@timowest would it help to make hashCode transient, and fill it with readObject()?
That way it doesn't get serialized, and you can ensure that the hashcode came from the current running JVM(so that it is consistent across the lifecycle), and not from the serialized form.
Of course, this will break existing serialized forms, so we'll have to think about this carefully.

Yes, I am doing that right now.

It would only make sense, since it is not a significant state value.

I haven't yet tested with multiple JVMs, but I could alter the hashCode values via reflection and make sure that they are properly reset via deserialization.

I am already testing with multiple JVMs, and can at least see that the fields are correctly restored.
I am now resetting to before the fixes to compare the behavior, but it looks good.

There are no cats in the database, do you see the same?
This makes the build fail.
Speaking of failing builds, Travis seems to be down.

Looks good, I also deserialized the predicate in a seperate VM and it works.
Nicely done Timo!

Seeing as the build succeeds, I will merge this.