June 14, 2017
For the past few weeks, I’ve been at the Recurse Center working on something I keep referring to as “the database project.” In this post I’ll outline my plans for it and what I’ve done so far – mostly for my own benefit and to gain clarity, but also as a primer for anyone else who might be interested. (And if it sounds so interesting that you want to help, let me know!)
The project is called Logos1. It’s a graph database heavily based on Datomic, which has an unusual set of features that I couldn’t find in any existing open-source databases.2 There’s an overview of Datomic’s design on the rationale page; I also found Rich Hickey’s talk “Deconstructing the Database” helpful in understanding the design and its benefits.
The set of features I’m interested in replicating is something like this:
The database consists of an append-only log of facts. Instead of deleting facts which are no longer true, you retract them; the retraction itself is a fact added to the database, which allows you to determine both that the fact is no longer true and the time before which it was true. Because the entire history of the database is present in the transaction log, you can derive the state of the database at any point in the past as well as the current state for querying.
The information model is a triple-store, like RDF – facts in the database consist of an entity (i.e. an ID referring to a unique thing), an attribute, and a value. Datomic adds to the classic triple-store an additional field for the transaction in which the fact was added; thus, time is an intrinsic property of the data. From the Datomic rationale:
Once you are storing facts, it becomes imperative to choose an appropriate granularity for facts. If you want to record the fact that Sally likes pizza, how best to do so? Most databases require you to update either the Sally record or document, or the set of foods liked by Sally, or the set of likers of pizza. These kind of representational issues complicate and rigidify applications using relational and document models. This can be avoided by recording facts as independent atoms of information. Datomic calls such atomic facts ‘datoms’. A datom consists of an entity, attribute, value and transaction (time). In this way, any of those sets can be discovered via query, without embedding them into a structural storage model that must be known by applications.
The query language for the database is a declarative variant of Datalog, similar to SPARQL. The query engine executes necessary joins without requiring the programmer to explicity declare them. Compare a query to get replies to all posts by a particular user in SQL and Logos’s Datalog variant:
SELECT c.contents FROM comments c JOIN posts p ON comments.parent_id = posts.id JOIN users u ON posts.user_id = users.id WHERE u.email = '[email protected]'
find ?comment where (?user email '[email protected]') (?post author ?user) (?comment parent ?post)
Relationships that SQL expresses via joins are expressed in terms of attributes in the Datomic/Logos/Datalog model. If the SQL query needed to be recursive – say, if the comments were threaded – the difference would be more stark; the Datalog model can easily extend to arbitrary graph relations, whereas SQL quickly breaks down when working with heavily interlinked data.
The database schema is defined and enforced in terms of what kinds of values an attribute can have, but not in terms of what attributes an entity can have, so the schema can be flexibly grown over time. In this respect it occupies a middle ground between relational databases and document stores; it supports data joins even more naturally than a relational database, and schema extensions as easily as a document store. (Of course, the performance situation is not trivial – several different indices are required in order to support different types of queries, and maintaining them can be expensive, so in practice some compromises are necessary.)
This one is, in my opinion, the kicker, and the thing that I haven’t been able to find in other open-source projects. The database is divided roughly into three components. The transactor is the process responsible for handling writes to the database. In order to maintain transactional consistency, only one thread in one process is responsible for executing all transactions and storing the new data, for which both Datomic and Logos rely on the second component, an external K-V store.
The transactor writes facts into the key-value store in the form of segments of each index, and each segment is large – hundreds or thousands of facts, like a page of a B-tree. (Logos’s current implementation uses a persistent B-tree with structural sharing, so that when new facts are added any process that still has a reference to the old root node of the index can continue to use it as long as they need to.) The third component of the database is a library that clients can use to execute queries and request transactions from the transactor. Because data added to the indexes in the backing store is immutable, clients reading from the database only need a handle to the root node of the latest index, and then they can read all the data they need directly from that key-value store without coordinating at all with the transactor. Moreover, the chunks of data in the key-value store can be cached client-side, so if a client wants to do a computationally-intensive query, they can do that by retrieving only the data they need from the backing store and then doing all the expensive computation locally, without interfering with other processes sharing the database. You could therefore run heavyweight analytics queries against the same database servicing low-latency transactions with relatively little performance interference (all processes do have to share the backing store if their caches do not contain the necessary data).
This is all a bit of a grandiose plan for a three-month project; what I’ve described above is the aspirational architecture of Logos. As it stands, Logos is a Rust application that can store and query data either in-memory, locally in a SQLite database, or externally in a Cassandra cluster; however, there is very little schema facility, no support for retracting facts or querying historical versions of the DB, no in-memory buffering of changes to the index (which is necessary to avoid
nlog(n) growth in space usage with a large constant factor as you add index segments), no separation between the client and the transactor, and many other smaller features that are missing from the query language.
However, I am more optimistic than when I started the project that I’ll be able to implement the separate transactor component and client library while I’m here at RC, and perhaps complete some other smaller features as well, though I will certainly not achieve anything like the performance, high-availability guarantees, or transactional capabilities of Datomic in that short time. My primary goal is to get the basic architecture (client, transactor, backing store) working without worrying too much about performance or space efficiency; if I can implement supplementary in-memory indexes and batched index updates I’ll be thrilled, but I think that’s a substantial amount of complexity. My stretch goal for the end of the batch is to implement enough features that I can write a small Twitter-style webapp in Python that uses the database libary via an FFI; we’ll see if I can actually get that far.
The name is a play on a combination of Datalog, transaction logs, log-structured merge trees, and the original meaning of the Greek word “logos” as something like “account” or “ground.” It was not intended as a play on my name, though I now realize the two are a little uncomfortably close.↩
The closest I could find is Project Mentat, a Mozilla project also drawing heavily on Datomic, but Mentat is designed to be embedded and lacks all of Datomic’s distributed features.↩