By analyzing the blame information on this pull request, we identified @7c6f434c, @edolstra and @domenkozar to be potential reviewers

👍

Well, if we slowly replace string parsing with separate explicit specification, at some point we will get rid of the sad stories like the recent node-7f one by being able to specify name and version separately (like many package managers already do).

+1 for making version as required attribute to derivation for Nix as comparing versions is what Nix does as a supported feature.

+1 for this PR as a step to move this way (many derivations already pass version parameter).

+1 for making version as required attribute to derivation for Nix as comparing versions is what Nix does as a supported feature.

+1 for this PR as a step to move this way (many derivations already pass version parameter).

Isn't it a step towards not passing version except as a part of
derivation name?

If version was a separate attribute, we wouldn't need to parse it.

And this PR specifically remove an instance of passing the version in favour of parsing.

Is there an issue, Wiki page, mailing list posting, or some other document that explains what exactly we'd gain by having version as a separate attribute? I am totally in favor of having that, if it makes sense, but it's not obvious to me yet what that change would allow us to do that we can't do today, and probably this issue is not the right place to discuss that.

We're currently extracting version from name, having a separate attribute for version would eliminate this error prone approach. No code or docs yet, this is just my suggestion for Nix.

For what it's worth, the commits 7ca8e13 and af8c1f3 make the parsing approach slightly less error-prone.

What's the argument for smashing version into name, and then having to heuristically sniff it back out every time?

Of course the current duplication is non-ideal, but this PR doesn't do anything to improve the situation.

@nckx, if you want to improve the situation, please don't hesitate to submit a PR that does!

@peti: That makes little sense if the attribute is being deprecated.

@nckx: I think they meant to have drv.version instead of drv.meta.version.

@bjornfor, that would be grand indeed, and if that's the case I misread it as defending the status quo: whatever the case, ‘completely redundant’ a separate version attribute is very much not.

Sorry @peti — I can be bloody literal-minded sometimes. :-)

@bjornfor : I see that the commit removed drv.meta.version without adding drv.version. This is moving us further from the result you describe. I think that an optional version specification is more in spirit of meta, making an obligatory version parameter for derivation would also be fine.

Leaving no version attribute and using derivation name parsing for building the source link seems to increase fragility. Especially after #12151. Given that OpenSSL 1.0.2e and 7f are both existing pieces of software on the same planet, derivation name parsing is not likely to be completely safe.

What's the argument for smashing version into name, and then having to heuristically sniff it back out every time?

I know one use case for parsing it – getting name and version just from a path, e.g. $ nix-env -i ./result or $ nix-env -i /nix/store/foo. Of course, we could introduce $out/nix-support/meta.nix or something similar.

My two cents:
The version attribute is useful because of parsing problems. There are many softwares and their versioning varies.
It could be made as I am doing for now: putting it as meta.version or something like.