Error "cycle detected" when building under specific circumstances

[Eisfunke]

Under certain circumstances I don't quite understand yet, I'm running intocycle detected errors with recent versions of haskell-flake.

I've narrowed it down to the combination of having a Paths_ (see here) defined under other-modules and using pandocunder build-depends in the cabal file, removing either one "fixes" the problem. I have absolutely no clue what happens there that doesn't seem to happen with the other packages I use.

I've created a minimal reproducing example here: https://git.eisfunke.com/software/experiments/-/tree/haskell-flake-mwe

This seems to be broken since the refactor for modular overrides in commit 996f5c2.
nix build works in my example with github:srid/haskell-flake?rev=74210fa80a49f1b6f67223debdbf1494596ff9f2 (the commit before the mentioned one).
But with github:srid/haskell-flake?rev=996f5c2cdc67285c4990df378976f9dbf26f8401 I get:

error: cycle detected in build of '/nix/store/3a103gj9pw8c9ibrhmc62r7df3vrqbwi-example-0.1.0.0.drv' in the references of output 'bin' from output 'out'

[roberth]

It sets separateBinOutput to true whenever there's an executable since that commit.
Support for this upstream in Nixpkgs Haskell is spotty, which is why it's opt-in. We probably shouldn't set separateBinOutput by default either.

[roberth]

Note that NixOS/nix#494 is a breaking change as it needs broad support in Nix's installed base before being useful, so even if someone were to prioritize an implementation effort, it would take a significant amount of time for the feature to materialize as a good default solution.

I'd suggest to remove the separateBinOutput default from haskell-flake.

[srid]

I'd suggest to remove the separateBinOutput default from haskell-flake.

Will do.

Support for this upstream in Nixpkgs Haskell is spotty, which is why it's opt-in.

Is there an upstream issue for this?

[roberth]

upstream issue

Not an issue, but the best context I could find

• Haskell optional separate bin output NixOS/nixpkgs#58525 (comment)

[Eisfunke]

Thank you all for the quick reaction, help and fix! <3 That fixes the problem for me! (although I'm still wondering why this only happened for me with the Paths_ module and the pandoc dependency...)

[roberth]

Normally I'd say nix why-depends, but I don't think that works for invalid paths.
That said, you might be able to grep it yourself. If it's not a remote build, the outputs are actually added to the store directory temporarily iirc; just not registered in the store db, so they'll be cleaned up during the next store gc.

[Eisfunke]

Yeah, while researching the problem I came across the same tip that the paths should be in the store directory, but for some reason they weren't for me (Nix 2.13.3). When I run nix build in my MWE above, the output store paths mentioned in the output ofnix log (nix path-info doesn't work for invalid paths) simply don't exist. Maybe it was changed or I've got something peculiar in my Nix config..

But anyway, it works now, so it's not really important ;)