Cozo runs (almost) everywhere#

Version 0.1 of Cozo can be used embedded from Python, NodeJS, Java, Rust and C, in addition to running standalone as a web server. Immediately after its release, many people asked about the feasibility of using Cozo embedded on mobile devices.

There was one major obstacle to supporting mobile: Cozo 0.1 used RockDB as the storage engine, and compiling RocksDB for mobile devices is not an easy task. We chose RocksDB because it could handle a huge amount of concurrency and is very fast, but the concurrency part may not be relevant for the mobile case: you almost always have only one process concurrently accessing the database.

So we ripped apart the storage engine code, made a nice and minimal interface out of it, and now Cozo supports swappable storage engines! At the time of this writing, you can choose from the following:

  • In-memory engine

  • SQLite engine

  • RocksDB engine

  • Sled engine

  • TiKV engine

They offer different trade-offs:

  • The in-memory engine is perfect if you just want to use Cozo as a computation engine. For us, it also made writing tests much easier. The downside is that it doesn’t persist data, and it doesn’t support much write concurrency.

  • The SQLite engine uses a minimal amount of resources, is easy to compile for almost all platforms including mobile, and is reasonably fast for reads. SQLite markets itself as a file storage format, and we took advantage of that by making SQLite the backup format for all engines. In this way when you backup your database, you get a single-file SQLite-backed Cozo database. You do not need to restore the backup to look inside: the backup is a fully functional database. As a backup format, it is also extremely space-efficient after you gzip it. The downside is as expected: SQLite is not very fast when it comes to writing and is effectively single-threaded when write concurrency is involved. But as we have said, these are usually not problems on mobile.

  • The RocksDB engine is crazy fast for both reads and writes and can handle an enormous amount of concurrency, while still being rather conservative on resource usage. In particular, its storage on disk is compressed, making its disk space requirements for live databases the smallest among all persistent options.

  • We included Sled as an engine just because we can. The only benefit is that it is pure Rust, and we are not Rust fundamentalists. It is not faster than SQLite for the usual workload that Cozo encounters and uses way more disk space.

  • The TiKV option is the slowest among all options (10x to 100x slower) since data must come from the network. The benefit is that TiKV is a distributed storage. We included it so that people may decide for themselves if it offers value for them. By the way, 100x slower than the other storage options may not be slow compared to the average graph databases in the market – see the next section.

As a result of the storage engine refactoring, Cozo now runs on a much wider range of platforms, including iOS, Android, and web browsers (with web assembly)! We have also expanded the officially supported languages where you can use Cozo embedded: Swift and Golang. Even if your platform/language combination is not supported, you can still use Cozo with the client/server mode. Or you can try to compile Cozo from source and interface it with your platform/language: let us know if you encounter problems, and we will help!