Ask HN: Implementing a graph database using Postgres tables for nodes and edges?

87 points by arunnanda ↗ HN
After hearing/reading about the negative experiences many people have had with upcoming graph databases, I am wary of jumping into a graph db. But almost any modern application has things that will be suited to a graph representation.

So I am considering a workaround to use a graph structure, but represent the underlying data in a reliable rdbms, in particular postgres.

The approach is: 1. Make two parent tables: node and edge. 2. Make separate tables for all "objects" (like people, places, things, etc.) which would be inherited from node, so they would all have a unique node ID. 3. Make separate tables inherited from edge which would be used to represent the many many relations. So each relation has an edge ID, and each table inherited from edge can be thought of as representing a specific kind of relation (like person lives in place, or person is friends with person).

One thing I have observed so far, is a large number of tables with few columns, but I think that lends itself to the advantage of easy indexing. There can be a large number of individual queries from the front end, but I believe I can use views to make represent tables with more comprehensive info to reduce the number of queries.

What do you guys think of an approach like this? What am I missing, what is wrong with it? I haven't come across this previously, and so am a bit nervous about the ramifications. Is someone else also doing this?

63 comments

[ 4.3 ms ] story [ 1400 ms ] thread
I use Neo4J and I'm curious what negative experiences you are referring to, and what benefit this might have over that. The biggest problem I foresee is querying and possibly data integrity. Querying since doing any kind of graph traversal would probably require writing hacks in software, and data integrity for much of the same reason. It may be easy to ensure a basic relationship but beyond that I would think it would be more difficult.
it's extremely memory hungry, and you need huge machines
I was never able to use Neo4J as a first level database on a busy website. Simple traversals were just too slow. I ended up maintaining a simple in-memory structure on a pair of machines behind a load balancer with all writes going to sql and then propagated to the graph instances. I realise this introduces issues around read-your-writes etc, but good enough for my use case.
The boring answer is that it depends on your data and what you need to do with it. Insights like "all data structures are graphs!" are dangerous - it can actually add vagueness that might prevent more useful understanding of your problem. Excessively generic handling of data often ends in pain when it comes to the real world (performance and complexity) compared to domain specific approaches.

If you have multi GB graphs and need to run intensive classic graph analysis algorithms then a lot of that can come optimised out of the box with a dedicated graph db. They will still leave you with plenty of hard problems.

However, if you have small graphs or only need conventional crud operations then using a new and unfamiliar graph db would probably be insane.

If your case is the latter and postgres is your comfort zone and you can get a working solution quickly, I would be tempted to prototype along the lines you suggest. When you start getting performance problems or need graph analysis, you can then make a better decision whether a graphdb or restructuring your relational db will solve it the best.

Thanks, that was almost the same reasoning I used to settle on this approach.

I definitely need acid transactions, and also some graph capability - which may or may not grow in the future. More pertinently, the kind of representation I was attempting led to a very normalized design anyway, and I think it also lends itself to all but the most sophisticated of graph functionalities! But I am only just starting, so hopefully it works out well.

I met with Kyle Kingsbury (Aphyr of Call Me Maybe the Jepsen database testing series) out in Berlin and I recall him mentioning at one point that Postgres was a good generic database especially with the JSON blob storage. So I am wondering if making the extensions you propose might not be worth it cause it will come with all that extra overhead.

I do however recommend exploring and even building your own graph database, it is a great learning exercise. But don't fool yourself by doing it on top of Postgres, actually jump in and become immersed. I built http://gunDB.io/ to scratch a similar itch, I wanted an easy graph database with realtime updates like Firebase. I would love to hear your feedback, or even have you contribute.

I think you want to google "triple store" because that is effectively what you are trying to do AIUI. Triple stores are the persistence strategy of choice for RDF and other Semantic-Web-type applications.

Each row in a triple store stores a relationship between 2 objects, with unique IDs for the subject, the relationship type, and the object. The IDs could refer to other tables that store entity and relationship details.

There are specialised triple store DBMSs, but implementing one in PostgreSQL should be pretty easy.

Do you have a gist that implements this approach?
Storing the data isn't an issue. It's fairly trivial to come up with a number of very good solutions to describing a graph.

The issue is querying the data, specifically when that involves walking the graph.

The real question then, is how do you want to query the data? If you are willing to make an assumption now that adds a tight constraint on your future work... i.e. "We will only ever perform queries that are at most 1 or 2 hops from the nodes/edges being queried" then perhaps PostgreSQL will work for you.

You can do recursive graph queries with PostgreSQL: http://www.postgresql.org/docs/9.5/static/queries-with.html
Agreed.

