hwanju edited a comment on pull request #14499:
URL: https://github.com/apache/flink/pull/14499#issuecomment-761572560
> I believe sources are relatively important, in particular because some
users have their own implementations there.
> @hwanju do you want to address this in this PR, or in a follow up?
This is nice catch! This should be addressed as part of this PR. When this
patch was written, the old Flink version did invoke `run()` in `invoke()` in
the same thread, but now with mailbox approach, it spawns a separate thread for
run(). Yes, `FlinkSecurityManager` supports thread-inherited monitoring flag,
so why this spawned thread wasn't protected? The reason is the thread local
variables are inherited at the time of thread construction, not thread run, and
the construction happens in `loadAndInstantiateInvokable` before `invoke`. By
wrapping `loadAndInstantiateInvokable` with exit monitoring, I see it's
protected as follows:
```
org.apache.flink.runtime.UserSystemExitException: Flink user code attempted
to exit JVM.
at
org.apache.flink.runtime.security.FlinkSecurityManager.checkExit(FlinkSecurityManager.java:184)
at java.lang.Runtime.exit(Runtime.java:107)
at java.lang.System.exit(System.java:971)
at
org.apache.flink.streaming.tests.FailureTestSource.run(FailureTestSource.java:65)
at
org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:110)
at
org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:66)
at
org.apache.flink.streaming.runtime.tasks.SourceStreamTask$LegacySourceFunctionThread.run(SourceStreamTask.java:245)
```
Added unit test (reproduced without fix, and fixed after it).
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
