The reason we do this is that there is potentially a bug in Flink RPC system, that CoordinationResponse can only be deserialized by AppClassloader instead of user code classloader, so we can't use customize CoordinationResponse class in RPCs (will lead to ClassNotFoundException). As CollectCoordinationResponse is predefined in Flink and shipped in flink-dist, it will be always visible to AppClassloader, and we use its payload for holding the actual serialized custom CoordinationResponse.
This closes#2744
When the CDC job runs on Flink cluster, DataSinkWriterOperator is
loaded by user code classloader, while SinkWriterOperator is loaded
by app classloader, and the protection mechanism forbids us from
accessing SinkWriterOperator across different classloaders.
Therefore we have to manually load the SinkWriterOperator in user
code classloader, and use reflections to control it.
This closes#2742
Co-authored-by: lvyanquan <decq12ybhl@gmail.com>
Qingsheng Ren
Kunni
Hang Ruan
joyCurry30
Leonard Xu