input-output-hk
Optimise IFD?
Currently, AFAICT, haskell.nix makes extensive use of IFD in the form of essentially importing the materialisation. This is problematic because of many obvious reasons:
- IFD blocks evaluation, so if you have
(import x.outPath) + (import y.outPath), they will be built sequentially, rather than in parallel, which is a huge issue. - Hydra doesn't work well with it.
A very related issue on GH is https://github.com/NixOS/rfcs/pull/109
By putting all IFD into a separate derivation, e.g. Hydra could build this before building everything else, thus ensuring correct functioning. It would also fix the issue with parallelism. This seems like a big undertaking to me however?
However, it seems to me like the problem also lies with Nix. There doesn't seem to be a good reason for the sequential behavior of IFD, see https://github.com/NixOS/nix/issues/6299.
By putting all IFD into a separate derivation
What does this mean? IFD is, necessarily, "in a separate derivation", it's "import-from-derivation" after all. I really don't know what you're getting at here.
