Veit's Blog

zepto, A Eulogy


Almost exactly a year ago today I was in a small room in a venture capitalist’s apartment, surrounded by 20 close friends and friends of friends, preparing to announce version 0.9.6 of zepto, the language’s first stable release. It was supposed to be a stable alpha that people could download and play around with. Version 0.9.7 was already in the works, with exciting things, like a more robust parser, a better traceback format, a faster hashmap implementation, and loads of other minor bugfixes. It never saw the light of day.

The commits came less and less frequently until at last, two months later, I pronounced the project dead to me. I haven’t worked on it since, although some of my systems, such as the static site generator of this blog and some other websites, the fast stack-based calculator I use when in the terminal, a tool I use for generating documentation for my work, the cowsay prompt that greets me every time I open a terminal, and some miscellaneous scripts I use, still depend on it. The language still works, although some of the native extension that depend on outdated Haskell packages do not anymore. The bitrot set in quickly.

Now, one year after my optimistic announcement, I want to tell you what zepto was, what it stood for, why it was great in many ways, and, most importantly, why I decided that zepto is, in fact, a failure.

A room of one’s own

I think most of us, at one point in our career, question our favorite language’s—or languages’—ways. We realize that the creators and BDFLs that we used to worship are not nigh-infallible; they are mere humans. And so, in youthful rebellion, we go full-on iconoclast and write a language of our own. Or we sit in our study—does anyone still have those?—and quietly write code, trying to figure out how things fit together, and how we could improve the systems we’re currently working with, incorporating the stone tablets of our mothers and fathers, full of weird type signatures, grammars specified in Backus-Naur notation, and oh so many parentheses. And then, maybe we have a great idea. Or maybe not.

zepto was one idea that I had. I had worked on a few Lisps before, in C and Haskell, and decided I liked them a lot. I didn’t know any Lisp when I embarked on that journey: I learned Haskell and Lisp both while writing zepto’s ancestors, and, to a large extent, zepto itself. You can still tell: the Haskell code that fuels zepto is pretty terrible, and so are the early libraries that I wrote. It was a bad idea to incorporate those into the standard library, but I did, accumulating cruft from zepto’s inception. Starting out with code from the now-infamous “Write Yourself A Scheme in 48 Hours” probably didn’t help either. The new version wasn’t around back then.

I ended up working on zepto for almost two years, devoting a lot of nights, weekends, and my bachelor thesis to it, while working full-time. My bosses were always very supportive, even encouraging. The CTO at Bright hosted a Clojure meetup and made me present an early version of zepto, to very lukewarm, paternalistic feedback. I was told things like “we all did that once”. Noone seemed to realize zepto was different! I now know my sales pitch tried to get people to try zepto in an underhand way, telling people how similar zepto was to the language they were currently using instead of telling them how it was different, even better than their tools. I lacked confidence, and for good reason: I myself didn’t know what zepto was or what it was meant to do.

About a year in, I eventually realized that I needed a focus, and I formed three core principles: naïveté, malleability, and friendliness. What at first felt constraining ended up being the most liberating thing I could’ve done for zepto and my work on it. Having a goal mattered, it motivated me to push towards them, and I had clear scales on which to grade conflicting solutions to the same problem.


Naïveté or, less pretentiously, naivety is an important goal in my world. If the inferior-but-simpler solution solves the problem at hand well enough and is not likely to cause any problems in the future, I’ll gladly stick with it. I’m very much not a purist.

This meant that most of the libraries I wrote for and in zepto are fairly naive. Some of them are very slow. I was able to get things working and, when I realized that there was a bottleneck, go fix it, instead of crafting the best solution for much longer and then realizing that it mattered very little for the overall performance of a tool.

I think naivety is important as a language starts out: the designers need to catch up with the more established languages, and need to start making theirs usable; even with a good FFI this might not always be trivial. So many libraries need to be wrapped in an idiomatic interface or, in the worst case, be written from scratch.

