Apparently it is not easy to create general purpose datalog engines that scale.
"In this paper, we started with the observation that Datalog engines do not translate across domains. We experimentally evaluated the advantages and disadvantages of existing techniques, and compared them with our own baseline, a general-purpose, parallel, in-memory Datalog solver (RecStep) built upon a rdbms.
"We presented the necessary optimizations and guidelines to achieve efficiency, and demonstrated that RecStep is scalable, applicable to a range of application domains, and is competitive with highly op- timized and specialized Datalog solvers."
Don't despair. Datalog is alive and well. Yes, it does compose beautifully. However, with composition solved, you'll discover that a naive implementation of relational algebra will suffer from spurious cross products.
This is pretty interesting stuff, thanks for sharing! Does temporal information get any special treatment in the economic graph at the moment? Are temporal queries of any interest?
I ask because I work on Crux [0] which at its core is a point-in-time bitemporal Datalog engine, and I see a lot of similarities (schemaless core, Worst-Case Optimal Join etc.)
"In this paper, we started with the observation that Datalog engines do not translate across domains. We experimentally evaluated the advantages and disadvantages of existing techniques, and compared them with our own baseline, a general-purpose, parallel, in-memory Datalog solver (RecStep) built upon a rdbms.
"We presented the necessary optimizations and guidelines to achieve efficiency, and demonstrated that RecStep is scalable, applicable to a range of application domains, and is competitive with highly op- timized and specialized Datalog solvers."
vldb.org/pvldb/vol12/p695-fan.pdf (2019)