Hacker News Reader

Comment by mark_l_watson | Hacker News Reader

I was invited to work at Google in 2013 on an internal project using their Knowledge Graph. I had just written two books on RDF/SPARQL/linked-data and at first I was taken aback by the query language used at the time. However, when I realized how well the KG scaled and how fast the queries were, I became a fan.

BTW, I think this paper is very well written. Graph queries is not an easy topic and their examples and text make the data types, syntax, and how to perform matches and aggregations easy to follow.

View on HN

More than a decade ago I worked for a few years on a product that was based on OWL/RDF/SPARQL and since then I still see a lot of problems I'm facing in terms of subject, predicate, object triples.

The semantic web feels like a fad that has clearly passed but there are plenty of good ideas from that field that are under-utilized. For example, a lot of what gets encoded as JSON can be described with these triples and once the data is in that form, there are interesting questions you can ask of it.

I think it's far from a fad, the adoption of the semantic web has been enormous in the past 5 years. Almost every major tech company (and many financial institutions) now have their own Knowledge Graph team.

I'm interested in industry history, so I wonder if you can comment on how much of this, if any, came into Google from Metaweb. There are no Metaweb-connected people among the authors (that I can tell) but Warren Harris is acknowledged. Was it a language that was in development before the acquisition, or was it needed mainly after?

As someone who worked on the Metaweb/KG team during that period, it's safe to say the development of PathQuery was a direct result of the acquisition. The lack of Metaweb-linked authors relates to the fact that PathQuery sits closer to 'serving knowledge' than 'representing knowledge'. But lessons learned from the development of MQL surely made their way into PathQuery, through Warren and others.

Mainly after. Metaweb brought MQL with them, which ran on their own single server graphd, which was eventually shutdown.

Warren started working on PathQuery as a replacement for MQL. PathQuery was originally executed on Pregel and was far from a realtime query language.

A realtime query engine was later developed for PathQuery in the search stack so interesting search queries could be answered a la minute from the Knowledge Graph, rather than just pre-generating results for a limited class of queries.

This was all done in the 2012-2013 timeframe. Most of the people involved are long gone.

I've forgotten the details: by real-time query engine are you referring to DGraph, Livegraph or something else?

Definitely not graphd. Maybe Livegraph? I forget what the project was called, but it was the first realtime query engine built for the Knowledge Graph and it used PathQuery as the query language.

Warren was certainly more familiar with the guts of MQL than anybody else, but PathQuery was designed for Google. PathQuery was actually the second system he built for extracting protobufs from graphs. He was also a longtime functional languages aficionado and the prototype implementation of PathQuery was written in Haskell. It took many smart-person-hours to get Warren's interpreter to work at Google scale and speed.

I think what differentiates PathQuery from many "query languages" is that it takes on the task of transforming that data into somebody else's schema (aka "turning protos into other protos"). Having a DSL for this is particularly attractive in a company that will otherwise expect you to do string formatting in C++. But even under normal circumstances there are wins from having data transformation vertically integrated with your query language.

PathQuery was started by Warren Harris. At Metaweb, Warren did a rewrite of MQL in OCaml which we never shipped because of the Google acquisition. He spent much of the first year at Google working on a prototype of PathQuery in Haskell. The prototype was rewritten in C++ and used in the first KG-serving infrastructure sometime around 2012/2013.

What are your thoughts on just using plain datalog/prolog?

I have only used datalog in the form of Clojure and Datomic on a consulting job about ten years ago.

I used to be a huge Prolog fan, but I don't think that anyone has paid me to do Prolog develop in over 20 years. My largest Prolog project was porting a prototype AI planning system that I wrote in Common Lisp to ExperProlog. It took about 6 weeks to prototype the system in Common Lisp, and re-writing it in Prolog only took about two weeks.

I noticed last month that someone was offering an online Prolog course using Swi-Prolog. That might be a good place to start.

Having written a chunk of my phd in prolog and even a simplified implementation of one, I'm definitely a fan. At the same time, our startup (graphistry) works w/ teams needing to look at all sorts of graph data, including knowledge graphs like the topic here, and I'd be nervous about meeting the performance & scale needs with out-of-the-box prolog systems. Likewise, a lot of prolog's power comes from ideas like automatically backtracking search, and I can't imagine using that in most team environments.

Datalog solves most of that :) My read of the paper is their semantics largely reduces to datalog. The most interesting semantic exception is their bounded recursion, which suggests they're doing optimizations normal datalog wouldn't. Likewise, graph workloads often have weird phenomena to optimize for in terms of expected queries & data, where I'm guessing datalog might fit semantically but not how people would normally implement an engine. The PathQuery paper carefully side-stepped all such discussion, so I've been curious.

At Graphistry, we do end-to-end GPU stuff, so I've been thinking about this particular problem, and came to a similar conclusion of a datalog-ish subset being the most straightforward way to get more performance/$ for this kind of task.