88fantasy opened a new issue, #12386:
URL: https://github.com/apache/seatunnel/issues/12386
### Search before asking
- [x] I searched the issues and found no similar issues. (#12139 is a
different restore hang.)
### What happened
With two master nodes, when the active master dies and the standby takes
over, a running MySQL-CDC job with `parallelism > 1` is restored as **RUNNING
but never triggers another checkpoint**. Source and sink tasks keep processing
(`SourceReceivedCount` / `SinkWriteCount` grow), but `SinkCommittedCount` stops
because the 2PC sink only commits on checkpoint completion. The job silently
freezes until some task fails and the pipeline is restarted.
### Root cause
On master active switch `SubPlan` restores the task execution states from
the IMap and reports
```java
jobMaster.getCheckpointManager().reportedPipelineRunning(pipelineId,
allTaskRunning.get());
```
where `allTaskRunning` is true only if **every** coordinator and physical
vertex is `RUNNING`.
For MySQL-CDC with parallelism 2, after the snapshot phase only one reader
keeps consuming the binlog and the other reader becomes idle and **FINISHED**
(log on the new master: `SourceTask (2/2) ... is in state FINISHED when init
state future`). So `alreadyStarted = false`,
`CheckpointCoordinator#restoreCoordinator(false)` sets `isAllTaskReady = false`
and waits for all tasks to report `READY_START`. The running tasks are not
redeployed, so they never report `READY_START` again, `allTaskReady()` never
passes and `scheduleTriggerPendingCheckpoint` is never called.
Control experiment: the same job with `parallelism = 1` (no idle-finished
subtask) is restored with `alreadyStarted: true` and checkpoints continue
immediately after the takeover.
### How to reproduce
1. Separated mode with 2 masters + 3 workers, checkpoint storage + IMap
persistence enabled.
2. Submit MySQL-CDC -> Doris (2PC) streaming job with `parallelism = 2`,
wait until the snapshot is done (one source subtask becomes FINISHED).
3. `kill -9` the active master.
4. Standby logs `This node become a new active master node`, `begin restore
job`, then `received restore CheckpointCoordinator with alreadyStarted: false`,
and no further `wait checkpoint id: N completed` lines. The job stays RUNNING,
`SinkCommittedCount` stays constant.
### Suggested fix (for discussion)
Treat subtasks that were closed as idle (tracked in `readyToCloseIdleTask` /
`FINISHED` after the reader signalled no more splits) as "started" when
computing `allTaskRunning`, or let `restoreCoordinator(false)` re-request
`READY_START` from tasks that are still running. Happy to work on a PR once the
direction is agreed.
### SeaTunnel Version
dev (`c7304ace6`, 2026-09-17); also reproduced with the fix for #12385
applied (state is then loaded correctly, but checkpointing still does not
resume).
### SeaTunnel Config
```conf
env { job.mode = "STREAMING", parallelism = 2, checkpoint.interval = 10000 }
source { MySQL-CDC { ... } }
sink { Doris { sink.enable-2pc = "true", ... } }
```
### Running Command
```shell
bin/seatunnel.sh --config cdc.conf --async
```
### Error Exception
```log
received restore CheckpointCoordinator with alreadyStarted: false
(no checkpoint triggered afterwards)
```
### Zeta or Flink or Spark Version
Zeta, separated cluster mode, 2 masters
### Java or Scala Version
JDK 8 (1.8.0)
### Screenshots
_No response_
### Are you willing to submit PR?
- [x] Yes I am willing to submit a PR!
### Code of Conduct
- [x] I agree to follow this project's Code of Conduct
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