For shallow or narrow queries of the graph this is the way to go.

But for deep and broad queries, if you want to perform some valuable analysis that involves traversing the graph some non-trivial distance from the point of origin, that's where PostgreSQL is going to let you down and you would probably be better off picking a more appropriate tool.

Same question though: How do you want to query your graph, and are you willing to limit yourself to not querying it in a certain way?

(comment deleted)
I'm pretty sure you can get comparable performance from a graph in SQL if you make sure you have covering indexes for the Node ids and the edges/adjacency list table.

This may be implementation-dependent but in theory, it's solid. I'm not sure what else a graph database would be able to do to beat an indexed adjacency list.

(comment deleted)
You raise a good point. I don't expect graph queries that go more than a few links away. To the extent that there will be such queries, I think appropriate indexing will help a great deal.

Implementation wise, it is not that different from a query with many joins anyway - which is a standard problem, and is addressed by materialized views as well as indexing. So as long as the complex queries don't need to be ad hoc and real time, I can prepare for them using materialized views.

However you are quite right in that complex, ad hoc, and realtime graph queries Will be a pain point. That is a risk I am prepared to tolerate in the short term. The thing is, I also need transactional integrity, so I am stuck with an rdbms as the primary solution. In the long term, I anticipate the need for both graph queries and transactions - thus, I think it is better to start with postgres and incorporate a graphdb as the need arises than the other way round.

Hopefully this line of reasoning make sense? I am really not experienced at this, so am trying to not bump into too many walls.

What are you actually trying to do? If you just need some queries on a small amount of read-only data, objects in memory (in a dedicated process, perhaps) are a perfectly sane approach, and much faster than ping-ponging queries with any database.
I had a similar challenge and ended up using ElasticSearch in a thoroughly denormalized data format Allowed me the flexibility of stateful (contextual) queries etc. Bit of a hack, but served my purpose. Hope that helps.
Yes, it might be a hack [actually I don't think it is], but it should offer some powerful functionalities...
Yes, it might be a hack, but it should offer some powerful functionalities... I'll remember this one.
I don't know much about this but Joe Celko has a whole book on representing trees in SQL: http://www.amazon.co.uk/Hierarchies-Smarties-Kaufmann-Manage...

The obvious first question would be: why are you representing your data as a graph? Can you represent it better as sets of predicates (ie: relations)?

Using views should not reduce the number of queries: a view is just a query with a name. If you can do it by combining views you can do it by combining queries.

Why re-invent the wheel ! All you need to do is properly understand your tool of choice. I have used Neo4j extensively for years and have had great results. Our biggest hurdle was understanding how to model the data for a graph properly. Get this right and everything falls into place. The one thing you are going to need is a good visualisation tool, so be prepared to roll one of these too :)
some alternatives to check :

* RDFLib Store for PostgreSQL ( https://github.com/RDFLib/rdflib-sqlalchemy )

* Sparqlify is a SPARQL-SQL rewriter that enables one to define RDF views on relational databases and query them with SPARQL. ( PostgreSQL supported http://aksw.org/Projects/Sparqlify.html )

* PostgreSQL-Neo4j Foreign Data Wrappers https://github.com/nuko-yokohama/neo4j_fdw

* PgRouting (network ) http://docs.pgrouting.org/dev/doc/src/tutorial/analytics.htm...

Thank you so much for this list. I am looking into these... Maybe there's something I can reuse, but at the very least I can get some inspiration from checking them out.
This sounds broadly sensible. It's basically a classic normalised relational design, which is as perfect as alligators or sewing needles, but with extracted supertables for nodes and edges.

I would challenge the idea that you need these supertables, though. What those allow you to do is to write graph queries which are generic (polymorphic?) over multiple kinds of node and edge. So as well as "find me all the people who are friends with my friends", you can write "find me all the people who are friends with my friends, or who have been to a place i've been, or who own a thing i own" without using a union. Do you actually need to write queries like that?

There are two downsides to the supertables. Firstly, more complexity, although it's a minor, or at least constant-factor, amount. Secondly, a loss of type safety. If your edges are defined in a supertable, then the columns which point to the ends of the edge have to be foreign keys to the node supertable. That means they can be any type of node; there's no way to constrain particular kinds of edge to connecting particular kinds of node. That seems like a considerable drawback to me.

True, as I was designing the tables, I was surprised to notice how normalized it looked.

The generic kind of queries you mention: actually, yes, I think I will need them.

Your second point on downsides is something I have/had to think about, thanks for raising it. I had some vague premonitions on those lines, but you helped make it concrete. The problem can be avoided by not making meaningless connections, but that's not a real solution. It won't be a preventive, but it could help to audit relationships - by having a script list out which tables FKs originate from.

