Scaling NetworkX to largest graphs with UKV

[Ashot Vardanian]

Dear NetworkX community,

For many years I have been enjoying the ease of using NetworkX and teaching it to my students. Today I want to return a favor and share something my team and I have been open-sourcing over the last year - UKV.

It brings NetworkX-like bindings to several storage engines, like LevelDB and RocksDB, and our more scalable alternatives. 

• Graph interface docs
• Broader project page on GitHub
• Benchmarks

This should allow everyone to deal with humongous graphs using an efficient underlying C++ implementation while keeping the application logic Pythonic and clean.

It is much faster than Neo4J, but we are still determining if 1, 2, or 3 orders of magnitude separate them. The numbers in our table above are probably wrong (even though we tried multiple JVM configurations and different Neo4J versions), but you will feel the speed once you start working with UKV. We are also trying to integrate cuGraph to bridge the gap between the SSD and the GPU, making the system even faster.

We support directed and attributed graphs but are not feature-complete or bug-proof yet. Still, we would appreciate your feedback and star on GitHub. Feel free to join our Discord if you have any questions or feature requests! Happy coding!

[Russell Jurney]

Can UKV work with Neo4j? Sit on top of it?