The race conditions have been eliminated and we greatly simplified the cluster setup.
With this "new" tool shipped with arangodb https://github.com/arangodb-helper/arangodb it is just a single line on each machine to get the cluster up and running.
Hey Michael from ArangoDB here. I was also a bit surprised. Especially for the graph case. We see that more memory is consumed but the throughput did not really increase in comparision to the benchmark we conducted 3 years ago. Nevertheless we see improvements on the shortest path queries, these challenge the database more. As we are no dedicated Neo4J experts we are intrested in every configuration that will speed things up, please share them with us.