Jiabao Sun created FLINK-41565:
----------------------------------

             Summary: Prevent MongoDB sink cleanup failures from terminating 
the TaskManager in Flink 2.2.x
                 Key: FLINK-41565
                 URL: https://issues.apache.org/jira/browse/FLINK-41565
             Project: Flink
          Issue Type: Bug
          Components: Connectors / MongoDB
    Affects Versions: 2.2.1, 2.2.0
            Reporter: Jiabao Sun
            Assignee: Jiabao Sun


Since FLINK-38180, Flink treats exceptions thrown during the cleanup of an 
already failed task as fatal, potentially terminating the entire TaskManager.

Currently, MongoWriter.close() attempts to flush any remaining buffered records 
to MongoDB and propagates exceptions if the write fails. When a task has 
already failed due to a MongoDB write error (e.g., duplicate key, document 
validation failure, or server unavailability), the subsequent flush during 
close() may encounter the same error again.

With Flink 2.2, this cleanup exception can cause the TaskManager to terminate 
instead of allowing the task failure to be handled through the normal recovery 
mechanism.

This introduces a behavioral regression compared to Flink 1.19/1.20. Although 
no released version is currently affected, the issue needs to be addressed 
before the MongoDB connector's first Flink 2.x release.

*Expected Behavior*

{{MongoWriter.close()}} should handle failures during its final flush without 
propagating exceptions that could terminate the TaskManager.

Such failures should be logged at ERROR level to ensure they remain visible for 
diagnostics.

The behavior should also account for the configured delivery guarantee:
 * {{{}AT_LEAST_ONCE{}}}: Records are flushed during checkpointing and normal 
end-of-input processing. Records remaining after a failed or cancelled task can 
be replayed from the last successful checkpoint.

 * {{{}NONE{}}}: The final flush during {{close()}} may be the only attempt to 
write the remaining buffered records, particularly when stopping with a 
savepoint without draining. Suppressing an exception in this case may result in 
record loss, which must be clearly documented and logged.

*Proposed Changes*
 # Update {{MongoWriter.close()}} to catch and log MongoDB write exceptions 
without rethrowing them during cleanup.

 # Document the implications for different delivery guarantees.

 # Add an integration test in {{MongoWriterITCase}} that triggers a MongoDB 
write failure and verifies that {{close()}} returns normally without 
propagating the exception.

 # Add a release note describing the change in cleanup failure handling.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to