RDFox uses datalog to apply inference rules to graph data, it has some state of the art algorithms for making datalog multi-core. It still uses SparQL, which is really similar to datalog for querying. There's some recent work on horizontally scaling datalog queries, but I hadn't seen anything in production.

Probably the most popular implementation is datomic and its free variants. Which has a lot of convenience to it as you can embed custom java functions.

It's not datalog but cypher has a lot of amazing ideas like easily combining algorithms to potentially recursive relationship queries, eg, "Give me the all shortest paths between these two peoples where there's at most 6 hops" -> allShortestPaths((n1:Person)-[*..6]->(n2:Person)) .

Take a look at the Souffle Datalog compiler:

https://souffle-lang.github.io/

Some benchmarks here showing it beating a lot of other engines:

https://souffle-lang.github.io/pdf/cc.pdf (this paper also cites LogixBlox)

They translate datalog into a straightforward relational bytecode, then use Futamura projections to compile that bytecode into a template-heavy "monolithic" C++ program.

As far as I remember one of use of C++ templates is to specialize for different sizes of tuples, so fast assembly is generated for particular queries/facts. They are doing at least 3 nontrivial transformations, and the last step really leverages C++.

The generated C++ is also multi-threaded, which is a big win.

I watched a talk on Rust where Niko M said he prototyped the new borrow checker with Souffle I believe. It's somewhere on YouTube.

Depending on what you are looking for, DDlog (https://github.com/vmware/differential-datalog) is interesting. I'm casually involved, for people weirded out by that, but it supports arbitrary changes to the input facts without recomputation.