You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[test] Unstable test FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath #4233
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.899 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.778 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.842 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 69.013 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.845 s <<< FAILURE!
java.lang.AssertionError:
I investigated the recent occurrences of #4233 using the detailed CI log
artifact. waitUntilTableReady() can narrow one trigger window, but the log
appears to expose a broader issue in the tiering enumerator.
The previous tiering job claimed the newly created table, generated zero
splits, and moved the table from tieringTableEpochs to finishedTables.
The enumerator was then closed before the next heartbeat. close() reports
the remaining active tables as failed, but it does not flush finishedTables, so the coordinator kept this table in the Tiering state
until the service timeout. The next tiering job consequently reported
"No available Tiering table found" for the entire test timeout.
The same lost-completion window can also apply to a normal FinishedTieringEvent, so this looks like a production liveness issue rather
than only a table-readiness race in this test.
Would you be interested in extending #4235 to flush pending completion reports
during enumerator shutdown and add a deterministic enumerator regression test?
I am happy to share the detailed timeline and test approach. If you would
prefer, I can prepare a separate root-cause fix and link it here.
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.854 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.813 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.887 s <<< FAILURE!
java.lang.AssertionError:
[ERROR] org.apache.fluss.lake.paimon.flink.FlinkUnionReadPrimaryKeyTableITCase.testUnionReadWithCustomLakeTablePath Time elapsed: 68.867 s <<< FAILURE!
java.lang.AssertionError:
Activity
ChengJie1053 commentedon Sep 4, 2026
I would like to work on this issue.
github-actions commentedon Sep 9, 2026
https://github.com/apache/fluss/actions/runs/34273236558/job/102219864609
github-actions commentedon Sep 16, 2026
https://github.com/apache/fluss/actions/runs/34958538719/job/104346331356
github-actions commentedon Sep 21, 2026
https://github.com/apache/fluss/actions/runs/35534668328/job/106141431587
github-actions commentedon Sep 22, 2026
https://github.com/apache/fluss/actions/runs/35602925512/job/106343071489
Jackeyzhe commentedon Sep 22, 2026
Hi @ChengJie1053
I investigated the recent occurrences of #4233 using the detailed CI log
artifact.
waitUntilTableReady()can narrow one trigger window, but the logappears to expose a broader issue in the tiering enumerator.
The previous tiering job claimed the newly created table, generated zero
splits, and moved the table from
tieringTableEpochstofinishedTables.The enumerator was then closed before the next heartbeat.
close()reportsthe remaining active tables as failed, but it does not flush
finishedTables, so the coordinator kept this table in the Tiering stateuntil the service timeout. The next tiering job consequently reported
"No available Tiering table found" for the entire test timeout.
The same lost-completion window can also apply to a normal
FinishedTieringEvent, so this looks like a production liveness issue ratherthan only a table-readiness race in this test.
Would you be interested in extending #4235 to flush pending completion reports
during enumerator shutdown and add a deterministic enumerator regression test?
I am happy to share the detailed timeline and test approach. If you would
prefer, I can prepare a separate root-cause fix and link it here.
github-actions commentedon Sep 24, 2026
https://github.com/apache/fluss/actions/runs/35914314872/job/107361806459
github-actions commentedon Sep 25, 2026
https://github.com/apache/fluss/actions/runs/35966673950/job/107526705686
github-actions commentedon Oct 13, 2026
https://github.com/apache/fluss/actions/runs/38302686028/job/114989180188
[flink] Flush finished tiering tables when closing enumerator
github-actions commentedon Oct 16, 2026
https://github.com/apache/fluss/actions/runs/38700031652/job/116305719076