Some databases design like extremely simple key-value databases cannot efficiently express joining relation unless load all the data in memory. The same could be said for column based databases etc. I guess that's quite a difference.
"key-value databases cannot efficiently express joining relation unless load all the data in memory"
With the right design can't you store the graph data across multiple keys and then load pieces of it selectively into memory? I get that a graph db specializes in this pattern and makes it more efficient, easier to query, etc., but that doesn't mean it's the only type of db that can model a graph.
Technically it could be, just loading all the keys in a value, which could be massive.
The design of relational databases indicates the database is responsible to do a lot of logic according to the query language. Graph databases usually also include their own query language in order to do the same thing.
For example, from trillions of records there are 100 records with field x equals to value y. And the job is to get all the 100 records instead of trillions of them. The database would take a short query and interpret the predicate logic in the database process, instead of sending the data back to client which usually located in another physical machine.
