Richard Towers

Terraforming the Advent of Code

I propped my laptop up on four coasters to stop it from melting the kitchen table. The thunderous noise of the fans almost drowning out the radio in the background.

For two hours, I waited. Every few minutes, a new number.

7947, 0, 7875, 40, 7539, 179, 2353, 2195…

Finally, an answer. Convergence. 2240.

The wrong answer.

I gave up on Advent of Code on day 11.

Advent of Code

Advent of Code is a series of small programming puzzles. Each December, programmers from around the world spend a little time solving puzzles - for fun, as practice, or just for the challenge.


Some folks like to pick a language that’s totally inappropriate for the job. That way, solving the easy puzzles feels like an achievement. And when the puzzles get too difficult, you can claim that it’s the language holding you back rather than your own inadequacy.

This year, I picked Terraform.

Terraform is an open-source infrastructure as code software tool that provides a consistent CLI workflow to manage hundreds of cloud services.

In other words, it’s not really a programming language at all. To my knowledge none of the Advent of Code puzzles so far have involved managing hundreds of cloud services.

Nevertheless, Terraform does include a configuration language, which is surprisingly powerful.

locals {
  # You can have lists:
  a_list = [1,2,3,4]

  # And you can do list comprehensions:
  evens = [for item in local.a_list: item if item % 2 == 0]

  # And there are lots of built in functions:
  squares = [for item in local.a_list: pow(item, 2)]

# And you can have outputs:
output "evens"   { value = local.evens   }
output "squares" { value = local.squares }

Applying the above with terraform apply -auto-approve gives:

Apply complete! Resources: 0 added, 0 changed, 0 destroyed.


evens = [
squares = [

For the easier puzzles, these features were all I needed. Some of the solutions were even quite elegant!

Terraform Triumphs

The most concise solution I found was to day 6’s puzzle. This required a little bit of string processing on the input, and then some set operations (union and intersection). Terraform has all these functions built in, so the solution was very nice:

locals {
  groups = [for group in split("\n\n", trimspace(file("input.txt"))):
    [for person in split("\n", group): split("", person)]
  anyone_yes_counts   = [for g in local.groups: length(setunion(g...)) ]
  everyone_yes_counts = [for g in local.groups: length(setintersection(g...)) ]

output "part_1_answer" {
  value = sum(local.anyone_yes_counts)

output "part_2_answer" {
  value = sum(local.everyone_yes_counts)

I also enjoyed day 5, which could be solved with a trick. The elaborate description in the question is a roundabout way of describing binary numbers, so converting the input (strings of Fs, Bs, Ls and Rs) into a binary string (0s and 1s) and then using parseint(str, 2) made the solution nice and simple.

Terraform’s surprisingly wide range of built-in functions helped out (I didn’t expect there to be a function for parsing numbers in base 2 in Terraform, but there is).

I also found Terraform’s list (and object) comprehensions to be very elegant and expressive. Anything that you would express with filter and map in a “normal” language can be nicely expressed in Terraform.

There were a few times in the easy puzzles when the solutions I came up with in Terraform were shorter and more readable than my friends’ solutions in much more well regarded languages (even Haskell). A triumph.

Terraform Troubles

Because Terraform’s purpose is managing infrastructure, they’ve deliberately avoided making the language more powerful than it needs to be. That means no user defined functions, no while loops, no recursion, and no equivalent to the reduce() function on lists.

I believe the absence of these features means that Terraform’s configuration language is not Turing Complete (although I have not proved that, so I could be wrong).

This is a sensible choice when the goal is managing hundreds of cloud services - you really don’t want your configuration going through arbitrarily complex machinations to work out what infrastructure you should be running (and paying for).

It does make Advent of Code quite tricky though.

My workaround for this was to run Terraform lots of times. On each run it would read some state from a local file, do some computations, and then write some new state to the file. The process was “Done” when Terraform reached a fixed point - i.e the file no longer changed.

Basically, wrapping Terraform with a little shell script like this:

#!/usr/bin/env bash
set -eu

output_file=$(mktemp -t aoc-terraform-output)
while ! grep -F 'Apply complete! Resources: 0 added, 0 changed, 0 destroyed.' "$output_file"
  terraform apply -auto-approve | tee "$output_file"

Diffing things is another strength of Terraform (the whole point is to converge your infrastructure into the state described by the code), so this worked pretty well too.

Unfortunately, this approach can be a bit …slow…

The Advent of Code exercises are all supposed to have solutions that complete in 15 seconds or less on modest hardware. Terraform is surprisingly fast at doing computation while it’s running. Unfortunately even an O(n) solution can be slow if you have to let Terraform refresh its state n times.

Having to wait minutes to run Terraform ~100s of times, only to discover I’d made a mistake and would need to do it all again was not fun.

Because Terraform computes it’s list comprehensions eagerly, there’s no way to say “keep going until you find something that looks like this, and then stop”. Basically this means you always get worst case performance - even if Terraform has already found the answer you’re looking for, it will keep going until it’s done all the computation you asked for.

Some Terraform functions can be slow if you’re running them in a very tight loop. I found that using try() in the middle of a nested loop that needed to run millions of times was a bad plan.


As I mentioned in the introduction, I was defeated by day 11. This one is basically Conway’s Game of Life, which was very fun to do in Terraform.

Sadly, I couldn’t come up with an O(sensible) solution for Part 2. Which lead to me letting my machine run hot for hours on end, and still coming out with the wrong answer.

While this was happening I was absent mindedly wondering about the carbon footprint of it all, and a lot of the fun evaporated.

Since then I think I’ve spotted my mistake - there’s a reasonably big bit of work that’s happening every time Terraform runs, but it’s the same answer every time. If I could just do it once, and cache the result in a local file I’d probably get through day 11. But first I’d have to work up the enthusiasm.

What did I learn?

I certainly learned a lot about Terraform, which is good because it’s a tool I use a lot at work.

Now that I’m more confident with the Terraform language, I’m starting to think about the whole tool a bit differently.

Maybe Terraform is not a Big Infrastructure Tool with a tiny language inside, but actually a Big Language in its own right 🤔

If that’s the case, then perhaps some of the best practices from the rest of software development are more applicable to Terraform than I thought.

One thing I’ve been thinking about is pure functions (functions with inputs and outputs, but no side effects) in the context of Terraform modules.

Sometimes we find ourselves using a module to create a single resource, just because there’s a bit of configuration that would be duplicated otherwise.

In a “real” language, a function that took some inputs, mapped them into a more complex structure, and then created an entire virtual machine as a side effect would be frowned upon.

It’s easier to reason about and test modules that only take inputs, do some computation, and return outputs. These modules can leave creating any resources to the caller.

I wonder if Terraform should reconsider its position on user-defined functions. If pure modules are a good idea, then perhaps giving them a nicer syntax might also be a good idea. Maybe some ideas from Dhall could be folded in. This would make future Advent of Codes much easier! (although I think that’s a non-goal for the team at Hashicorp).

Did I have fun?

Yes. Although to be fair, what else was I going to do in December 2020?

My solutions

If you’re curious, all of my solutions are in this git repository.