This is my go at Advent of Code (2024) in pure nix.

In general, nix eval '.#dayX' (where 'X' is the number of the day, padded to length 2, such as nix eval '.#day03') will display the answer to a given day.

Note, many solutions require increasing the max call depth and increasing the stack size. For all solutions, consider running them as:

$ ulimit -s unlimited
$ nix eval --option max-call-depth 4294967295 ...

You also likely want to run parts one at a time, such as nix eval '.#day01.part2' in order to reduce total resource usage; I only ever run them like such, so some solutions may OOM if both parts are run together.

I'm also running this on a machine with 64GiB of memory, so solutions may take up to that amount of memory.

I also have experimental-features = ca-derivations dynamic-derivations flakes nix-command recursive-nix set in my nix show-config, though I think only flakes nix-command are used in practice in this repo.

If for some reason you want to see how slow tvix is on something, you can do that too:

$ time ./run-tvix.sh day06.part1

19.77s
max memory:                2879 MB

# as compared to nix
$ time nix eval --option max-call-depth 4294967295 '.#day06.part1'
3.07s
max memory:                2139 MB

In general tvix seems to be significantly slower, but it's interesting to compare!

Since tvix doesn't support fetchGit yet, run-tvix will populate a 'nixpkgs' symlink in the repository to use automatically.

This is done using nix, so you need nix installed as well.

On my machine, running all solutions serially takes about 2 hours:

$ time nix run --option max-call-depth 4294967295 '.#check.all'

6537.53s  user 55.99s system 99% cpu 1:50:16.46 total
max memory:                40045 MB

This is on a decently fast desktop cpu (Ryzen 7 7700X), so YMMV.

I'll update these numbers if I ever bother to make them much faster, but it's all dominated by running reasonably efficient graph algorithms (dijkstra's) just being really slow in nix, so I feel like I'd need a breakthrough on how to implemnet Dijkstra's in a way that doesn't tank nix's performance in order to get the runtime under an hour.

Part 2 takes an hour and 20GiB of memory to run on my machine.

I know there's a more optimal solution, but you know, it's fine.

Both parts take around 12GiB of memory and 1.5 minutes.

~10 minutes, but not much memory!

Thank gosh for builtins.deepSeq to avoid 60+ GiBs of thunks!

Part1 is 5 minutes.

Part 1 is ~5 minutes.

Part 1 took 37GiB of memory on my machine

Both parts take ~7.5 minutes (with 20GiB of memory usage).