Lately I've been thinking that what people want out of graph databases is a better query experience for queries that are built around pattern matching and transitive closure. For those cases, datalog is the most expressive language, and compilation of datalog to SQL is probably the best implementation.