Over the last 2 years or so, I’ve been working on a ground-up rewrite of Alda, the music composition programming language that has been my passion project since 2012. Now that I’m finally almost done(!) with the rewrite and just about ready to release Alda 2 to the world, I figured I should explain why I made the somewhat surprising decision to rewrite it in Go and Kotlin, given that Alda 1 is mostly written in Clojure.
Alda and Clojure
In all honesty, the main reason that I ended up writing Alda in Clojure to begin with is that Clojure is my favorite programming language. At the time in my life when I wrote the first working version of Alda, I just felt like I wanted to write everything in Clojure. And there is certainly something to be said for writing your hobby project in the language that makes you the happiest or the language that you’re most familiar with.
Clojure did prove to be a good choice for me to write the first version of Alda because:
The excellent Instaparse library made it really easy for me to write a little EBNF grammar and generate a working parser.
One of Clojure’s superpowers is making it easy to transform deeply nested data structures from one shape to another, which is basically what Alda does with the parser output to turn it into a score data structure ready to be performed.
Clojure is a JVM language, and the Java Virtual Machine comes with a built-in MIDI synthesizer and sequencer that I was able to use to make sound without requiring end users to install anything extra.
But implementing Alda in Clojure also ended up having another interesting benefit. I made an early decision to include Lisp as a subset of the Alda language. With Clojure itself being a Lisp, I was able to do this trivially by parsing Clojure’s syntax as a subset of Alda’s syntax and then evaluating the Clojure code within the context of the score. This didn’t take much effort to implement, but it unlocked tremendous possibilities for algorithmic music composition with Alda.
To make it more clear what I’m talking about, consider this snippet of Alda code where a piano is playing the notes C through G:
In Alda 1, you can also write your scores (either partially or completely) using a Clojure DSL:
Because we’re exposing the ability to
eval arbitrary Clojure code, you can do
anything available in the Clojure language:
And this is a lot of fun, because the Clojure standard library is chock full of interesting and useful functions for operating on sequences:
Even though it wasn’t part of my original plan for the Alda language, this feature of Alda became very important to me, as I grew to love writing music programmatically by writing Clojure code within an Alda score.
So why, then, did I decide to do a total rewrite of Alda in Go?
One of the major pain points of Alda 1 is that its architecture is rather complex, and the burden of that complexity is foisted onto the user. I wrote about this in detail a few months ago, but the short version is that Alda 1 does most of its work in background “worker” processes, and you can’t do anything useful unless you explicitly start an Alda server first. Alda can’t even tell you if you have a syntax error unless you first have a server running, because even parsing is done in the worker process.
The only reason that I chose to do so much of the work in the worker process (and so little in the client) is that the Clojure runtime is infamously slow to start, which makes it unsuitable for writing command line applications where fast start-up time is important. I still wanted to keep all of the Clojure code that I’d written and continue to develop Alda in Clojure, so as a compromise, I moved all of that code into a background process, and wrote a lightweight Java client that delegates most of the work to that background Clojure process.
In short, I chose to make the architecture more complicated and the user experience worse just so that I could continue to develop Alda in Clojure! But as time went by, I grew increasingly dissatisfied with the complex client/server/worker architecture and the fact that users need to start a server before they can do anything useful with Alda. After mulling it over for a while, I finally decided that it would be worth simplifying the architecture so that most of the work is being done in the client, even if that meant that I had to rewrite Alda in a different language.
As an aside: I’m well aware that nowadays, GraalVM can be used to compile Clojure programs into fast, self-contained, native binaries. However, at the time when I was beginning the rewrite of Alda and I was deciding which language/runtime to use, GraalVM either didn’t exist yet, or it was brand new, so it wasn’t really an option.
Even now, in 2021, GraalVM is probably still not mature enough for me to feel comfortable using it as the backbone of Alda. (Besides, I’ve just spent two years rewriting Alda in Go. I’m not in any hurry to rewrite it again!)
I wanted to move most of the work into the client, and startup time and performance were both super important. It became imperative that I rewrite the client in a low(-ish) level programming language, one that could produce native executables on every platform (at least Windows, macOS and Linux) that start up instantly and run fast.
Go is by no means my favorite language (I could say more about what I don’t like about Go, but that’s a topic for another time!), but it proved to be a good pragmatic choice because out of the options that I tested, which also included Rust and Crystal, Go was the only language that made it easy for me to create 100% static, cross-platform executables.
I also ended up using Kotlin to write the Alda 2 “player” process, a new background process that listens for low-level instructions sent by the Go client and plays audio using the JVM’s MIDI sequencer and synthesizer. I could have stuck with Clojure to write this new player process, but I wanted to cut down on startup time as much as possible, and Clojure is simply a non-starter (no pun intended!) in that area. When it comes to JVM languages, I’m a big fan of Kotlin, because it does a lot in the way of developer happiness (FP affordances, null safety, terseness, actual lambdas, etc.), and it has reasonably good startup time to boot, which makes it well suited for writing command line applications.
