Mastodon, the free, open-source social network server based on ActivityPub is
written in Ruby on Rails and the network is experiencing influx of new users.
Mastodon administrators are finding how difficult and costly is to scale Ruby
on Rails applications the hard way. I’ve spent a deacde working on a large Ruby
on Rails project, much larger than Mastodon. Let me quickly describe what is
going on. Disclaimer: This post is solely based on my experience with scaling
other Ruby on Rails applications, take this with a grain of salt. Also, I’d
appreciate comments at
People often think that Ruby on Rails is slow because Ruby is slow, according to various benchmarks and shootouts. Well, while Ruby is not fast at all, it is not the primary reason why that is. See, most web applications, including Mastodon, do not perform CPU-intensive tasks. Most of the time, CPU is actually waiting for data to be read or written from network (client, database, redis) or disk (cache). Even if you upgrade to the most recent Ruby 3.2, which is the fastest of all Ruby versions, it won’t help at all.
Backend software these days need to respond to many of HTTP requests, when a service is under load we are speaking about hundreds or even thousands of requests per second. A single process (instance of software running on an operating system) with naive implementation can handle as much as one request. To be able to handle more, requests must be dealt concurrently, which is a very complex topic but let’s keep in simple. The solution is to create multiple execution threads that can run concurrently on the program level and ideally in parallel on the operating system level to utilize as much CPU resources available.
There is a little problem in Ruby (and Python and many other interpreted languages) called “global interpreter lock”, also known as GIL which limits the amount of parallelism reachable through concurrency of a single interpreter process with multiple threads. In short, while GIL allows some parallelism when threads are waiting for I/O (input and output, e.g. writing data into database) in theory, in practice it does not scale. Most applications have the sweet spot of 5-15 threads per process when the most performance can be achieved. This highly depends on the hardware or virtual machine (CPU type, number of cores, hyper-threading).
Don’t be mistaken that other languages do not have this problem, even server software written in C (e.g. Apache httpd 2.x) is known to have scalability issues with many threads. For example the ThreadsPerChild Apache setting is set to 64 by default (25 on Windows, whoops). But performance is not the only concern, isolation and security is also at stake here.
Mastodon, like many other Ruby on Rails applications, needs to schedule backround tasks. The backend API communicates with some kind of job scheduler through ActiveJob API and one or more backend worker instances pick the tasks and perform the work (e.g. delivering emails, updating database). The same problem applies - one worker process can effectively run only the mentioned 5-15 threads, but there can be thousands of jobs in the queue.
What to do? It is farily simple. You spawn multiple processes which are simply “copies” of the main process doing the same work. This is what Apache httpd does too, so it can’t be wrong. So roughly speaking, 10 processes can handle 150 tasks concurrently. That is a decent number, is it?. Well, there’s a snag.
Ruby and Ruby on Rails specifically are not light on memory. Ruby itself suffers from memory fragmentation problem, another complex topic but in short: when Ruby allocates a memory, it is quite unlikely to ever return it back. And Ruby on Rails is designed to be rather memory hungry, particularly its ActiveRecord DAO API. One process of the application I have experience with consumes 300 MB after start, and then it grew up to 500 MB during operation as other parts of the app load up. Mastodon is not as big, but we are speaking about hundreds of MBs per process after hours of operation. There is a way to workaround the memory problem on some operating systems and that is copy on write concept, but in practice that does only help a little after application is started.
And that is only ideal state. Programmers are humans, they make mistakes and one very common mistake is a memory leak. These memory leaks together with Ruby fragmentation makes process memory consumption to steadily grow up until the point that the process must be restarted. Most Ruby applications servers do have some kind of memory limiter feature which triggers automatic restart of the process when a threshold is reached.
Once a hardware or a VM runs out of memory, the OS starts swapping that starts killing the whole instance and everything is getting slower up to the point that the application is unusable.
Now, Ruby has multiple implementations. The one that Mastodon runs on is called CRuby and it has all of these problems I described in my post. There is also JRuby, a JVM-based Ruby implementation which is substationally faster and does not suffer from GIL. But there are drawbacks: it is not compatible with Ruby on Rails and while some older versions of RoR can run on JRuby, most applications won’t due to bugs in dependencies which are only tested against CRuby. Also, JRuby is even more memory hungry than CRuby from my experience.
So there you have it: Ruby on Rails requires multiple processes as the application grow, these processes require substantial amount of memory which only grow with time. Memory is a precious resource, so Mastodon instances can burn a lot of money along the way. Be prepared for this!
Wait a moment, you just said that Ruby and Ruby on Rails sucks! That is totally not what I mean. This is a wonderful platform that enables millions of people to rapidly develop web applications and operate them, many major sites are written in Ruby on Rails including github.com and even twitter.com started as a RoR app. And there’s hey.com and basecamp.com and they all run buttery smooth. Like everything in life, it has some highlights and drawbacks. And the major drawback of Ruby on Rails is that it is not easy to operate at all.