Automate formatter on PRs

[AoifeHughes]

We should automate the running of the julia formatter on PRs.

I propose 2 options:

1. We use https://pre-commit.ci/ and add a pre-commit config
2. Add a custom workflow - I've added an example PR for this here: formatter example #2595

Possibly CRef: #2241

[AoifeHughes]

https://github.com/TuringLang/Turing.jl/actions/runs/15778945304/job/44479679379?pr=2595

[AoifeHughes]

I think the fact that the main branch had formatting issues: c064825 fixed: https://github.com/TuringLang/Turing.jl/actions/runs/15779419598/job/44481147605?pr=2595

Justifies this being needed. This also takes the burden away from developers.

[penelopeysm]

I think the fact that the main branch had formatting issues

This is probably because the workflow is using JuliaFormatter v2. The repo is formatted according to v1 and the current github action to suggest changes uses v1 as well (https://github.com/TuringLang/actions/blob/e49af9fafe0df446f473ffeb936f987b4b1c17b7/Format/action.yml#L19-L22)

v2 is still quite buggy (I just checked this morning) and IMO not ready for production use

[mhauru]

I'd be in favour of 1. I already use pre-commit locally.

I'm not keen on 2. because auto-commits sound a bit intrusive to me. We already nag about formatting issues with ReviewDog.

I think the formatting issues on main are probably due to it using JuliaFormatter v2, which we've made a conscious decision not to upgrade to yet, because it's quite buggy. (I feel like there's an issue or old PR about that conversation, but I can't find it.) EDIT: Ah, Penny got there first.

[penelopeysm]

i actually hate the automated format comments haha

[AoifeHughes]

I'm not keen on 2. because auto-commits sound a bit intrusive to me. We already nag about formatting issues with ReviewDog.

If you run your pre-commit locally then it wouldn't matter or affect you.

[AoifeHughes]

I think the formatting issues on main are probably due to it using JuliaFormatter v2, which we've made a conscious decision not to upgrade to yet, because it's quite buggy. (I feel like there's an issue or old PR about that conversation, but I can't find it.) EDIT: Ah, Penny got there first.

Fair, I could change that.

[AoifeHughes]

I'd be in favour of 1. I already use pre-commit locally.

Though, this shifts the burden to the user/dev and if we're going to do that then we need to provide that in a contributing.md as dev instructions

[penelopeysm]

As discussed on Slack, I'd be very happy to have a default pre-commit config on the repo, because that's an opt-in mechanism (even if you clone the repo, you still have to run pre-commit install to actually use it).

Agree with having setup instructions somewhere.

This is the most basic setup:

repos:
- repo: local
  hooks:
  - id: format
    name: format
    language: system
    entry: julia -e 'using Pkg;  JuliaFormatter; format(".")'

It's annoyingly ugly to enforce the version, though:

Pkg.dependencies()[Base.UUID("98e50ef6-434e-11e9-1051-2b60c6c9e899")].version.major == 1 || error("Install v1 please")

Advanced setup to avoid TTFFormat

.pre-commit-config.yaml

repos:
- repo: local
  hooks:
  - id: format
    name: format
    language: system
    entry: bash -c 'source ~/.bash_profile; jf'

.bash_profile

# JuliaFormatter binary ... https://github.com/domluna/JuliaFormatter.jl/issues/633#issuecomment-1518805248
export _JULIA_FORMATTER_SO=$HOME/.julia/formatter.so
jf() {
    project=${1:-$PWD}
    OLD=$PWD
    if [ -e "$_JULIA_FORMATTER_SO" ]; then
        :
    else
        echo "Could not find $_JULIA_FORMATTER_SO, so will build it..."
        _build_jformat
    fi
    cd $project
    julia --startup-file=no --threads=auto -J $_JULIA_FORMATTER_SO -O0 --compile=min -e 'using JuliaFormatter; format(".")'
    if [ $? -ne 0 ]; then
        printf "\n\nFailed to run JuliaFormatter; you may need to regenerate the sysimage. To do this, run the following command:\n\n    rm -f \"$_JULIA_FORMATTER_SO\"; _build_jformat\n"
        return 1
    fi
    cd $OLD
}
_build_jformat() {
    # Build a formatting image using an example project
    OLD=$PWD
    WORKDIR=$(mktemp -d)
    cd $WORKDIR
    git clone --depth 1 --quiet https://github.com/TuringLang/Turing.jl  # Not used; just an example project
    cd Turing.jl
    { 
        julia --startup-file=no --compile=yes -O3 --threads=auto -e 'using Pkg; Pkg.activate(; temp=true); Pkg.add("PackageCompiler"); Pkg.add(name="JuliaFormatter", version="1"); open("precompile_file.jl", "w") do io; write(io, "using JuliaFormatter; format(\".\")"); end; using PackageCompiler; create_sysimage(["JuliaFormatter"]; sysimage_path="'$_JULIA_FORMATTER_SO'", precompile_execution_file="precompile_file.jl")'
    } || {
        echo "Building format file failed. Exiting."
    }
    cd $OLD
}

[AoifeHughes]

As discussed on Slack, I'd be very happy to have a default pre-commit config on the repo, because that's an opt-in mechanism (even if you clone the repo, you still have to run pre-commit install to actually use it).

Agree with having setup instructions somewhere.

This is the most basic setup:

repos:

• repo: local
  hooks:
  • id: format
    name: format
    language: system
    entry: julia -e 'using Pkg; JuliaFormatter; format(".")'
    It's annoyingly ugly to enforce the version, though:

Pkg.dependencies()[Base.UUID("98e50ef6-434e-11e9-1051-2b60c6c9e899")].version.major == 1 || error("Install v1 please")

Doesn't this also add python as a requirement to the library?

I'm still unsure about having an "opt-in" version of something if we consider it mandatory? That just seems to create more steps overall

[penelopeysm]

a requirement to the library

There is a requirement to format your files, but there is no requirement to format your files with pre-commit. You can format them perfectly fine using

julia> using JuliaFormatter

julia> format(".")

So I don't see how this makes pre-commit (or Python) a project dependency.

That just seems to create more steps overall

That's the rationale behind having the GitHub bot suggest changes, so that people don't have to format locally, and can rely on the bot's suggestions to do it. I find it annoying, but in principle I am much more comfortable with that than having the bot push changes automatically.

Consider the case where there's a JuliaFormatter bug, where formatting isn't idempotent. (Note that this isn't a hypothetical; this is exactly what happened with v2, and it's also why I had to disable formatting for the README in #2575.) In that case, formatting a formatted document will lead to changes, and it will make the bot push new commits endlessly.

[AoifeHughes]

There is a requirement to format your files,

Is there?