Can flake-parts solve cross-system evaluation with IFD?

[shlevy]

Consider a Haskell project built with haskell.nix. We want flake outputs and hydra builds for Linux and Darwin. Flake outputs are fine, just use perSystem or the equivalent functionality in flake-utils. But Hydra presents a problem.

Hydra jobs are evaluated once for all systems, on a Linux machine. So if we have the hydraJobs just build all of the packages, we'll run into a problem with the IFD of haskell.nix: It will require a Darwin builder at evaluation time, which may not be available then and certainly will be slow. haskell.nix does allow you to specify an evalSystem for your project evaluation, but if we set the evalSystem unconditionally to Linux then local darwin users will not be able to evaluate the flake outputs.

My current approach to solving this is to define a function taking system and evalSystem, with the latter defaulting to the former, then call the function just setting system for each system's packages but call it setting each system but the Linux evalSystem when building hydraJobs.

Could flake-parts give us some general way to do this? Maybe some kind of general way to re-call self with an extra module, or alternatively have perSystem be callable twice somehow?

[roberth]

Yes, this is possible, but don't you want to make sure that IFD works on the native system?

[shlevy]

We can have a job in the jobset that ensures all the deps build and even evaluate.