[Nix-dev] fixed-output derivation that *also* depend on (some of) its inputs?

[Shea Levy]

On Wed, Oct 01, 2014 at 10:32:57AM +0200, Bjørn Forsman wrote:
> On 27 September 2014 11:12, Vladimír Čunát <vcunat at gmail.com> wrote:
> > On 09/24/2014 09:58 PM, Bjørn Forsman wrote:
> >>
> >> On 22 September 2014 17:16, Shea Levy <shea at shealevy.com> wrote:
> >>>
> >>> I suppose a flag that says to rebuild if the output is valid but doesn't
> >>> have the right deriver could be what you want? But for now the best
> >>> answer
> >>> seems to be "change the hash when it needs to change"
> >>
> >>
> >> I don't know what "deriver" is.
> >
> > Deriver is the derivation that built a path (i.e. the build recipe).
> > If you built a path yourself, you can get it by
> > nix-store -q --deriver /nix/store/path
> 
> Ah, thanks.
> 
> Shea, it looks like the deriver (*.drv) doesn't change unless "name"
> or the output hash attrs changes. So rebuild-if-deriver-changed won't
> make any difference.
> 

Nope, deriver will change if any attribute changes. Or can you show me
an example otherwise?

> 
> >> Even though it's a bit hackish, using
> >>
> >> stdenv.mkDerivation rec {
> >>    extraHash = builtins.hashString "sha256" (buildCommand);
> >>    name = "foo-${extraHash}-0.0";
> >>
> >> works just like I want it to.
> 
> Ouch, it broke real bad when using it in the real world[1]:
> 
>   fetchGrailsAppDeps = { stdenv, grails, src, sha256 }:
>     stdenv.mkDerivation rec {
>       srcHash = builtins.substring 0 20 (builtins.hashString "sha256"
> (builtins.toString src));
>       name = "grails-app-deps-${srcHash}";
>       [SNIP]
>     };
> 
> $ nix-build grails-app.nix
> error: the string `grails-app-deps-b0f29513403e6dca5372' is not
> allowed to refer to a store path (such as
> `!out!/nix/store/80bd47sghjl4i41sd783b2w49lqzkm93-git-export.drv'), at
> /nix/store/rjvnj4skg17is5x9jhzq1yl96mdvch0c-nix-1.7/share/nix/corepkgs/derivation.nix:8:12
> 
> This only happens when "src = fetchgit { ... }" and not when "src =
> ./path/to/source". Why?
> 

Use builtins.unsafeDiscardStringContext on the result of toString src
before hashString.

>
> And what does this check protect against?
> 
> Now I have a big fat warning "IMPORTANT: Remember to force re-download
> deps whenever you update the source!" in my expression :-(
> 
> Oh, this is bad...
> 
> > I think I understand what you want: the output should not change except on
> > updating the package itself, but one can only find out the hash by
> > performing the build. On update you want to do that.
> >
> > I personally force such rebuilds by changing the first few sha256 characters
> > to zeros. The extraHash way and Shea's suggestion are other fine ways. For
> > this use case, I find no clear winner, and I don't think we can do
> > significantly better.
> 
> What about a new, optional derivation input like "extraHashInputs = [
> list of attrs that will be included in calculation of $out ];"?
> 
> 
> [1]: https://gist.github.com/bjornfor/3f2fda1b1ce077225df5
> 

Please no. This is fixing the problem in the wrong place.

> 
> Best regards,
> Bjørn Forsman