Github user StefanRRichter commented on the issue:
https://github.com/apache/flink/pull/5934
Yes, there is a theoretical problem if the serialization would not be
threadsafe. I think currently the silent assumption that holds is that
serializers are immutable w.r.t. serialization after they have been created and
passed to create a state. I think Stephan's suggestion would be the right
approach to materialize that assumption into code and also gain a bit of
efficiency.
I think the other concurrency problem is a problem that comes from lazy
state registration, and I assume most jobs use de-facto eager state
registration, so all the state is registered in `open()` or
`initializeState(...)`, and they are never changed at runtime. For the few
remaining jobs, there must also be a certain timing that they need to register
a new state while a checkpoint is running. Overall, I think those are the
reasons why there is no reported case of a problem from this in practice. But I
agree this should be addressed and is also very easy to fix.
---
