3 ms·
I'm having trouble understanding how exactly graph databases differ from relational databases, especially seeing as how graphs are isomorphic to relations. All
by colanderman 16y ago
I'm having trouble understanding how exactly graph databases differ from relational databases, especially seeing as how graphs are isomorphic to relations.
All the examples on this page look equivalent to how I'd model them in SQL: http://wiki.neo4j.org/content/Domain_Modeling_Gallery http://wiki.neo4j.org/content/Domain_Modeling_Gallery
The best I gather is that graph databases are schemaless (big whoop, more room for error), and that their implementations tend to perform well with transitive closures. I'm not seeing anything that can't be solved with materialized views in an RDBMS.
I'm tempted to think that just represents NoSQL folks coming to realize that relations are actually a good thing. Soon they'll be talking about how wonderful it is to take the Cartesian product of two graphs.
Someone care to enlighten me?
- aristidb 16y agoIf you have stored a tree with adjacence (parent-child) relations, how would you retrieve all nested children? Or would you propose a different storage mechanism that keeps the property that reparenting entire subtrees is cheap? (All this in your RDBMS of choice.)
- colanderman 16y agoYou mean all descendants of a node / transitive closure? WITH RECURSIVE r AS (SELECT parent, child FROM tree UNION SELECT r.parent, tree.child FROM tree, r WHERE tree.parent = r.child) SELECT parent, child FROM r; transforms this: (A,B), (B,C), (B,D) into this: (A,B), (A,C), (A,D), (B,C), (B,D)
- aristidb 16y agoNice! I did not know about this feature. How efficient is it, and which databases support it? :-)
- colanderman 16y agoIt is SQL standard. I use PostgreSQL, and I'm pretty sure Oracle supports it. MySQL doesn't seem to. Performance wise it's not magical; it quite literally repeatedly scans the database, broadening the search tree until there is no fringe. On a table I have with nearly exactly that structure of about 300 rows and limited nesting depth, the transitive closure takes 4 ms on a crappy VPS. Of course I'm sure graph databases are somehow caching the results or performing some other optimization; you can achieve the same effect in RDBMSes with materialized views (which just cache the results of this query & keep it updated). Oracle natively supports materialized views; you can hack it easily in PostgreSQL and MySQL (though it should really be built in).
- yummyfajitas 16y agoIn a connected graph, you don't want to cache this [1]. It will take up O(N^2) space. If you have 100k nodes, that would require 10^10 cached edges. Further, insertion time becomes worst case O(N^2). Consider a graph with two connected components (each having O(N/2) nodes). Once you add an edge that connects them, updating the view will require O(N^2/4) operations - you need to connect every node in the first component to every node in the second. A graph database is more specialized than SQL, so they can perform a very simple optimization. You don't scan the database, you just follow links. Links are typically stored with the node, so it's O(1) time. [1] In a connected graph, there is a constant time algorithm - return True. But that won't work if you also want edge distance.
- colanderman 16y agoYou're right about caching. I wasn't sure if graph databases took the typical NoSQL approach of massive non-normalized tables to avoid computational overhead. SQL doesn't scan either -- tables are indexed via either a hash (O(1)) or a tree (O(log edges)). While I do realize the speed advantage of a specialized structure, I feel this could be better implemented as a specialized index/storage mechanism in SQL than as an entirely new database system.
- yummyfajitas 16y agoSuppose you have a graph with edge weights. Query: find all the nodes a distance < D from some reference node. With SQL92 or earlier, this will require an unbounded number of queries. Proof: A SQL92 query can only traverse a finite number of edges, since it has only a finite number of joins (say K). Consider the graph 1->2->3->...->N with edge weights 1/N. This will require N/K queries to traverse. N is arbitrary. It's probably doable in SQL99 and various vendor specific extensions (I think it's Turing Complete), but it is unlikely to be fast. So, as is the normal situation, you can do whatever you want in SQL. But the performance is likely to suck. (That said, I agree with you that many people are using NoSQL because it's the hip new thing, when they should just use SQL. "Schemaless flexibility" is BS - SQL is very flexible if you use an ORM + migration tool. The only flexibility you lose is the flexibility to screw up the data in application logic.)
- colanderman 16y agoTrivial with SQL-standard recursive queries: WITH RECURSIVE r AS (SELECT parent, child, distance FROM graph UNION ALL SELECT r.parent, graph.child, r.distance + graph.distance AS distance FROM graph, r WHERE graph.parent = r.child) SELECT parent, child FROM r WHERE distance < 3.14159; Testing your graph with N=1000 on a teeny VPS. 477,897 rows returned in 6.1s. If we move the WHERE clause inside the sub-SELECT (which unfortunately makes the query slightly more awkward to use), selecting lower distances reduces the compute time. For example, a bound of 0.5 returns 196,238 rows in 2.5s. What sort of performance gains does a graph database give for this same query?
- yummyfajitas 16y agoAs I said, you can probably do it with SQL99 query, them being turing complete and all. I don't know how your database is implementing things under the hood, so I don't know. I'll just list the major advantages of a graph database: 1) Edges stored with nodes. You can almost certainly read an edge and all it's neighbors with a single disk seek. This is unlikely to be the case for a SQL system. 2) Compute time is always O( portion of graph traversed), not O(graph size), and is always under the control of the user. In SQL, this will vary quite a bit with implementation. 3) You can specify various heuristics for the traversal - depth first, breadth first, etc. This is important if the specific nature of your data will affect the speed of the query.