Does framework provide libraries / helper functions to partition graph with high data locality?

I'm imagining framework users may need to specify some arguments (e.g., selecting from a list of pre-defined algorithms, or optionally implements one and register it as a part of the framework), but since partition graph is a complex problem, having a reusable library could help productivity.

How does this sharded graph representation perform on dense vs sparse graphs?

While the sharing make sense to produce a nice contiguous load from each shard, how come there is no cap on the memory size an iteration can have. Isn't the point to constrain memory usage?