Right now serialization of high level graphs is a bit of a mess. It was evolved organically and without much architectural thought. This seems to be a source of technical debt.

A lot of the challenge to serializing high level graphs is in order to keep the scheduler protected in a few ways ...

1. We don't want to unpickle client code in a potentially protected environment
2. We don't want to couple the scheduler software version to the client and worker versions
3. We want to allow the scheduler and client and workers to be written in different languages in the future

I like these constraints, but perhaps they're causing more havoc than they're worth. It might be time to reconsider these constraints and instead allow the client and scheduler to communicate by pickle. I think that this would allow us to remove a lot of currently painful code, and accelerate high level graph work in the future.

cc @quasiben and team