Thanks John, very interesting reading.
No luck for us in inventory, as there's essentially updates rather than inserts, which is a bit more complicated than the solution described: "generally the client needs to be smart enough to merge updates based on a timestamp, with a periodic batch job that cleans out obsolete inserts"
But now we're considering the alternative of reading/writing the whole graph at once and process queries in memory. If "the whole graph" is too big to fit in memory without problems, then we should find a way to partition it.