[
https://issues.apache.org/jira/browse/FLINK-33224?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18113948#comment-18113948
]
Sylwester Lachiewicz commented on FLINK-33224:
----------------------------------------------
Triage finding (2026-09-10): the exception in the original 1.14.6 report no
longer exists on any maintained connector line; the bounded path in
{{KafkaPartitionSplitReader}} removes empty splits up front
({{removeEmptySplits}}) and {{testAssignEmptySplit}} is green on main. The
remaining symptom, the job staying up because the final MAX_WATERMARK is never
emitted for a source that only saw empty partitions, is the runtime issue
[FLINK-40499|https://issues.apache.org/jira/browse/FLINK-40499] that Martijn
linked, with the fix pending in [apache/flink PR
#29037|https://github.com/apache/flink/pull/29037].
Related reports checked at the same time:
[FLINK-33001|https://issues.apache.org/jira/browse/FLINK-33001] was fixed by
[06789ec7|https://github.com/apache/flink-connector-kafka/commit/06789ec7b86a3b2edae1dc7f9f29d672ebd3610b]
in 3.1.0, and [FLINK-35565|https://issues.apache.org/jira/browse/FLINK-35565]
by [FLINK-34470|https://issues.apache.org/jira/browse/FLINK-34470] in 3.3.0; a
new unit test for an out-of-range starting offset reset onto the stopping
offset stays green on main.
Recommendation: nothing to change in the connector. Close this once FLINK-40499
ships, or move it to the runtime component now.
_This comment was created with AI assistance._
> Flink Batch Job does not complete when using Kafka topics with empty
> partitions
> --------------------------------------------------------------------------------
>
> Key: FLINK-33224
> URL: https://issues.apache.org/jira/browse/FLINK-33224
> Project: Flink
> Issue Type: Bug
> Components: Connectors / Kafka
> Affects Versions: 1.14.6
> Reporter: Kanthi Vaidya
> Priority: Major
>
> Kafka Source Connector no longer throws an exception when running a flink job
> in batch mode when using Kafka topics with empty partitions. However, flink
> does not emit the final watermark of Long.MAX_VALUE even if all the messages
> have been read. Note that we are using bounded sources, since that is a
> requirement for batch mode. It appears that the Kafka consumer continues to
> run even though the source is bounded.
> Same is the case with an empty topic. We do have optional sources, where data
> need not be present all the time. However if we mark the source as bounded,
> the Kafka connector should recognize empty partitions and mark it as
> completed.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
