Mate Czagany created FLINK-41254:
------------------------------------
Summary: Two-phase window aggregation loses buffered records when
a source subtask becomes idle
Key: FLINK-41254
URL: https://issues.apache.org/jira/browse/FLINK-41254
Project: Flink
Issue Type: Bug
Components: Table SQL / Runtime
Reporter: Mate Czagany
Two-phase event-time window aggregation produces wrong results when a source
subtask receives records and then becomes idle.
LocalSlicingWindowAggOperator holds pre-aggregates in-memory and flushes it
only when its subtask's watermark passes the slice end, or in
prepareSnapshotPreBarrier. When the chained source emits WatermarkStatus.IDLE,
the
subtask's watermark stops advancing, so the buffer is never flushed. The global
operator excludes the idle channel, advances on the remaining subtasks, and
fires the window.
More info can be found here:
[https://lists.apache.org/thread/128qfwtdgtj6yfhftkj7r15polqqz6mn]
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