Since I started writing this post, I have had two ideas on how to address this, it'd be great to have your opinion on them - I'm thinking out loud here.

1. Use referential integrity. Since each relationship/edge will be in its own inherited table, one could impose a constraint that each column_in_the_specific_edge_table REFERENCES another_column_in_a_specific_node_table. Like the columns in the person_lives_in_place relationship table must reference columns in the person table and the place table - each of which is also its own inherited node table.

2. More convoluted and a bit cumbersome, but if the previous/simpler approach is insufficient, one could create data types corresponding to each node type. And impose that as a type constraint on the edge tables.

But maybe the 1st option could actually work...what do you think?

A lot of the early academic RDF graph databases used this approach. However, performance was no where near to what is required. This let to for the SPARQL/RDF graph databases to now a whole set of independently developed stores. Some on top of existing solutions e.g. Oracle semnet, Virtuoso and DB2 sparql. More on their own solid foundations.

You could test out 3 or 4 different SPARQL solutions in the time that it would take you to develop something graph like on your own.

On the other hand, cutting edge approaches, actually take a graph representation of data and lay it out in a relational manner. http://ercim-news.ercim.eu/en96/special/monetdb-rdf-discover... giving the best of both worlds.

In short you can build something yourself. But don't expect that it will be better than something build and supported by someone else.

So investigate the competition: BlazeGraph, Virtuoso, StarDog, Oracle 12c EE/semnet, DB9, Dydra before deciding to build your own. Building your own because its fun to do is great, but unless it pays your bills not a good idea for production environment.

PS. The edge table (EAV) is the major problem, it leads to a lot of self joins and difficult exercises for the query planner.

You can improve a lot on this if you can put "different" edges into different tables or partitions.

Triple stores with support for JSON-LD framing such as Dydra can also make it easier to have a front-end on top of your DB without extensive middle layer code.

A store like StarDog and BlazeGraph on the other hand gives you a lot of flexibility by both supporting SPARQL and TinkerPop. (both cluster and scale out, although BlazeGraph has GPL option. StarDog is only Commercial)

edge is not stored as IAV the table hierarchy is:

> Node <- Entities

> Edge <- M2M

which is strange since an edge is already a M2M.

Thanks for the monetdb link.

Ah ok, so in RDF terms each predicate gets its own table, which makes sense. Then you have a subject foreign key relation ship and a object foreign key relation ship to the node tables.

That would be better than one big table in performance, which was a major problem in the RDF on SQL databases.

Of course those accepted any graph, if one constrains the number of possible predicates/relations then this solution could more efficient.

Building a graph database is time consuming. Why are you putting edge's children (person, place, friend) in a separate table instead of using a json field? Do you plan to query the person, place, friend table directly? You need to benchmark the different queries.

My understanding is that big5 and others do not use graphdb for all their stuff since they are not as good as rdbms to do queries with a single hop or JOIN. They embrace microservices and maintain a graphdb (in memory, persistent or distributed) to answer domain specific queries. That approach is similar to your schema except that graph queries run on a single node without superfluous network roundtrips.

It would be nice to be able to use a single database for all data related stuff to have atomic writes and simpler architecture. That's what multi-model databases are tackling have a look at OrientDB and ArangoDB [1].

Also, Tinkerpop, already mentioned in the thread, is a ready-to-scale graphdb with much love I recommend you have a look at Tales of the Tinkerpop [2].

[1] https://news.ycombinator.com/item?id=10180185

[2] https://news.ycombinator.com/item?id=10316140

Of course, JSON is a good way to store ad hoc, or new relations which haven't yet been structured. But for frequently occurring relations, I'm preferring separate tables instead of JSON for ease of querying and joins. If number of relationships are large, querying within a long JSON field is not the best approach.

Multi model db's are a good idea indeed. They were the first things I had considered, and extensively so - both OrientDB and ArangoDB. I came back to traditional solutions simply for practical reasons of maturity, and ease of finding professional support.

Someone here mentioned about a Tinkerpop implementation that uses Postgres as the storage engine, I think that might offer the best of both worlds, but I am yet to look into it.

The point about micro services is a really valid one. In principle, if the application is neatly split into a group of micro services, it would make things easier. This is something I am going to look into once the application has evolved somewhat - it is hard to make the split from the get-go, without knowing how it is going to be actually used - by the end user.
I was also wondering about the same thing:

Implementing a graph on top of a RDBMS is trivial, and if the semantics are correct (the same as the ones exposed by a graph db?), then I'm not sure why would people want to use a proper graph db.

