nielifeng commented on issue #10997:
URL: https://github.com/apache/seatunnel/issues/10997#issuecomment-5086473326
I completed a controlled validation of this issue on current `dev` commit
`dbead9b` using a separated Zeta Master/Worker deployment.
### Patch direction
The candidate fix changes only `ReportMetricsOperation` scheduling:
```java
public class ReportMetricsOperation extends TracingOperation
implements IdentifiedDataSerializable, Offloadable {
@Override
public String getExecutorName() {
return Offloadable.OFFLOADABLE_EXECUTOR;
}
}
```
The payload, serializer ID, invocation semantics, metrics storage, and
`SeaTunnelServer.updateMetrics()` behavior remain unchanged.
### Results
- A focused regression confirms that `ReportMetricsOperation` is offloaded
from
Hazelcast generic-operation threads and uses the built-in offloadable
executor.
- The existing metrics IMap integration path still stores metrics correctly.
- Matched unpatched and patched bounded-batch runs completed without metrics
failure/interruption, failed jobs, or persistent pending jobs.
- A controlled reduced generic-operation-thread comparison also completed on
both versions; the patched run kept the metrics operation off generic
threads.
- Both versions reached the same workload-capacity knee at a higher
submission
rate, so that behavior was treated as capacity evidence rather than a
product
failure.
- A longer patched run remained healthy before a separately planned Worker
fault.
The original production-style permanent RUNNING hang was **not directly
reproduced** in this lab environment. Therefore, the cluster result is
directional evidence for isolating metrics work from control-plane generic
threads, not a definitive before/after reproduction of the reported hang.
The planned Worker restart exposed an independent checkpoint recovery NPE in
`CheckpointCoordinator.isNoErrorCompleted()`. It is unrelated to
`ReportMetricsOperation` offloading and is tracked separately in #11563.
Local targeted tests for the metrics offload and the independent checkpoint
regression pass. No PR has been submitted yet.
--
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]
