<img src="static/logo_c.png" width="200" height="175" alt="Logo">

CozoDB

Table of contents

1. Introduction
2. Getting started
3. Install
4. Architecture
5. Status of the project
6. Links
7. Licensing and contributing

🎉🎉🎉 New versions 🎉🎉🎉

Version v0.7: after HNSW vector search from 0.6, in 0.7 we bring to you MinHash-LSH for near-duplicate search, full-text search, Json value support and more! See here for more details.

---

Version v0.6 released! This version brings vector search with HNSW indices inside Datalog, which can be integrated seamlessly with powerful features like ad-hoc joins, recursive Datalog and classical whole-graph algorithms. This significantly expanded the horizon of possibilities of CozoDB.

Highlights:

• You can now create HNSW (hierarchical navigable small world) indices on relations containing vectors.
• You can create multiple HNSW indices for the same relation by specifying filters dictating which rows should be indexed, or which vector(s) should be indexed for each row if the row contains multiple vectors.
• The vector search functionality is integrated within Datalog, meaning that you can use vectors (either explicitly given or coming from another relation) as pivots to perform unification into the indexed relations (roughly equivalent to table joins in SQL).
• Unification with vector search is semantically no different from regular unification, meaning that you can even use vector search in recursive Datalog, enabling extremely complex query logic.
• The HNSW index is no more than a hierarchy of proximity graphs. As an open, competent graph database, CozoDB exposes these graphs to the end user to be used as regular graphs in your query, so that all the usual techniques for dealing with them can now be applied, especially: community detection and other classical whole-graph algorithms.
• As with all mutations in CozoDB, the index is protected from corruption in the face of concurrent writes by using Multi-Version Concurrency Control (MVCC), and you can use multi-statement transactions for complex workflows.
• The index resides on disk as a regular relation (unless you use the purely in-memory storage option, of course). During querying, close to the absolute minimum amount of memory is used, and memory is freed as soon as the processing is done (thanks to Rust's RAII), so it can run on memory-constrained systems.
• The HNSW functionality is available for CozoDB on all platforms: in the server as a standalone service, in your Python, NodeJS, or Clojure programs om embedded or client mode, on your phone in embedded mode, even in the browser with the WASM backend.
• HNSW vector search in CozoDB is performant: we have optimized the index to the point where basic vector operations themselves have become a limiting factor (along with memcpy), and we are constantly finding ways to improve our new implementation of the HNSW algorithm further.

See here for more details.

Introduction

CozoDB is a general-purpose, transactional, relational database that uses Datalog for query, is embeddable but can also handle huge amounts of data and concurrency, and focuses on graph data and algorithms. It supports time travel and it is performant!

What does embeddable mean here?

A database is almost surely embedded if you can use it on a phone which never connects to any network (this situation is not as unusual as you might think). SQLite is embedded. MySQL/Postgres/Oracle are client-server.

A database is embedded if it runs in the same process as your main program. This is in contradistinction to client-server databases, where your program connects to a database server (maybe running on a separate machine) via a client library. Embedded databases generally require no setup and can be used in a much wider range of environments.

We say CozoDB is embeddable instead of embedded since you can also use it in client-server mode, which can make better use of server resources and allow much more concurrency than in embedded mode.

Why graphs?

Because data are inherently interconnected. Most insights about data can only be obtained if you take this interconnectedness into account.

Most existing graph databases start by requiring you to shoehorn your data into the labelled-property graph model. We don't go this route because we think the traditional relational model is much easier to work with for storing data, much more versatile, and can deal with graph data just fine. Even more importantly, the most piercing insights about data usually come from graph structures implicit several levels deep in your data. The relational model, being an algebra, can deal with it just fine. The property graph model, not so much, since that model is not very composable.

What is so cool about Datalog?

Datalog can express all relational queries. Recursion in Datalog is much easier to express, much more powerful, and usually runs faster than in SQL. Datalog is also extremely composable: you can build your queries piece by piece.

Recursion is especially important for graph queries. CozoDB's dialect of Datalog supercharges it even further by allowing recursion through a safe subset of aggregations, and by providing extremely efficient canned algorithms (such as PageRank) for the kinds of recursions frequently required in graph analysis.

As you learn Datalog, you will discover that the rules of Datalog are like functions in a programming language. Rules are composable, and decomposing a query into rules can make it clearer and more maintainable, with no loss in efficiency. This is unlike the monolithic approach taken by the SQL select-from-where in nested forms, which can sometimes read like golfing.

Time travel?

Time travel in the database setting means tracking changes to data over time and allowing queries to be logically executed at a point in time to get a historical view of the data.

In a sense, this makes your database immutable, since nothing is really deleted from the database ever.

In Cozo, instead of having all data automatically support time travel, we let you decide if you want the capability for each of your relation. Every extra functionality comes with its cost, and you don't want to pay the price if you don't use it.

For the reason why you might want time travel for your data, we have written a short story.

How performant?

On a 2020 Mac Mini with the RocksDB persistent storage engine (CozoDB supports many storage engines):

• Running OLTP queries for a relation with 1.6M rows, you can expect around 100K QPS (queries per second) for mixed read/write/update transactional queries, and more than 250K QPS for read-only queries, with database peak memory usage around 50MB.
• Speed for backup is around 1M rows per second, for restore is around 400K rows per second, and is insensitive to relation (table) size.
• For OLAP queries, it takes around 1 second (within a factor of 2, depending on the exact operations) to scan a table with 1.6M rows. The time a query takes scales roughly with the number of rows the query touches, with memory usage determined mainly by the size of the return set.
• Two-hop graph traversal completes in less than 1ms for a graph with 1.6M vertices and 31M edges.
• The Pagerank algorithm completes in around 50ms for a graph with 10K vertices and 120K edges, around 1 second for a graph with 100K vertices and 1.7M edges, and around 30 seconds for a graph with 1.6M vertices and 32M edges.

For more numbers and further details, we have a writeup about performance here.

Getting started

Usually, to learn a database, you need to install it first. This is unnecessary for CozoDB as a testimony to its extreme embeddability, since you can run a complete CozoDB instance in your browser, at ne