Use a subset of the relationships your application actually queries in JanusGraph. Load enough vertices and edges to expose skew, then run the traversals that must be fast. Inspect indexes, write throughput, query latency, and the impact of the selected storage backend.
Graph databases reward a well-defined relationship problem; they do not remove data modeling. Test a schema change and a backup restore, and document how you will reconcile graph records with the source of truth. Decide from measured traversals, not a small toy graph.
Akousa