Abstract, optimized spatial partitions are not necessary, and difficult to communicate to the citizens who need to know whose district they're in.
Boundaries should conform to existing political boundaries, even Zip codes, or recognized geographical features, such as rivers or street centerlines. Neither is it necessary that all districts contain precisely the same number of voters, a few percent variation is acceptable because the census population data is not perfect either. They should be compact polygons with easily identifiable boundaries and roughly equal populations. No other criteria should be considered.

If the algorithm splits an ethnically or economically homogeneous neighborhood in two, or combines two radically different ones, concentrating or diluting voting blocs, too bad! As long as it is not done deliberately, so be it. Next year new district boundaries will be generated anyway. The way it's done now, inequities are not only probable, they are guaranteed.