Comment by _zoltan_

so the workers understand cross worker topology? Do they always just see individual tiles? what if there is a dependency - eg I'm streaming data and want to compute something later on an earlier computation? How would that get distributed?