Why I Built Litestream
Posted February 10, 2021 by Ben Johnson ‐ 6 min read
tl;dr—Despite an exponential increase in computing power, our applications require more machines than ever because of architectural decisions made 25 years ago. You can eliminate much of your complexity and cost by using SQLite & Litestream for your production applications.
When I was your age…
I can tell I’m getting old because I talk about the “good old days” of computing. Back when knowing a single programming language and SQL was good enough to work at most jobs. Back when you could build all your web pages with basic HTML skills.
But the good old days weren’t actually that good.
They kinda sucked.
In 1995, the LAMP stack emerged. It was a combination of Linux, Apache, MySQL, & PHP. It’s a simple stack that all lived on one box. Unfortunately, that box was stocked with a CPU that was measured in megahertz and memory that was measured in megabytes. It was dog slow.
But Moore’s Law promised a better future as computers became exponentially faster. So why do we need more computers than ever before?
Where it went wrong
Since early web languages like PHP and Ruby were slow, you didn’t want them running on your database server so you moved them to their own servers in what is called an n-tier architecture. Now you can keep adding more stateless Ruby servers to this presentation layer and scale your application.
Database servers like Oracle, PostgreSQL, and MySQL tended to be complicated to operate so this architecture helped to isolate them. Scaling and managing a fleet of database servers is a nighmare so organizations scale up their database machine vertically as much as possible before they scaled horizontally to multiple servers.
This n-tier architecture sounds simple at first but it has hidden complexity. On a single machine, our server could add an in-memory cache to speed up requests but now data is shared across multiple machines so we must add a memcached or Redis server to share cached data. Database servers are easily overloaded by numerous connections so we have to add intermediate services like PgBouncer to pool connections. If we have events in our system that must be communicated to all our nodes then we need a cluster of Kafka machines.
Now we have a fleet of machines to manage. These systems have then caused an explosion in infrastructure automation with tooling like Kubernetes. All of these layers slow us down from our job of writing software systems that solve real problems.
What started as a simple two-tier software system has grown into a behemoth with a dozen layers of complexity.
Complexity begets complexity.
For a while, I used Bolt when writing applications and it was refreshing because there were no depedencies to set up and performance was blazingly fast. However, it lacked features like schema migration, a query language, or a REPL so it made application development difficult. But instead of going back to database servers like Postgres, I turned to SQLite.
Moving to SQLite
“But nobody writes production applications with SQLite, right?”
No, there’s a growing movement of people that see the value of single process applications. Expensify has run tests of 4 million queries per second on a single node using SQLite. David Crawshaw has a conference talk and blog post on building single process applications on SQLite. Django co-creator Simon Willison built a data exploration and publishing tool called Datasette that’s built on SQLite.
SQLite is known for being bulletproof and having an absurdly in-depth testing suite. Its motto is “Small. Fast. Reliable. Choose any three.”
So why is SQLite considered a “toy” database in the application development world and not a production database?
That one big issue
The biggest problem with using SQLite in production is disaster recovery. If your server dies, so does your data. That’s… not good.
Other database servers have replication so they can stream database changes to another server in case one goes down. The best you can hope for with standard SQLite is to run a nightly backup. Solutions like rqlite are great but it requires a 3-node cluster.
Why can’t SQLite have a replication tool that’s as easy to use as SQLite?
The problem Litestream solves
I built Litestream to bring back sanity to application development. Litestream is a tool that runs in a separate process and continuously replicates a SQLite database to Amazon S3. You can get up and running with a few lines of configuration. Then you can set-it-and-forget-it and get back to writing code.
You might think this sounds expensive to continuously write to cloud storage that provides 99.99% uptime and 99.999999999% durability but it’s astoundingly cheap. Typical costs are only about $1 per month. Litestream is free and open-source too so there’s never a license to pay for.
But I need 100% uptime…
The software industry has chased uptime as a goal in and of itself over the last several decades. Solutions such as Kubernetes tout the benefits of zero-downtime deployments but ignore that their inherent complexity causes availability issues. There’s even a web site dedicated to public postmortems related to Kubernetes.
Most cloud providers provide multiple layers of redundancy in their systems to protect against individual node and network failures. This doesn’t provide a 100% guarantee but will provide you with very high uptime servers. Anecdotally, I’ve run several VPS servers over the years which all have well over 99.9% uptime and have suffered no catastrophic failures.
Sounds good, but how do you scale this?
Developers always want to know how to scale but that depends on their particular application. Typically, you want to scale vertically first by simply increasing amount of CPUs cores and RAM on your machine first.
Servers these days have a ton of power. I recently wrote a Go web application with a SQLite database that would serve an HTTP request with multiple database queries in under 50µs on a $5/month VPS. That translates to thousands of requests per second per core. SQLite scales reads well with the number of cores on a machine. Amazon AWS has machines that can then scale up to 96 CPU cores and hundreds of gigabytes of RAM.
If you exceed the capacity of a single node, sharding your data can allow you to scale horizontally to multiple nodes. This works particularly well for SaaS applications where each customer is isolated from one another. Because SQLite and Litestream simplify deployment, managing a cluster of several isolated nodes is easy to maintain.
The Future of Litestream
Litestream is helping to simplify application development but that’s only the start of it. There are exciting features coming including replication to read-only replicas. This will give you the ability to run local copies of your database at the edge to deliver requests instantly.
If you’re interested in trying out Litestream, the Getting Started can get you up and running in less than 10 minutes.
I’m interested to hear from others that want to simplify and improve application development. If you have ideas or thoughts about the future of Litestream, please get in touch on the GitHub Discussions board and drop me a line.
Update: This post sparked a lot of discussion on Hacker News that you can read here.
—Ben Johnson (@benbjohnson)