140:
--------------------------------------------
Summary: Extract reader recovery gating from
DynamicKafkaSourceEnumerator into a ReaderRecoveryCoordinator
Key: FLINK-40140
URL: https://issues.apache.org/jira/browse/FLINK-40140
Project: Flink
Issue Type: Improvement
Components: Connectors / Kafka
Reporter: Sylwester Lachiewicz
DynamicKafkaSourceEnumerator (~1,200 lines) accumulated three interacting
pieces of recovery gating state with FLINK-39980:
initialReaderRegistrationPending, pendingReportedSplitsByReader, and
pendingMetadataUpdateReaders. The gating semantics (when to defer split
assignment and metadata update events during recovery) are implicit in field
interactions spread across several methods, and cannot be tested without
driving the whole enumerator.
Proposal: extract this state into a dedicated ReaderRecoveryCoordinator
collaborator with intent-revealing methods (hasPendingRecovery,
shouldDeferMetadataUpdateEvents, drainReportedSplits,
drainDeferredMetadataUpdateReaders) covered by direct unit tests. No behavior
change intended; orchestration stays in the enumerator.
This is preparatory cleanup toward further decomposition of the enumerator
(e.g. bounded-mode no-more-splits termination signaling), whose coupling
contributed to the idle-reader hang observed in
SourceTestSuiteBase#testIdleReader.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
