Description
Currently, our CI cabal installs cabal-plan on every validate job. Here're timings of every step in a Validate ubuntu-20.04 ghc-9.4.2 run (only longer-than-minute steps are shown):
| Step | Time |
|---|---|
| Run haskell/actions/setup@v2 | 2m 19s |
| Install cabal-plan | 5m 42s |
| Validate build | 15m 5s |
| Validate lib-test | 1m 9s |
| Validate lib-suite | 4m 33s |
| Validate cli-test | 4m 31s |
| Validate cli-suite | 11m 29s |
My rough calculation is that the cabal-plan step takes more than 10% of time (~13% perhaps). I think it's a non-negligable price. Why do we pay it?
If you look into validate.sh -- the actual validation script, it uses cabal-plan for two use cases:
cabal-plan list-binto find executables built bycabalcabal-plan topto list all packages planned for the build topologically sorted.
The first use case is obsolete since we have cabal list-bin. It should be easy to swap one for another.
The second use case doesn't feel useful to me. Does anyone have a different feeling? Does it make sense to pay 13% of time for it?
If the answer for the latter question is No, I think we should consider removing dependency on cabal-plan.
There's one more reason I'd like to remove this dependency: cabal-plan does not update for newer GHCs particularly fast. E.g. GHC 9.6.1 has been released more than one month ago, and cabal-plan still does not build without allow-newer.
If people think that we should have cabal-plan in CI, I think we could do it and solve the two problems I mention here via prebuilt binaries: downloading those will both save most of the time and relieve us of compatibility considerations.