Switching from Clojure to Go and Kotlin did have some downsides:
It took me a long time (over 2 years, working off and on in my free time outside of work and family responsibilities) to write a complete port of Alda in a new language.
I could have spent that time implementing one of many feature ideas that I’ve been wanting to add to Alda for ages, but I felt that it was more important to improve the foundation of the Alda platform before I added even more code that I would eventually need to port.
I work comparatively slower in Go and Kotlin than I do in Clojure, which made the process take even longer. This was largely due to not having as much experience with those languages, compared to Clojure, a language that I’ve been in love with since 2012, and that I’ve been using at work since 2014.
On the other hand, I think pushing myself to write a lot of code in Go and Kotlin has made me a more well-rounded programmer. I can truly say now that I have some real world Go and Kotlin experience, which could very well be useful for me someday when working on other projects.
Go makes it hard for me to program the way that I want to. I think it’s fair to say that learning Clojure has warped my brain. Whenever I have to write code in another language, I’m tempted to try to write it in a functional and data-oriented style, or at least as close to that as the language will allow me to get. I’ve had a pleasant experience with Kotlin in that regard. Thanks to its immutable data structures and other FP goodies available out of the box in the standard library, I didn’t need to adjust my thought patterns too much as I was writing the audio engine portion of Alda 2 in Kotlin.
Go was another beast entirely. The language does have a pretty good story for programming with functions as values and writing code that’s oriented around “verbs” (first class functions that compose together) instead of “nouns” (the proliferation of unnecessary classes and objects that you typically see in more OOP-heavy languages like Java). But Go is also an unapologetically imperative language. It’s statically typed, but it lacks generics, which can make it difficult sometimes to implement the abstraction that you have in mind.
At first, I tried to do the sorts of FP things that I do all the time in Clojure-land, like
mapa function over a list, or
filtera list based on a predicate. But thanks to Go’s lack of generics, the closest you can get is to write a function with a very specific type like
FilterString. It would be silly to have to implement such a function again and again, every time I needed to do it with a different combination of types. So, I gave up and embraced the Go Way instead, which is to just use
forloops and do things imperatively.
Another thing about Go that I find cumbersome is that there is more ceremony involved in passing data around. In Clojure, the hash map is practically the only data structure you ever need. It’s idiomatic (even encouraged!) to add arbitrary data to a map on a whim as the map gets passed around from function to function. In Go, on the other hand, there are maps, but they are homogenous (e.g. the entries in a
map[string]intmust all have string keys and integer values) and if you need to collect together various values of different types, you can define a
structthat contains the values you need, but you can’t add an arbitrary field to an existing struct in the same way that you can use
associn Clojure to add any key-value pair you want into a map.
At the end of the day, I have found Go to be quite comfortable, and it provides me with a lot of value as a pragmatic choice for writing native, cross-platform CLI applications with minimal toil. I think it’s well worth the mental adjustments that I have to make to try and think like a Go programmer instead of a Clojure programmer.
When I rewrote Alda in Go, I was careful to preserve the ability to compose music programmatically in Clojure. I achieved this by writing a Clojure library (alda-clj) that provides a Clojure DSL on top of the Alda language. What the library does under the hood is dirt simple. A Clojure expression like…
…evaluates to a record that implements a
Stringify protocol. For example,
the value of the expression above is a
Note record that knows how to represent
itself as a string of Alda code. And alda-clj provides a
->str function that
returns the string of Alda code:
These “domain objects” (part declarations, notes, chords, etc.) compose together
easily, allowing Clojure programmers to construct an Alda score in pure Clojure.
Then, playing the score is simply a matter of wrapping the objects in a call to
Under the hood, all that’s doing is stringifying the domain objects into a
single string of Alda code (e.g.
c d e ) and piping it into the
command to play it, just like any other Alda score.
There are two wonderful things about this.
First, it means that I get to have my cake and eat it too. I can write Alda in the language that it needs to be written in for performance reasons, but I can still use Alda as a vehicle for programmatic Clojure music, the same way that I always have. (Actually, it’s even better now!)
The other thing is that the general pattern that I’ve come up with here can be used to create an Alda DSL in practically any programming language. Contributors have created Alda libraries in Ruby (alda-rb) and Julia (Alda.jl), and I’ve even written a guide to help programmers roll their own Alda library for their favorite language. I can’t wait to see what other libraries people will come up with!
Alda 2: coming soon!
Hopefully I’ve made it clear enough why I chose to reimplement Alda in Go and Kotlin. I also gave you a little taste of what’s to come with the next version of Alda, which I’m very excited about. I’ll be writing more soon about the upcoming release of Alda 2.0, so stay tuned!
Reply to this tweet with any comments, questions, etc.!