Okay, so assuming that I never want to use Kryo, but prefer to use POJO for state evolution, this is the summary if I understand correctly:

What must be serialisable:

• instance variables if they contain data initialised during operator construction

• the entire Job Graph, including the operator's instance variables

Things that must be serialisable with POJO:

• States
• Records that pass between operators (i.e., both input and output)