Jim Hughes created FLINK-39729:
----------------------------------
Summary: CorrelateRestoreTest.testRestore[8]
(CORRELATE_WITH_LITERAL_AGG) is flaky due to non-deterministic changelog
ordering
Key: FLINK-39729
URL: https://issues.apache.org/jira/browse/FLINK-39729
Project: Flink
Issue Type: Bug
Components: Table SQL / Planner
Reporter: Jim Hughes
Assignee: Jim Hughes
`CorrelateRestoreTest.testRestore[8]` (`CORRELATE_WITH_LITERAL_AGG`) is
intermittently failing because the test asserts exact changelog record
ordering, but the multi-source streaming plan produces non-deterministic record
interleaving after restore.
The plan has three independent sources feeding into two streaming joins
(InnerJoin, LeftOuterJoin, SemiJoin). Post-restore, when `source_t2`'s record
reaches the SemiJoin before the retract cascade from the
GroupAggregate+InnerJoin pipeline completes, the `-D[3]` and `+I[3]` changelog
records are never emitted in the expected window.
The test was introduced in FLINK-37882.
{code}
Expecting actual:
["+I[2]", "-D[2]", "+I[2]", "+I[3]"]
to contain exactly in any order:
["+I[2]", "+I[3]", "-D[2]", "-D[3]", "+I[2]", "+I[3]"]
but could not find the following elements:
["-D[3]", "+I[3]"]
{code}
*Fix:* Switch from changelog assertion to materialized data assertion using the
existing `testMaterializedData()` + `consumedValues()` pattern (already used by
`MultiJoinTestPrograms` and `DeltaJoinTestPrograms` for the same class of
non-determinism). The expected materialized state is `[+I[2], +I[3]]`.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