It also encourages the actual writing of software in the new language, and that makes catching errors and discovering needs so much easier than looking down the idiomatic ivory tower to see what the people on the bottom might need.


Everything in zepto was malleable: it had #lang constructs similar to Racket, but much more primitive. Functions were able to hook themselves up into the system by promising that they would translate the file into zepto’s AST, and zepto took over from there. It’s a simple idea, but fairly powerful.

(zepto:implements-lang <my-fun> "my-language")

Fig. 1: Hooking up a new parser can be so easy.

On the other side of the implementation, I also planned—but never really finished—a nanopass compiler that would work on zepto’s AST. The user would have been able to add passes and backends. I actually managed to write—mostly—working backends for Clojure and JavaScript.

In the end I think this was one of the factors that most contributed to zepto’s death. But we’ll talk about that later.


The most important feature of zepto was friendliness, at least for me. This meant that a friendly community was encouraged—although the “community” never grew larger than a handful people—, but also a certain developer-friendliness in all the features of the language. Intuitive, beautiful APIs played a part in that, of course, but I think every language thinks it has great APIs. More important were tools like zeps, the language package manager, written in zepto. It supported tests, bootstrapping new projects from templates, sandboxing, documentation generation, and all of the features that a developer now expects from a language. I used it a lot, and it was actually a pleasure. It was also tremendously helpful for onboarding new people.

Why it failed

This is the part where I’ll start speculating. As I see it there were many reasons why zepto failed, and some of them also plague languages that are multiple orders of magnitude more popular.

But let’s get the real main reason out of the way: working on a project as big as zepto for two years, nights and weekends only, is draining. Cloning the repo now reveals that the zepto repository and all of its submodules clocks in at around 15.000 lines of code, and zeps is another 2.500, both mostly written in zepto. And that doesn’t even include a fair share of the modules that I wrote and didn’t include in the standard library or any unmerged feature branches, and there are quite a few of those. These numbers are not gigantic, but enough where, after two years, you ask yourself “why am I doing this?”.

And that gets to the heart of it: while in the beginning I was content to answer that question simply with “because it’s fun”, I realized that it got less fun as time dragged on. And it didn’t seem as if anyone would be interested in taking over the development of my little baby—I can’t blame them, what with the aforementioned cruft and all. I was hesitant to recommend using or trying it to anyone, because I know all of its ugly nooks and crannies, and there are oh so many. My sales pitches were plagued by insecurity.

I got more serious and more ambitious over time, but I failed to meet most of the goals I set out for myself. Some of them were even at odds with each other. A nanopass compiler with pluggable backends, but also a foreign function interface? How’s that supposed to work exactly? And what about all of the complex data types zepto supports out of the box? Do we release large preludes for every language we support? Those questions never got an answer, even though I knew they needed one if I was to push forward.

Having everything be malleable didn’t help. It complicated every part of the design, and made the language more opaque. If everyone can have their own language, noone has any way to communicate. I’m still not a fan of limiting a language just because a feature can be abused, instead hoping that people are responsible, in the Kantian humanist sense1. But I’ll write about operator overloading and macros another time.

In summa, two main reasons are to blame for the death of zepto: my own inability to work on it anymore, and the lack of a technical goal for it. It had plenty of goals, and a general ideology, but I didn’t know how to best reach a state where the system was close to its Platonian idea.


Maybe I gave up on zepto too soon; maybe a general regrouping and rewrite would’ve helped. But as readers of this blog know, I never lack ideas for side projects, and I’m currently in a love affair with Carp, as you’ll realize if you scroll through my recent post output. I’m just grateful for the time I spent on zepto, for I learned a lot, and happy I have time to tackle other challenges again.

(define (goodbye)
  (write "Goodbye, zepto!"))
Fig. 2: Saying goodbye.

1. The word I’m looking for is a translation of the word “Mündigkeit”, a German term Kant used extensively, notably in his “Beantwortung der Frage: Was ist Aufklärung?”. It’s very close to “responsibility”, but I don’t think it was translated that way.