For my understanding of the node (a good implementation could hide disk I/O by prefetching data for next iteration of loop), I assume this is the job of framework (not programmers who use the framework).

We didn't talk about multi-core implementation of such algorithms - do we also assign work statically to each core (instead of dynamically) so the cores won't compete to load the same shard and invalidate each other's caches, which would result in worse performance?

I'm wondering if this would be considered as graph compression or streaming or both? Are there are other graph streaming algorithms we can employ depending on how many times we visit? (i.e. In most contexts I've seen streaming before, elements are visited exactly once, but this introduces a more dynamic load-and-reconstruct idea.)