I thought that probably it'd be an issue of performance: the "right tool for the job" that "does one thing and does it well" probably is leaner and more efficient. After all, unlike a trivial implementation, getting a graph on a RDBMS to perform well might not be that simple after all (still, your idea of inheriting tables might make things more flexible and maybe more efficient)

But then, when looking up some Neo4j benchmarks, the numbers seems to not be good at all:

http://baach.de/Members/jhb/neo4j-performance-compared-to-my...

I'd like to hear from someone that used Neo4j (also, other graph databases are interesting) in production, and benchmarked it against a RDBMS prototype, finding the former as the better solution of the two.

While the author admits that the documentation advises to use the traversal API (which gives the best performance) he goes with cypher over the REST interface which would never match a direct db connection. It's a fundamental flaw in his comparison.
AFAIK:

A proper graph db is almost mandatory if there are complex, ad hoc queries that need to be made on real time data.

Using an RDBMS is great if the graph query types are going to be known in advance - so they can be prepared for using materialized views and indexes, and if they aren't too complex - so one doesn't descend into JOIN-hell.

But most applications aren't like that, and not all applications can be completely satisfied using only a graphdb. Hence the rise of the new multimodel databases like OrientDB and ArangoDB. So I think it is a question of what risk one is prepared to tolerate.

I've done this before with MySQL in 2011 after throwing my hands in the air with two different graph DBs.

By far the most productive thing I did. People will say that you need recursive queries, etc, but I found most of the time just keeping the query set in the application layer was easy enough and super fast. Scaling it was easier too because there are lots of articles on scaling the MySQL / Postgres.

Agreed, my reasons for coming back to Postgres are quite similar. In general, one only needs to go a few layers in the graph. Recursive queries are needed for specific use cases. Even then, querying from the app, like you said, and/or building materialized views and materialized paths are very good ways of working around this, as long as the query is anticipated in advance and the view/path built before the query. So that reduces the real use of graph db's to an even smaller subset - complex ad hoc graph queries.
If directed acyclic is sufficient for your use case, you can use this approach, which has worked perfectly for me in the past: http://www.codeproject.com/Articles/22824/A-Model-to-Represe...
This is a really interesting approach, thanks! I have a relational-to-graph problem that can be done as a DAG, and I was planning on using Postgres as a source of truth and mirroring out the data to Neo4J just for queries, but I think I might try this instead.
I believe Redis and Hyperdex have constructs such as bitarrays and sets that can be easily used for representing and querying graphs/hypergraphs. For Redis there are some success stories and even ready to use extensions.
I've done this a few times. Depending on the size/density of your data, it can be fine. If you have billions of entities and/or highly dense relations, it's not very efficient though.

I didn't use table inheritance, and I don't distinguish relationships at the query stage, so if that's core to your method you should think about it. I assume with that setup it'd be easier to take advantage of partial indexes though.

Recursive queries work but if the intermediate tables gets large it explodes fast. You can't use aggregates in the recursive function either (e.g. to count number of leaf nodes in a tree), you have to apply them at the end, in which case the intermediate table has to be large...

I've had decent experience so far with GIN indexes and @> operators on an ARRAY[] column adjacency list. In my case I stored "ancestors" so I could select by object id and get all ancestors, or use `ancestors @> ARRAY[parent_id]` to select all object ids as descendants.

Of course, if you don't need any "self" relationships, then none of the really complicated stuff matters...

Thanks for the feedback.

Indeed, one of the reasons behind inheriting tables was to use partial indexes - which should help with performance. Another was ease scaling out, if needed.

Using the @> operators on array columns is something I also looked into, mainly for materialized paths - I expect them to remain static, so no expensive updates to the paths/arrays. But actually, I don't think I will need many self relationships - that was the third reason for using inherited tables, so I could split entities into separate groups.

Of course, I have no idea how any of it will work out, since this was just a semi-serious line of inquiry initially, and everything was conceptual so far, except the feedback of experienced people such as yourself, on this thread - which has led me to pursue this seriously and actually try to build it out. After reading the responses on this thread, I now think it will be worth the effort.

No-one seems to have yet mentioned Facebook's TAO paper, which describes scaling basically this approach to 1 billion reads per second. https://cs.uwaterloo.ca/~brecht/courses/854-Emerging-2014/re...

The nice thing about your approach is that you can shard your objects by their id and your edges by their "from" id, and have all lookups go to one box when you shard. Throw in some caching and it scales to multiple boxes really well.

You said it!

Actually, Facebook's TAO was an early inspiration for me as well, but it is really customized and complex too, so probably not something for a tiny team. But the concepts from TAO are applicable to many similar scenarios.

I would love to see a SQLite extension to support graphs...