nixfmt
nixfmt copied to clipboard
Multithreading and performance optimization
It seems that we are only using a single thread when formatting tons of files. But this can be trivially parallelized. This should also ease the adoption of format-checking in git commit hook as mentioned in https://github.com/NixOS/nixpkgs/pull/322537#issuecomment-2196405557
Eg: fd -e nix --exec-batch nixfmt (--exec-batch passes as much files to nixfmt as possible in a single exec)
It takes takes 1:49.27 (~110s) on my machine with only a single thread (~99% CPU) being occupied.
We previously decided to deprecate the directory mode of nixfmt and instead encourage treefmt, for which I recently implemented parallel formatting.
So I don't think there's anything to do here, let's keep the complexity of this codebase at a minimum :)
I think that there is quite a bit of code performance which Nixfmt currently leaves on the table. Someone with knowledge of profiling Haskell applications would likely find quite a few low-hanging fruits.
Off the top of my head I can already think of https://github.com/NixOS/nixfmt/blob/master/src/Nixfmt/Predoc.hs#L178-L192 which has overall quadratic complexity, but by delaying computations into a single pass at the end this could be reduced to linear. (I didn't do it when I wrote the code because it requires a minor data types refactoring and I wanted to get stuff to work first.) Of course I can't tell how much this actually affects overall performance, but there are probably quite a few like this which add up.
You can also run it with GNU parallel: parallel "nixfmt {}" ::: *.nix
I think we can close this actually, because we now deprecated recursive mode, and are recommending treefmt instead, which does run in parallel.
This issue has been mentioned on NixOS Discourse. There might be relevant details there:
https://discourse.nixos.org/t/formatting-team-meeting-2024-12-10-and-2025-01-07/58466/1