This looks nice enough. It may be worth mentioning that the real challenge for creating a generalized pipeline is rarely the control flow. Usually performance

This looks nice enough. It may be worth mentioning that the real challenge for creating a generalized pipeline is rarely the control flow. Usually performance is not achieved due to; poor intermediate data locality, poor system utility balancing. Unless pipeline code has the ability to match estimated job utility to device capabilities - it won't be useful in many non trivial cases. Unless there is an automated way to store intermediate assets such that data locality between stages is (at least somewhat) optimal, significant amounts of all time will be spent in process migration.